You are on page 1of 7

Automatica 48 (2012) 844–850

Contents lists available at SciVerse ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Brief paper

Design of a modified repetitive-control system based on a continuous–discrete


2D model✩
Jinhua She a,b , Lan Zhou a , Min Wu a,1 , Jie Zhang a , Yong He a
a
School of Information Science and Engineering, Central South University, Changsha 410083, China
b
School of Computer Science, Tokyo University of Technology, Tokyo 192-0982, Japan

article info abstract


Article history: This paper concerns the problem of designing a modified repetitive-control system for a class of strictly
Received 28 November 2009 proper plants. Repetitive control involves two types of actions: control and learning; but the insertion of
Received in revised form a low-pass filter in a modified repetitive controller, which is employed to guarantee the stability of the
23 August 2011
system, mixes the two actions together. In this paper, a continuous–discrete two-dimensional model is
Accepted 26 September 2011
Available online 19 March 2012
first constructed. Next, the continuity of repetitive control and Lyapunov stability theory are applied to
the model to establish two linear-matrix-inequality (LMI) based sufficient stability conditions, one for the
Keywords:
design of the cutoff angular frequency and one for the design of the feedback gains. The features of these
Repetitive control conditions are exploited to develop an iterative algorithm that searches for the best combination of the
Two-dimensional (2D) system maximum cutoff angular frequency of the low-pass filter and the feedback gains. A numerical example
Linear matrix inequality (LMI) illustrates the design procedure and demonstrates the validity of the method.
Lyapunov stability theory © 2012 Elsevier Ltd. All rights reserved.

1. Introduction Ryoo, & Chung, 2006), process control (Álvarez, Yebra, & Berenguel,
2007), etc.
Repetitive control (RC) performs learning by means of a Iterative learning control (ILC) is another well-known method
repetitive controller that contains a pure-delay positive-feedback that makes use of previous control trials and is closely related to
line. This control method simulates the behavior of human RC. However, as pointed out by Arimoto (1998), Hara (1986) and
learning. For a given periodic reference input, a repetitive others, there are significant differences between them. First, the
controller gradually improves the tracking precision through initial state of a period is different. For an RC system (RCS), the
repeated learning actions, which involve adding the control input state at the beginning of a period is the same as the final state
of the previous period to that of the present period to regulate of the system in the previous period. However, in an ILC system
the present control input. As a result, the tracking error is reduced (ILCS), the reference trajectory is defined over a finite time interval,
step by step; finally, the output tracks the reference input without and the state of an ILCS is usually reset after every period. The
steady-state error. Due to its easy implementation and high control second difference between an RC and an ILC concerns the design
precision, RC provides an effective and practical solution to many
problem. In an RCS, learning occurs through periodic delay-based
control problems in the fields of mechanical systems (Crudele &
updates in a neutral-type delay system (Fig. 1). An RCS can only
Kurfess, 2003), power systems (Roncero-Sanchez, Acha, Ortega-
be stabilized when the relative degree of the plant is zero. To
Calderon, Feliu, & Garcia-Cerrada, 2009), motion control (Doh,
guarantee the stability of a strictly proper plant, which is the type
that most control engineering applications deal with, a repetitive
controller has to be modified by the insertion of a low-pass filter in
✩ The work of J. She was supported by the Kurata Memorial Hitachi Science and
the delay line. The resulting system is called a modified RCS (MRCS)
Technology Foundation. The work of L. Zhou, M. Wu, and J. Zhang was supported
(Hara, Yamamoto, Omata, & Nakano, 1988). This restriction does
by the National Natural Science Foundation of China under Grant Nos. 60974045
and 60674016, the National Science Fund for Distinguished Youth Scholars of China not apply to an ILCS, which is easy to stabilize even for a strictly
under Grant No. 60425310, and Hunan Natural Science Foundation under Grant No. proper plant. Thus, the stability analysis of an ILCS in Galkowski
11JJ4059. The material in this paper was not presented at any conference. This paper et al. (2003), Galkowski, Paszke, Sulikowski, Rogers, and Owens
was recommended for publication in revised form by Associate Editor Yasumasa (2002) and Hladowski et al. (2010), which use the 2D system
Fujisaki under the direction of Editor Roberto Tempo.
approach and linear matrix inequalities (LMIs), cannot be applied
E-mail addresses: she@cs.teu.ac.jp (J. She), zlly98@yahoo.cn (L. Zhou),
min@csu.edu.cn (M. Wu), zhangjie@wuhua.csu.edu.cn (J. Zhang), directly to an RCS.
heyong@csu.edu.cn (Y. He). On the other hand, Hladowski, Rogers, Galkowski, and Virendra
1 Tel.: +86 731 88836091; fax: +86 731 88836091. (2008) considered the problem of discrete linear repetitive process
0005-1098/$ – see front matter © 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2012.02.019
J. She et al. / Automatica 48 (2012) 844–850 845

enables preferential adjustment of the control and learning actions.


The resulting control system exhibits both satisfactory robustness
and good control performance. However, in an MRCS, the low-
pass filter mixes these two actions together. Thus, the method in
Wu et al. (2011) and Wu et al. (2010) cannot simply be extended
to the design of an MRCS for a strictly proper plant. To perform
Fig. 1. Configuration of a basic repetitive-control system.
preferential adjustment of control and learning in an MRCS, we
need to devise a new 2D design approach that exploits the inherent
control with dynamic boundary conditions. Their method is 2D structure of the system.
applicable to a discrete RCS. Note that this paper deals with the
This paper presents a method of designing an MRCS for strictly
problem of continuous repetitive control. The level of difficulty
proper plants that solves the trade-off problem. We devise an
involved in discrete RC is entirely different from that for con-
iterative algorithm that optimizes both the maximum cutoff
tinuous RC. That is, it is easy to stabilize a discrete RCS, but hard
angular frequency of the low-pass filter and the feedback gains to
for a continuous RCS. Furthermore, there is an issue regarding
handle the trade-off problem, and employ two tuning parameters
control performance between sampling points. As pointed by
in an LMI-based stability condition, which enable preferential
Chen and Francis (1995), for a discrete system, even though the
control performance is satisfactory at sampling points, there is no adjustment of control and learning, to improve the transient
guarantee that it is so between them, and undesirable oscillations control performance. First, a new continuous–discrete 2D model
may occur there. is established that accurately describes the characteristics of an
An MRCS has two kinds of design parameters: feedback gains, MRCS. The parameters of a 2D control law can be used directly
which are mainly related to robust stability; and the cutoff angular to preferentially adjust the control and learning actions. Next, two
frequency of the low-pass filter, which is mainly related to the LMI-based sufficient stability conditions for the control system are
tracking precision. These parameters affect each other; and there derived. They are used to convert the problem of determining the
is a trade-off between them. All reported methods of designing an maximum cutoff angular frequency of the low-pass filter into an
MRCS are mainly concerned with designing a stabilizing controller SGEOP, which is presented in an LMI, and to convert the problem of
for a fixed low-pass filter, or vice versa. Since they do not take the determining the gains of the control law into a feasibility problem
relationship between the filter and controller into consideration, for another LMI. Two tuning parameters in one of the LMIs are
they cannot resolve the trade-off effectively (Chen & Liu, 2005). used to adjust control and learning. Then, an iterative algorithm
Doh and Chung (2003) presented a design algorithm based on is presented that produces the best combination of low-pass filter
a linear matrix inequality (LMI) that finds the maximum cutoff and feedback gains. Finally, a numerical example illustrates the
angular frequency of the low-pass filter. Since it employs a fixed design procedure and demonstrates the validity of the design
weighting matrix to solve a standard generalized eigenvalue method.
optimization problem (SGEOP), the algorithm is conservative. In Throughout this paper, R+ is the set of non-negative real num-
addition, that report only considered the design of a low-pass filter bers; Cp is p-dimensional vector space over complex numbers; Z+
for a given stabilizing controller; but there was no discussion of the is the set of non-negative integers; ℵ is the linear space of all the
design of the controller itself. Thus, how to maximize the cutoff functions from [0, T ] to Cp . L2 (R+ , Cp ), or just L2 , is the linear space
angular frequency of the low-pass filter and to find stabilizing of square integrable functions from R+ to Cp ; and ℓ2 (Z+ , ℵ), or just
feedback gains, e.g., simultaneously optimize these parameters, is l2 , is the linear space of all the functions from Z+ to ℵ (discrete-
a challenging problem. time signal).
A close examination of repetitive control shows that it
involves two different actions: continuous control within each
repetition period and discrete learning between periods. A 2. Problem description
design method that enables independent adjustment of the
two can potentially provide better transient performance and This paper considers the MRCS in Fig. 2. The plant is a single-
better tracking performance. However, most design methods were input single-output linear system:
developed for one-dimensional (1D) space – the time domain –
ẋp (t ) = Axp (t ) + Bu(t ),

and focus mainly on the stability issue. They consider only the
(1)
overall effects of the two actions. That makes it very difficult, if y(t ) = Cxp (t ),
not impossible, to dramatically improve the transient performance
(She, Wu Lan, & He, 2010). A better solution is the method in where xp (t ) ∈ Rn is the state of the plant; u(t ), y(t ) ∈ R are
Wu, Zhou, and She (2011) and Wu, Zhou, She, and He (2010), the control input and output, respectively; and A, B, and C are real
which employs two-dimensional (2D) system theory (Bose, 2003) constant matrices. Note that the relative degree of the plant is
to design a robust RCS. Describing the control action with a greater than zero; that is, there is no direct path from u(t ) to y(t ).
continuous differential equation and learning with a discrete r (t ) is a periodic reference input with a period of T , and e(t ) =
difference equation yields a continuous–discrete 2D model. Unlike r (t ) − y(t ) is the tracking error. CMR (s) is the modified repetitive
a 1D method, the one in Wu et al. (2011) and Wu et al. (2010) controller, and q(s) is a low-pass filter that ensures the stability of

Fig. 2. Configuration of a modified repetitive-control system.


846 J. She et al. / Automatica 48 (2012) 844–850

the control system. Without loss of generality, we assume that q(s) Formulating the design problem in 2D space allows us to adjust
is a first-order filter; that is, control and learning by tuning Fp and Fe , respectively. This provides
q(s) = ωf /(s + ωf ). (2) a way of improving the tracking performance. It is a significant
advantage over other methods.
The state-space representation of the modified repetitive con-
troller is
3. Design of MRCS
ẋf (t ) = −ωf xf (t ) + ωf xf (t − T ) + ωf e(t ),

(3)
v(t ) = e(t ) + xf (t − T ), Substituting (7) into (6) gives the following representation of
where ωf is the cutoff angular frequency of the low-pass filter, the closed-loop MRCS:
xf (t ) ∈ R is the state variable, and v(t ) is the output of the modified
ẋ(k, τ ) = Āl x(k, τ ) + Ādl x(k − 1, τ ),

repetitive controller. (11)
The feedback control law for the control system is e(k, τ ) = C̄ x(k, τ ),
where
u(t ) = Ke v(t ) + Kp xp (t ), Ke ∈ R, Kp ∈ R1×n , (4)    
and Ke and Kp are feedback gains. This paper considers the A + BFp 0 0 BFe
Āl = , Ādl = . (12)
following design problem: −ωf C −ωf 0 ωf
Find a combination of the largest cutoff angular frequency, ωf , of (11) shows that control and learning affect each other by means of
the low-pass filter in (2) and the feedback gains, Ke and Kp , in (4) that Āl and Ādl . The following lemmas are employed in the derivation of
stabilizes the system and ensures satisfactory tracking performance in
stability conditions for the system (11).
both the transient and steady states.
(3) and (4) describe the MRCS in the time domain. Note that (4) Lemma 1. If there exists a semi-positive definite functional V (k, τ )
is for the MRCS in Fig. 2 that is continuous and decreases monotonically
u(t ) = Ke [e(t ) + xf (t − T )] + Kp xp (t ) in every interval [kT , (k + 1)T ) , k ∈ {0, 1, 2, . . . }, then the MRCS is
asymptotically stable.
= {Ke r (t ) + (Kp − Ke C )xp (t )} + Ke xf (t − T ). (5)
The control law (5) contains two kinds of information: information Proof. From the characteristics of repetitive control, we know that
on the present period (first term on the right side) and information V (k, T ) = V (k + 1, 0) holds for the MRCS in Fig. 2. So, V (t ) =
on the previous period (second term). Since we can view the past V (kT + τ ) := V (k, τ ) is continuous and decreases monotonically
state as a kind of experience, the words control and learning mean in [0, ∞). From Merkin (1997), the MRCS in Fig. 2 is asymptotically
that we use information on the present and previous periods, stable. 
respectively, to produce the present control input. So, we cannot
individually adjust the control and learning actions by directly Lemma 2 (Khargonekar, Petersen, & Zhou, 1990). For any real matrix
tuning the control gains, Ke and Kp , in (4). To solve this problem Σ = Σ T , the following assertions are equivalent:
 
and thereby improve the control performance, below we present a (1) Σ =
S11 S12
< 0;
2D description of the MRCS in Fig. 2. ⋆ S22

First, we employ the lifting technique (Yamamoto, 1994) to slice (2) S11 < 0 and S22 − S12 S11 S12 < 0; and
T −1

the time axis, [0, +∞), into intervals of length T and convert a (3) S22 < 0 and S11 − S12 S22 S12 < 0.
−1 T
vector-valued continuous-time signal, ξ (t ), into a function-valued
discrete-time sequence, {ξk (τ )}. Its element is denoted ξ (k, τ ) in Lemma 3 (Doh & Chung, 2003). If there exist symmetrical positive
this paper. That is, definite matrices P and Q such that
ξ (k, τ ) = ξk (τ ) := LC [ξ (t )], t = kT + τ , τ ∈ [0, T ], k ∈ Z+ ,  
P Āl + ĀTl P + Q P Ādl
where LC is an isometric and isomorphic transformation between <0 (13)
⋆ −Q
L2 (R+ , Cp ) and ℓ2 (Z+ , ℵ). Since the stability of the system does not
depend on an exogenous signal, we set r (t ) = 0. Then, we obtain holds, then the closed-loop system (11) is asymptotically stable.
the following 2D representation from (3) and (4):
Since Āl and Ādl in (12) contain the design parameter ωf , (13) is
ẋ(k, τ ) = Āx(k, τ ) + A¯d x(k − 1, τ ) + B̄u(k, τ ), not an LMI. Thus, it cannot be used directly to design Fp , Fe , and

(6)
e(k, τ ) = C̄ x(k, τ ), ωf . However, as shown below, for given Fp and Fe , (13) can be
transformed into an LMI, thereby allowing us to use it to determine
u(k, τ ) = Fp 0 x(k, τ ) + [0 Fe ] x(k − 1, τ ),
 
(7) the cutoff angular frequency, ωf . Let
where
ωf = ω̂f + δωf , (14)
x (k, τ ) = xTp (k, τ ), xTf (k, τ ) ,
 T  

     where ω̂f and δωf are a rough estimate and the corrected value to
 A 0 0 0
Ā = , ¯
Ad = , (8) be determined, respectively. Then, we can write Āl and Ādl in (12)
−ωf C −ωf 0 ωf

 as
0 , C̄ = − C 0 ,
 T  T   
B̄ = B
Āl = Āl0 + δωf × Āl1 , Ādl = Ādl0 + δωf × Ādl1 , (15)
Fp = −Ke C + Kp , F e = Ke . (9) where
The relationships between the feedback gains in (4) and (7) are 
A + BFp 0
 
0 0

Ke = Fe , Kp = Fp + Fe C . Āl0 = , Āl1 = ,
(10) −ω̂f C −ω̂f −C −1
Based on the continuous–discrete 2D model, (6) and (7), the    
0 BFe 0 0
problem of designing the MRCS in Fig. 2 can now be formulated Ādl0 = , Ādl1 = .
0 ω̂f 0 1
as follows:
Find a combination of the stabilizing feedback gains, Fp and Fe , For Q in (13), we assume that
in the 2D control law (7) and the largest cutoff angular frequency,
Q = Q0 − δωf × Q1 > 0, (16)
ωf , that stabilizes the continuous–discrete 2D system (6) and yields
satisfactory control and learning actions. where Q0 and Q1 are symmetrical matrices.
J. She et al. / Automatica 48 (2012) 844–850 847

From (15) and (16), we can write the stability condition (13) in where
terms of the LMI 
1
 
1

P12 = diag P1 , P2 , Q12 = diag Q1 , Q2 ,
Ξ0 + δωf × Ξ1 < 0, (17) α β
where P1 = X1−1 , P2 = X2−1 , Q1 = Y1−1 , Q2 = Y2−1 .
 
P Āl0 + ĀTl0 P + Q0 P Ādl0 From (11), V (k, τ ) defined in (23) is a continuous function of τ .
Ξ0 = ,
⋆ −Q0 Lemma 1 shows that we only need to consider the monotonicity
(18)
of V (k, τ ) with respect to τ within the interval [kT , (k + 1)T ). A
 
P Āl1 + ĀTl1 P − Q1 P Ādl1
Ξ1 = .
⋆ Q1 simple calculation shows that

So, we have the following theorem. ∂ V (k, τ )


= 2xT (k, τ )P12 ẋ(k, τ ) + xT (k, τ )Q12 x(k, τ )
∂τ
Theorem 1. For given ω̂f , Fp , and Fe , if there exist symmetrical
− xT (k − 1, τ )Q12 x(k − 1, τ )
matrices Q0 and Q1 , and a positive-definite matrix P such that
LMIs (16) and (17) hold, then the cutoff angular frequency ωf = ηT (k, τ )Φ η(k, τ ), (24)
in (14) guarantees the asymptotic stability of the closed-loop
system (11). where
T
η(k, τ ) = xT (k, τ ), xT (k − 1, τ ) ,

Furthermore, if we let (25)
δωf = 1/γ , (19)  1

Φ + Q1 Φ12 0 P1 BFe
we can obtain the maximum cutoff angular frequency, ωfm (=ω̂f +  11 α 
1/γ ), for given Fp and Fe , and a rough estimate ω̂f , by solving the 1
 
 ⋆ Φ22 + Q2 0 ωf P2 
SGEOP Φ= β , (26)
 

min γ > 0 s.t. Q1 < γ Q0 and Ξ1 < −γ Ξ0 . ⋆ ⋆ −Q1 0 


 
(20) 
 1 
Note that this SGEOP can be solved by using the function gevp ⋆ ⋆ ⋆ − Q2
β
in the Robust Control Toolbox of Matlab (Balas, Chiang, Packard,
& Safonov, 2005). 1 1
Φ11 = (AT + FpT BT )P1 + P1 (A + BFp ),
α α
Remark 1. How to find a Q and employ it in Lemma 3 to calculate Φ12 = −ωf C T P2 , Φ22 = −2ωf P2 .
an ωf is a problem. Doh and Chung (2003) and Doh et al.
(2006) used a fixed Q to transform (13) into a standard linear Clearly, if Φ < 0, then V (k, τ ) decreases monotonically in
objective minimization problem. In this study, we reduce the [kT , (k + 1)T ) for any k ∈ {0, 1, 2, . . . }. Thus, the system in Fig. 2
conservativeness by adjusting Q using δωf , and transform the is stable. Also, from Lemma 2, Φ < 0 is equivalent to the LMI
problem of calculating the maximum cutoff angular frequency,
ωfm , into an SGEOP. This produces a less conservative result than
 1 
Φ11 Φ12 0 P1 BFe Q1 0
a fixed Q does. However, the transformation introduces some  α 
conservativeness because, to solve the SGEOP (20), Q0 must be a  ⋆ 1
ωf P2
 
Φ22 0 0 Q2 
positive definite matrix, while it only has to be symmetrical in the 
 β 

stability condition of Theorem 1 and Lemma 3.  ⋆ ⋆ −Q1 0 0 0 
 < 0. (27)
 
On the other hand, for a given cutoff angular frequency, ωf , we 1

 ⋆ ⋆ ⋆ − Q2 0 0

obtain the following stability condition of the system (11) by using β
 
 
the Lyapunov stability theorem.  ⋆ ⋆ ⋆ ⋆ −Q1 0
 

1
 
Theorem 2. For given positive scalars ωf , α , and β , if there exist ⋆ ⋆ ⋆ ⋆ ⋆ − Q2
symmetrical positive definite matrices X1 , X2 , Y1 , and Y2 together with β
arbitrary matrices W1 and W2 such that Pre- and post-multiplying the matrix on the left side of (27) by
diag {α X1 , X2 , Y1 , β Y2 , Y1 , β Y2 } yields the LMI (21). 
Θ11 T
β BW2 α X1
 
−αωf X1 C 0 0
 ⋆ −2ωf X2 0 βωf Y2 0 X2  Remark 2. Theorem 2 constitutes an LMI-based design algorithm
 ⋆ ⋆
 
−Y 1 0 0 0 
<0 (21) for obtaining the control gains that is based on the continuity and
 ⋆ ⋆ ⋆ −β Y2 0 0 

 ⋆ the time-delay property of the MRCS. Even though the form of the
⋆ ⋆ ⋆ −Y 1 0  Lyapunov functional (23) is different from that in Wu et al. (2010),
⋆ ⋆ ⋆ ⋆ ⋆ −β Y2 the two functionals have the same effect. To see this, we let
holds, where −1 T
C + = C̄ T C̄ C̄ T , Q̄12 = C + Q12 C + ,

Θ11 = α X1 A + α AX1 + α
T
W1T BT + α BW1 ,
and use (11) to write (23) as
then the MRCS (11) is asymptotically stable. Furthermore, the feed-  τ
back gains in (7) are given by
Ṽ (k, τ ) = x (k, τ )P12 x(k, τ ) +
T
eT (k, s)Q̄12 e(k, s)ds. (28)
τ −T
Fp = W1 X1 ,−1
Fe = W2 Y2 .−1
(22)
Then, we have
Proof. We choose the following Lyapunov functional candidate ∂ V (k, τ )
= 2xT (k, τ )P12 ẋ(k, τ ) + eT (k, τ )Q̄12 e(k, τ )

 τ ∂τ
V (k, τ ) = xT (k, τ )P12 x(k, τ ) + xT (k, s)Q12 x(k, s)ds, (23)
− eT (k − 1, τ )Q̄12 e(k − 1, τ ) .

τ −T (29)
848 J. She et al. / Automatica 48 (2012) 844–850

Fig. 3. Flowchart of an algorithm.

The second part of (29) is equivalent to 1V2 (k, τ ) in Wu of the maximum cutoff frequency, ωfm , and the control gains,
et al. (2010), which guarantees that the tracking error decreases Fp and Fe .
monotonically from period to period. In addition, the LMI-based
stability condition (21) can be used directly in designing the control Remark 4. The two conditions, (20) and (21), are independent
gains. of each other and have different purposes: One is for the design
of the cutoff angular frequency, and the other is for the design
of the feedback gains. To optimize the cutoff angular frequency
Remark 3. Condition (21) in Theorem 2 contains two tuning and the feedback gains simultaneously, the algorithm in Fig. 3
parameters, α and β , that make it possible to preferentially adjust stops only when there are no more solutions to SGEOP (20) or
control and learning. More specifically, we can use α and β to (21). On the other hand, it is possible to use just Theorem 2 to
adjust the weighting matrices X1 and Y2 in (21), respectively, and construct a bisection search algorithm to find the largest ωfm
Th2
and
thus the feasible solutions Fp and Fe in (22). This is how we adjust the corresponding FeTh2 and FpTh2 . However, if we insert ωfm
Th2
, FeTh2 ,
control and learning.
and FpTh2 into SGEOP (20) and it has a solution, then we can find a
Theorems 1 and 2 present two stability conditions for the MRCS new ωfm (>ωfm Th2
) that is better than the one given by Theorem 2.
in Fig. 2. By exploiting the features of these two conditions, we can So, our algorithm produces a larger cutoff angular frequency for the
use the iterative algorithm (Fig. 3) to find the best combination low-pass filter than using just Theorem 2 does.
J. She et al. / Automatica 48 (2012) 844–850 849

Fig. 4. Tracking error for parameter set (a) α = 0.1, β = 0.2; (b) α = 0.1, β = 0.5; (c) α = 1, β = 1; (d) α = 1.8, β = 1.

Fig. 5. Simulation results for α = 1.8 and β = 1.2.

4. Numerical example time. So, adjusting β mainly effects learning, rather than control.
The tracking error in the first period is smaller in (d) than in (c). So,
Consider the speed control of a rotational system that consists adjusting α mainly affects control.
of two DC motors: one is the controlled object, and the other Based on the evaluation of the performance index (31), we
is a disturbance generator (She, Fang, Ohyama, Hashimoto, & finally selected
Wu, 2008). Their axles are coupled together with a spring. The
parameters of the plant are α = 1.8, β = 1.2. (34)

28.06
 
−31.31 −2.833 × 104 The resulting control gains and maximum cutoff frequency of the
 
0
A= 0 −10.25 8001 , B= 0 , low-pass filter (2) are
1 −1 0 0
Ke = 21.4443, Kp = −9.6 1009.5
 
0 (35)
0 .
 
C = 1 0
ωfm = 540.5651 rad/s (36)
We employ the above algorithm to design an MRCS (Fig. 2) to track
the reference input and the performance index is
r (t ) = sin π t + 0.5 sin 2π t + 0.5 sin 3π t . J10 = 0.0936. (37)
Set The simulation results (Fig. 5) show that the rotational control
ε = 10 , −3
h = 0.1, ω̂f = 20 rad/s. (30) system is stable and quickly enters the steady state.

Choose the performance index


5. Conclusion
10  kT
1
J10 = e2 (t )dt (31) To enable repetitive control to be applied to a wider range
2 k=1 (k−1)T
of control engineering problems, we developed a method of
as a criterion for the selection of the tuning parameters, α and designing an MRCS for a class of strictly proper plants. Since the
β . We show what effect adjusting the tuning parameters has by low-pass filter in an MRCS mixes the control and learning actions,
carrying out simulations on four parameter sets: methods for an RCS cannot simply be extended to an MRCS. We
(a) : α = 0.1, β = 0.2; (b) : α = 0.1, β = 0.5; derived a new 2D representation of the system in which the

(c) : α = 1, β = 1; (d) : α = 1.8, β = 1. 2D control law contains the direct sum of the effects of control
and learning. This allows preferential adjustment of control and
(32) learning through the regulation of the corresponding gains in the
The performance indices and the maximum cutoff frequencies control law. Then, we used the Lyapunov stability theory to derive
are two LMI-based sufficient stability conditions that can be used
to design the low-pass filter and the 2D feedback control law.
(a) : a
J10 = 0.5209, ωfm
a
= 99.8364 rad/s


 Combining the features of these two conditions, we established
(b) :
 b
J10 = 0.4363, ωfm
b
= 718.2389 rad/s an iterative algorithm that finds the maximum cutoff angular
(33) frequency of the low-pass filter and the control gains. Specifically,
(c) : c
J10 = 0.2438, ωfm = 868.6327 rad/s
c

two tuning parameters in one of the LMIs enable the preferential




( d) : d
= 0.1068, ωfm
d
= 837.2533 rad/s.

J10 adjustment of control and learning. Finally, a numerical example
Fig. 4 shows that the tracking error converges faster for set (b) than shows that the resulting system has satisfactory stability and good
for (a), and that both enter the steady state at almost the same tracking performance.
850 J. She et al. / Automatica 48 (2012) 844–850

References and M. Nakano). His research interests include the application of control theory,
repetitive control, process control, Internet-based engineering education, and
Álvarez, J. D., Yebra, L. J., & Berenguel, M. (2007). Repetitive control of tubular heat robotics.
exchangers. Journal of Process Control, 17(9), 689–701.
Arimoto, S. (1998). A brief history of iterative learning control. In Z. Bien, &
J.-X. Xu (Eds.), Iterative learning control analysis, design, integration and
applications (pp. 1–7). Boston, Dordecht, London: Kluwer Academic Publishers.
Lan Zhou received the B.S. degree in 1998 from Hunan
Balas, G., Chiang, R., Packard, A., & Safonov, M. (2005). Robust control toolbox user’s
Normal University, Changsha, China, and the M.S. degree
guide. Natick, Mass: The Mathworks, Inc., pp. 134–157.
in 2006 from Central South University, Changsha, China.
Bose, N. K. (2003). Multidimensional systems theory and applications. The Nether-
From 2008 to 2010, she has been a Joint Cultivation
lands: Kluwer Academic Publishers.
Doctoral Candidate of Japan and China. She received her
Chen, T., & Francis, B. (1995). Optimal sampled-data control systems. London:
Springer. Ph.D. degree in control science and engineering from
Chen, J. W., & Liu, T. S. (2005). H∞ repetitive control for pickup head flying height in Central South University in 2011. She is an Associate
near-field optical disk drives. IEEE Transactions on Magnetics, 41(2), 1067–1069. Professor of control theory and control engineering with
Crudele, M., & Kurfess, T. R. (2003). Implementation of a fast tool servo with the School of Information and Electrical Engineering,
repetitive control for diamond turning. Mechatronics, 13(3), 243–257. Hunan University of Science and Technology, Xiangtan,
Doh, T. Y., & Chung, M. J. (2003). Repetitive control design for linear systems with China. Her current research interests include robust
time-varying uncertainties. IEE Proceedings of Control Theory and Applications, control, repetitive control, and control application.
150(4), 427–432.
Doh, T. Y., Ryoo, J. R., & Chung, M. J. (2006). Design of a repetitive controller:
an application to the track-following servo system of optical disk drives. IEE
Proceedings of Control Theory and Applications, 153(3), 323–330.
Galkowski, K., Paszke, W., Rogers, E., Xu, S., Lam, J., & Owens, D. H. (2003). Stability Min Wu received the B.S. and M.S. degrees in engineering
and control of differential linear repetitive processes using an LMI setting. IEEE from Central South University, Changsha, China, in 1983
Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 50(9), and 1986, respectively, and the Ph.D. degree in engineer-
662–666. ing from Tokyo Institute of Technology, Tokyo, Japan, in
Galkowski, K., Paszke, W., Sulikowski, B., Rogers, E., & Owens, D. H. (2002). LMI based 1999. Since July 1986, he has been a Faculty Member with
stability analysis and controller design for a class of 2D continuous–discrete Central South University, where he is currently a Professor
linear systems. In Proceedings of the 2002 American control conference. Vol. 1 (pp. of automatic control engineering with the School of Infor-
29–34). mation Science and Engineering. He was a Visiting Scholar
Hara, S. (1986). Repetitive control. Journal of the Society of Instrument and Control at the Department of Electrical Engineering, Tohoku Uni-
Engineers, 25(12), 1111–1119 (in Japanese). versity, Sendai, Japan, from 1989 to 1990, a Visiting Re-
Hara, S., Yamamoto, Y., Omata, T., & Nakano, M. (1988). Repetitive control system: search Scholar at the Department of Control and Systems
a new type servo system for periodic exogenous signals. IEEE Transactions on Engineering, Tokyo Institute of Technology, Tokyo, Japan, from 1996 to 1999, and
Automatic Control, 33(7), 659–668. a visiting professor at the School of Mechanical, Materials, Manufacturing Engi-
Hladowski, L., Galkowski, K., Cai, Z., Rogers, E., Freeman, C. T., & Lewin, P. L. (2010). neering and Management, University of Nottingham, Nottingham, UK, from 2001 to
Experimentally supported 2D systems based iterative learning control law 2002. He received the control engineering practice paper prize of the International
design for error convergence and performance. Control Engineering Practice, Federation of Automatic Control (IFAC) in 1999 (jointly with M. Nakano and J. She).
18(4), 339–348. His current research interests include robust control and its application, process
Hladowski, L., Rogers, E., Galkowski, K., & Virendra, R. S. (2008). On control laws control, and intelligent control. He is a member of the Nonferrous Metals Society of
for discrete linear repetitive processes with dynamic boundary conditions. China and the China Association of Automation.
Multidimensional Systems and Signal Processing, 19, 477–488.
Khargonekar, P. P., Petersen, I. R., & Zhou, K. (1990). Robust stabilization of
uncertain linear systems: quadratic stabilizability and H∞ control theory. IEEE
Transactions on Automatic Control, 35(3), 356–361.
Merkin, D. R. (1997). Introduction to the theory of stability. New York: Springer. Jie Zhang received his Bachelor’s degree of engineer-
Roncero-Sanchez, P., Acha, E., Ortega-Calderon, J. E., Feliu, V., & Garcia-Cerrada, A. ing from Central South University (CSU), Changsha, China
(2009). A versatile control scheme for a dynamic voltage restorer for power-
in 2008. He is currently pursuing his doctor’s degree in
quality improvement. IEEE Transactions on Power Delivery, 24(1), 277–284.
Graduate School of Environment and Energy Engineering,
She, J., Fang, M., Ohyama, Y., Hashimoto, H., & Wu, M. (2008). Improving
Waseda University, Japan.
disturbance-rejection performance based on an equivalent-input-disturbance
His research interests include robust repetitive con-
approach. IEEE Transactions on Industrial Electronics, 55(1), 380–389.
trol, power system, and robotics.
She, J.-H., Wu, M., Lan, Y.-H., & He, Y. (2010). Simultaneous optimization of the low-
pass filter and state-feedback controller in a robust repetitive-control system.
IET Control Theory and Applications, Institution of Engineering and Technology,
4(8), 1366–1376.
Wu, M., Zhou, L., & She, J. (2011). Design of observer-based H∞ robust repetitive-
control system. IEEE Transactions on Automatic Control, 56(6).
Wu, M., Zhou, L., She, J., & He, Y. (2010). Design of robust output-feedback
repetitive controller for class of linear systems with uncertainties. Science China:
Information Sciences, 53(5), 1006–1015.
Yamamoto, Y. (1994). A function space approach to sampled data control systems
and tracking problems. IEEE Transactions on Automatic Control, 39(4), 703–713. Yong He received his B.S. and M.S. degrees in applied
mathematics from Central South University, Changsha,
China in 1991 and 1994, respectively. In July 1994, he
Jinhua She received a B.S. degree in engineering from joined the staff of the university, where he is currently
Central South University, Changsha, Hunan, China, in 1983, a professor in the School of Information Science and
and an M.S. degree in 1990 and a Ph.D. degree in 1993 Engineering. He received his Ph.D. degree in control theory
in engineering from the Tokyo Institute of Technology, and control engineering from Central South University
Tokyo, Japan. In 1993, he joined the Department of in 2004. From January 2005 to March 2006, he was
Mechatronics, School of Engineering, Tokyo University a research fellow in the Department of Electrical and
of Technology, and in April, 2008, he transferred to the Computer Engineering, National University of Singapore.
University’s School of Computer Science, where he is From March 2006 to January 2007, he was a research
currently a professor. He received the control engineering fellow in the Faculty of Advanced Technology, University of Glamorgan, United
practice paper prize of the International Federation of Kingdom. His current research interests are time-delay systems, networked control
Automatic Control (IFAC) in 1999 (jointly with M. Wu systems, etc.

You might also like