Sunteți pe pagina 1din 8
Seediscussions,stats,andauthorprofilesforthispublicationat: A Filtered


Article in IEEETransactionsonAutomaticControl·February2011





CITATIONS 7 2authors ,including: QuanQuan BeihangUniversity(BUAA) 71 PUBLICATIONS 260








A Filtered Repetitive Controller for a Class of Nonlinear Systems

Quan Quan and Kai-Yuan Cai

Abstract— In this note, a filtered repetitive controller (FRC) is designed to compensate for periodic disturbances in a class of nonlinear systems. First, a new model of periodic disturbances is proposed. By this model, the FRC is designed and the resulting closed-loop error dynamics are analyzed with the help of a Lyapunov-Krasovskii functional. Finally the method is applied to periodic disturbance rejection in a class of robotic manipulators. Compared with repetitive controllers, the FRC provides the flexibility to choose filter parameters to achieve a tradeoff between tracking performance and stability. More importantly, FRC can deal with small input delay while the corresponding RC cannot.

Index Terms— Repetitive control, nonlinear systems, robotic manipulator, additive decomposition.


Repetitive control (RC) is an internal-model-based control

approach in which the infinite-dimensional internal model


sT gives rise to an infinite number of poles on the imagi- nary axis. It was proved in [1] that, for a class of general linear plants, exponential stability of RC systems could be achieved only when the plant is proper but not strictly proper. Moreover,


sT may destabilize the system. To

the internal model

enhance stability, a suitable filter is introduced as shown in

Fig. 1, forming a filtered repetitive controller 1 (FRC, or filtered repetitive control, also designated FRC) in which the loop gain

is reduced at high frequencies. Stability results only with some

sacrifice of high frequency performance. With appropriate design, however, an FRC can often achieve an acceptable tradeoff between tracking performance and stability, a tradeoff which broadens the application of RC in practice.

In fact, RC can also be considered as a special case of FRC with an all-pass filter. In the past two decades or more, FRC for linear time-invariant systems has reached maturity [2]–[4]. There has been little research, however, on RC, let alone FRC,

for nonlinear systems. This is the initial objective of this note.

A linear RC system is a neutral type system in a critical case

[5],[6]. The characteristic equation of the neutral type system has an infinite sequence of roots with negative real parts

approaching zero, i.e. sup{Re s |F (s) = 0 } = 0 where F (s)

is the characteristic equation. This implies that a sufficiently

small uncertainty may lead to sup{Re s |F (s) = 0 } > 0.

It is proved that a linear RC system will lose its stability

when subject to an input delay no matter how small the delay is (see Appendix A). Therefore, the stability of RC



This work was supported by the 973 Program (2010CB327904), the National Natural Science Foundation of China (60904066), and the Innovation Foundation of BUAA for PhD Graduates. The authors are with National Key Laboratory of Science and Tech- nology on Holistic Control, Department of Automatic Control, Bei- jing University of Aeronautics and Astronautics, Beijing 100191, China (qq;

1 In this note we have replaced the term “modified” in [1] with the more descriptive term “filtered”.



1 1 e sT e sT U s s + + B s
1 e
U s
B s

(a) Repetitive Controller

1 1 Q s e sT Q s e sT s + U s +
1 Q s e
Q s e
U s
B s

Q s : a suitable filter

(b) Filtered Repetitive Controller

Fig. 1. A suitable filter Q(s) is introduced into a repetitive controller to form an FRC

systems is insufficiently robust. On the other hand, with a

low-pass filter, an FRC system is a retarded type system. For

a stable FRC system, we have sup{Re s |F (s) = 0 } < 0.

Continuity arguments would suggest that for a sufficiently small uncertainty, the FRC system is still stable. Therefore,

it is expected that, with a low-pass filter, a nonlinear FRC

system with a small input delay is still stable. This is our second objective.

To achieve the two objectives we must analyze the tracking performance of a closed-loop system composed of an FRC and a nonlinear plant. The theory of FRC proposed in [1]

is derived in the frequency domain and can be applied only

with difficulty, if at all, to nonlinear systems. For this reason, tracking performance needs to be analyzed in the time domain. The contributions of this note are: 1) a method to design

and analyze an FRC for a class of nonlinear systems; 2) the ability of our FRC to cope with small input delay; and, 3) a tradeoff achieved by tuning filter parameters between tracking performance and stability. We use the following notation. R n is Euclidean space of dimension n and R + denotes the space of nonnegative reals in R. · denotes the Euclidean vector norm or induced matrix norm. C ([T, 0] ; R n ) denotes the space of contin- uous n-dimensional vector functions on [T, 0] . x t c

θ[T,0] x (t + θ) , where

[T, 0] . C PT ([0, ) ; R m ) is the space of nth-order contin- uously differentiable functions g : [0, ) R m which are T -periodic, i.e. g (t + T ) = g (t) . λ min (X) and λ max (X) denote the minimum and maximum eigenvalues respectively of a positive semidefinite matrix X. I n is the identity matrix with dimension n.


x t

x t (θ)


x (t + θ) ,





A. Problem Formulation

To illustrate the generality of FRC, we consider the follow- ing error dynamics examined in [7],[8]:

e˙ (t) = f (t, e (t)) + b (t, e (t)) [v (t) vˆ (t)] .



Here e (t) R n is a tracking error vector, v (t) R m is a disturbance, vˆ (t) R m is a signal designed to compensate for v (t) . For simplicity, we set the initial time t 0 = 0. Throughout this note, we assume the following for the system (1). Assumption 1[7],[8]: The functions f : [0, ) × R n R n and b : [0, ) × R n R n×m are bounded when e (t) is bounded on R + . Moreover, there exists a differentiable function V 0 : [0, )×R n [0, ) , a positive definite matrix M (t) = M T (t) R n×n with 0 < λ M I n < M (t) and a matrix R (t) R n×m such that

V 0 (t,e (t))

c 0 e (t) 2




V 0 (t,e (t)) ≤ −e T (t) M (t) e (t)


+ e T (t) R (t) [v (t) vˆ (t)]



where c 0 > 0. Assumption








C PT ([0, ) ; R m ).

Under Assumptions 1-2, our objective is to design a single

FRC with the following two properties: 1) with certain filter

t e (t) = 0; 2) with another set of appropriate

filter parameters, for any value of e (0), e (t) is uniformly ultimately bounded (for the definition see Definition 1 below), and the FRC can cope with small input delay.

parameters, lim

B. A New Model of Periodic Signals

A new model to describe periodic signals is proposed. In

order to show the difference, the usual model of periodic signals is given first. Any v ∈ C PT ([0, ) ; R m ) can be

generated by the model [1]





x (t T)




= x (t)


(θ) = v

(T + θ) , θ [T, 0]


where t R + and x (t) R m is the state. By the internal model principle [9], it is expected that asymptotic rejection of

a periodic disturbance can be achieved by incorporating the

above model (4), i.e.

into the closed-loop system. This is also the basic idea of RC


new model to describe v ∈ C PT ([0, ) ; R m ), which

will help in design of the FRC for the nonlinear system (1),

is given in Lemma 1.

Lemma 1: If A 0, > 0 and A 1, < 1, then for any v ∈ C PT ([0, ) ; R m ) there exists a signal δ ∈ C PT ([0, ) ; R m ) such that

sT I m (the transfer function of (4)),







A 0, x˙ p (t)


x p (t) + A 1, x p (t T) + δ (t)

v (t)


x p (t) + δ (t)

x p (θ) = v (T + θ) , θ [T, 0]



where A 0, , A 1, R m×m . If A 0,



and A 1,


I m ,


then any v ∈ C PT ([0, ) ; R m ) can be generated by (5) with δ (t) 0.

Proof: See Appendix B.

Example 1: Let A 0, = I m and A 1, = (1 α) I m . Then (5) becomes

x˙ p (t)


x p (t) + (1 α) x p (t T) + δ (t)

v (t)


x p (t) + δ (t)


The system (6) is a retarded system driven by a bounded external signal δ (t) , t [0, ). We examine the bound on δ (t) by the frequency response method. By (6), the Laplace transform from v (s) to δ (s) is defined by

δ (s) = G (s) v (s)

where δ (s) and v (s) are the Laplace transforms of δ (t) and

v (t), respectively; here G (s) =

) . The

x p (θ) = v (T + θ) , θ [T, 0] .

1+ s(1 α) exp(sT ) 2+ s(1 α) exp(sT

frequency responses of G (s) with T = 2π and = 0.01, 0.001 are shown in Fig.2.

T = 2 π and = 0 . 01 , 0 . 001 are shown in

Fig. 2.

Frequency responses of G (s)

As shown in Fig.2, G (s) has a comb shape with notches matching the frequencies of the periodic signal v (t) , t

[0, ). This makes the frequency response of G (s) close to 0 at k, k = 0, 1, 2, · · · . As the parameter decreases, the periodic components, especially in the low frequency band,

will be attenuated by G (s) more strongly. Since the low frequency band is dominant in most periodic signals, this will result in a smaller upper bound on δ (t) . In particular, if = 0, then δ (t) 0 by Lemma 1.


A. Controller Design

According to (5), to compensate for the disturbance v (t) , the FRC is designed as


with vˆ (θ) = 0, for θ [T, 0] , where t R + , A 0, , A 1, R m×m , k e R + , and R (t) is as in (3). The FRC (7) can be considered not only as a repetitive controller, but also as a “modified” repetitive controller as described in [1].

A 0,

vˆ(t) ˙ = vˆ(t) + A 1, vˆ(t T) + k e R T (t) e (t)



In particular, the controller (7) with A 0, = 0 and A 1, = I m reduces to


Therefore, the controller (7) with A 0, = 0 is a repetitive controller.



k e R T (t) e (t) ,

vˆ(t) = vˆ(t T) + k e R T





vˆ (s)


(t) e (t)


r e (s)

2) Denote

vˆ (t)

Then the

Laplace transform

from r e (s)


vˆ (s)


defined by

vˆ(s) = 1 Q (s) e sT 1 B (s) r e (s) .


Here Q (s) = (sA 0, + I) 1 A 1, is called the Q-filter

in [1] and B (s) = (sA 0, + I) 1 . This controller is therefore a “modified repetitive controller” in the sense of [1].

Subtracting (7) from (5) yields


A 0,

v˜(t) = v˜(t) + A 1, v˜(t T)


where v˜ x p v.ˆ In (10), the initial condition on v˜ is bounded. We do not concern ourselves with the concrete value of the initial condition as the following results hold globally. Combining (1) and (10) yields a closed-loop error dynamics in the form

Ez˙ (t) = f a (t, z (t)) + f d (z (t T )) + b a (t, z (t)) σ (t) .



k e R T (t) e (t) + δ (t)

z =

f a (t, z (t)) =

f d (z (t T ))


E =

b a (t, z (t)) =



R m+n , σ =


0 n×1

R m+n ,

(t, e (t)) + b (t, e (t)) v˜ (t) R m+n ,

v˜(t) k e R T (t) e (t)


A 1, v˜(t T)


A 0, 0 n×m

I m

0 n×m

R m+n ,


0 m×n

I n

0 m×m

b (t, e

R (m+n)×(m+n) ,

(t)) R (m+n)×2m .

Remark 1: The closed-loop error dynamics (11) is a retarded type system if A 0, is nonsingular; whereas it will become a neutral type system if A 0, = 0.

B. Stability Analysis

The closed-loop error dynamics (11) will be analyzed with the help of a Lyapunov-Krasovskii functional, with the results stated in Theorem 1 below. The following preliminary result is needed.

Definition 1[10]: A solution z (t, z t 0 ) of (11) with z t 0 ∈ C ([T, 0] ; R n ) is said to be ultimately bounded with bound

, uniformly with respect to t 0, if for each δ

exists T

when z t 0 c < δ, then z (t, z t 0 ) for all t t 0 + T .

> 0 there 0 such that,

= T ( , δ)

> 0 independent of t 0

Lemma 2[10]: Assume that there exists a continuously differentiable functional V (t, z t ) : [0, )×C ([T, 0] ; R n ) [0, ) such that

γ 1 z (t) 2


V (t, z t ) γ 2 z (t) 2 + tT z (s) 2 ds


where γ 1 , γ 2 are positive real numbers. If there exists b R + such that


V ˙ (t, z t ) ≤ −γ 3 z (t) 2 + b

where γ 3 is a positive real number, then the solutions of (11) are uniformly ultimately bounded with respect to the


γ 3 (γ 2 + T ), where µ is an arbitrarily small

positive real number.

(1 + µ)




With the help of Lemma 2, we have

Theorem 1: For the error dynamics (1), suppose Assump-





and if V 0 (t,e) in (2) further satisfies V 0 (t,e)

with c 1

ultimately bounded. Proof: For (11), choose the nonnegative function V (t, z t ) to be

> 0, then, for any value of e (0), e (t) is uniformly

c 1 e (t) 2

which the parameters satisfy k e

and A 1,

t e (t) =

tions 1-2 hold, and

= I αA 0, .





vˆ (t) is generated

We claim

A 0,





that 1)



the FRC (7)

A 0,


= A T


A 0,

= 0,

ensures A 1,


V (t, z t ) = 2k e V 0 (t, e (t))


+ v˜ T (t) A 0, v˜(t) + tT v˜ T (s) v˜ (s) ds


where V 0 satisfies Assumption 2 and A 0, = A T 0, 0. Under

Assumption 2, taking the time derivative V (t, z t ) along (11)



V (t, z t ) ≤ −2k e e T (t) M (t) e (t)

+ v˜ T (t) Hv˜ (t) + g (t) + 2˜v T (t) δ (t)


H = I m + A T

1, A 1,

g (t) = v˜ T (t) A 1, T A 1, v˜ (t) + 2˜v T (t)

v˜ T (t T) v˜(t T) .

A 1, T v˜(t T)


Under Assumptions 1-2, using the fact that g (t) 0, we obtain

V (t, z t ) ≤ −2k e e T (t) M (t) e (t) + v˜ T (t) Hv˜(t)


Based on the results above, conclusions 1) and 2) will be proved in detail in the Appendix C.


+ 2 v˜ (t) δ (t) .


To show its effectiveness, we apply the proposed controller design to the tracking problem for a class of robotic manipu- lators. The latter are described by

D (q (t)) q¨(t) + C (q (t) , q˙ (t)) q˙ (t)

+ G (q (t)) + τ d (t) = τ (t)




q (t)

R n



vector of generalized coordi-

nates, τ (t)

R n


the vector of

input torques, τ d

C PT ([0, ) ; R n ) is the vector of T -periodic disturbances, D (q (t)) R n×n is the symmetric inertia matrix satisfying


C (q (t) , q˙ (t)) R n×n is the centripetal-Coriolis matrix, and

G (q (t)) R n is the gravity vector.

R + ,



λ D I n

D (q (t))


λ D I n




A. Error Dynamics

In this section, the tracking problem will first be converted into error dynamics as in (1). Define the filtered tracking error as in [11]:

e (t) = q˜(t) + q˜(t)

where q˜(t) = q d (t) q (t) and q d (t) is a desired trajectory. Our control design is based on the fact that the parameters appear linearly in the robot dynamics (17). For convenience of control synthesis, the linear-in-the-parameters property [11] is written as [12]–[14]:


D (q (t)) q¨ e (t) +

C (q

(t) , q˙ (t)) q˙ e (t)


G (q

(t)) = Ψ (q, q,˙ q˙ e , q¨ e , t) p



where q˙ e (t) = q˙ d (t)+˜q (t) and q¨ e (t) = q¨ d (t)+ q˜(t) , p R m

is the vector of constant parameters, and Ψ (q, q,˙ q˙ e , q¨ e , t) R n×m , denoted by Ψ (t) for brevity. Design τ (t) according to

τ (t) = M (t) e (t) + Ψ(t) pˆ(t) + τˆ d (t) . (19)

Here M (t) R n×n is a positive definite matrix, pˆ(t) is the designer’s estimate of p, and τˆ d (t) R n is the designer’s estimate of τ d (t) . By employing (18) and (19), the filtered error dynamics are obtained as follows

D (q (t)) e˙ (t) + C (q (t) , q˙ (t)) e (t)

= M (t) e (t) + Ψ (t) [p pˆ(t)] + [τ d (t) τˆ d (t)] .

Furthermore, the system above can be written in the form of (1) with



(t, e (t)) = (t, e (t)) = D 1 (q (t)) Ψ(t)

D 1 (q (t)) [C (q (t) , q˙ (t)) + M (t) e (t)]

I n

v (t) =





, vˆ(t) =


τˆ d




B. Verification of Assumptions

Define the positive definite function

V 0 (t,e (t)) =


2 e T (t) D (q (t)) e (t) .


c 0 e (t) 2 V 0 (t,e (t)) c 1 e (t) 2 ,

1 λ D are positive reals. By


time derivative of V 0 (t,e) along (1) with (20) is evaluated as (3) with R (t) = Ψ(t) I n . Therefore, Assumption 1 is satisfied. Since p is a vector of constant parameters and

skew-symmetry of matrix

where c 0



2 λ D

and c 1





D (q (t)) 2C (q (t) , q˙ (t)) ,

τ d (t) is the vector of T -periodic disturbances, v (t) in (20) satisfies v ∈ C PT ([0, ) ; R n+m ). Therefore, Assumption 2 is



C. Numerical Simulation

The robot, the initial condition and tracking task used

for this

disturbance is

example are the same as in [11]. We assume the

τ d (t) =

0.5 1 cos 2πt

0.5 sin 2πt t 0.5 cos 2πt 1




which is unknown for the controller design except for the fact that τ d ∈ C PT [0, ) ; R 3 . According to the dynamics in [11], we obtain the expressions:


p =

Ψ =

J p

q¨ e,1



l 2

l 3 T


q¨ e,2 + g





q¨ e,1 + q 3 q˙ 3 q˙ e,1

+q 3 q˙ 1 q˙ e,3

q¨ e,2 + g q¨ e,3 q 3 q˙ 1 q˙ e,1



m 2 /s and x j denote the jth element in the vector x, x = {q, q,˙ q˙ e , q¨ e } . The parameters J p , l 2 , l 3 are assumed unknown for the controller design.

Here J p = 0.8 kg/m 2 , l 2




l 3






In (19), choose M (t) to be M (t) 10I 3 and design vˆ (t)

as (7) with k e = 1, A 0, = I 3 and A 1, = (1 0.0001 ) I 3 . For tracking performance comparison, we introduce the per-

t[(k1)T,kT ] q˜(t) , where k =

1, 2, ··· . 1) Without Input delay: If = 0, then these parameters above satisfy the conditions of Theorem 1. As seen in Fig.3, the performance index J k approaches 0 as k increases. It implies that q˜(t) approaches 0 as t → ∞. This result is consistent with conclusion 1) in Theorem 1. If = 2, 1, 0.5, then the chosen parameters satisfy the conditions of Theorem 1. As seen in Fig.3, the performance index J k is bounded. This result is consistent with conclusion 2) in Theorem 1.

formance index as J k =


2) in Theorem 1 . formance index as J k = sup Fig. 3. Tracking performance

Fig. 3.

Tracking performance with different

As seen in Fig.3, the tracking performance improves as decreases. This is consistent with the conclusion for linear


systems that, recalling (9), as the bandwidth of Q (s) = I 3 increases, i.e. decreases, the tracking performance

improves, and vice versa [1],[4]. 2) With Input delay: Consider the robotic manipulator with an input delay as follows:



D (q (t)) q¨(t) + C (q (t) , q˙ (t)) q˙ (t)

+ G (q (t)) + τ d (t) = τ

(t d ε )

where d ε > 0 is the input delay. The controllers are designed as above. When RC is adopted ( = 0) and d ε = 0.001s, it is observed from Fig.4 that a state of the closed-loop system diverges at about 30s. This confirms that the stability of RC systems is insufficiently robust. Using the FRC with = 0.5 and prolonging the input delay from d ε = 0.001s to d ε = 0.01s, it is observed from Fig.4 that the closed-loop system can still track the desired trajectory. This shows that an FRC system is more robust than its corresponding RC system.

FRC system is more robust than its corresponding RC system. Fig. 4. dotted–desired trajectory] Time response

Fig. 4.

dotted–desired trajectory]

Time response of q d,3 (t) and q 3 (t) with input delay. [solid–response,

Remark 2: From the simulations, the FRC provides the flexibility to choose filter parameters to achieve a tradeoff between tracking performance and stability. More importantly, it is shown that FRC is effective because input delay exists widely in practice.



Compared with existing repetitive controllers, the FRC provides flexibility to choose different filter parameters such as A 0, to satisfy performance requirements and tolerate some small uncertainties, such as input delay. If a periodic distur- bance is present, then the proposed controller with A 0, = 0 causes the tracking error to approach 0 as t → ∞. In order to demonstrate its effectiveness, the proposed method is applied to periodic disturbance rejection in a class of robotic manipulators. Simulation data show that a tradeoff between tracking performance and stability can be achieved by tuning the filter parameters. Furthermore, simulation data show that

FRC can deal with small input delay while the corresponding RC cannot, which demonstrates the effectiveness of FRC.


A. RC Systems with an Input Delay

Consider the following simple RC system:




x (t) + u (t)


(t) = u (t T ) x (t)


where x (t) , u (t) written as

R. The RC system above can be also

x˙ (t) x (t T ) = 2x (t) + x (t T ) .


From [6], the system above is asymptotically stable. In the following, we claim that the RC system (21) with an input delay τ ε > 0, namely




x (t) + u (t τ ε )


(t) = u (t T ) x (t)


will lose its stability no matter how small τ ε is.

For simplicity, let τ ε =

2N T , where N is a positive integer.

The characteristic equation of (23) is

1 e Ts = e


2N s


s + 1 .


From [15, p. 28, Lemma 7.1], the roots of (24) will approach

j as k → ∞. Therefore, if k is sufficiently large, then a

root of (24) is

± 2


s = 2


j + α ε + β ε j

with sufficiently small α ε , β ε R. Setting k = 2N k + N, where k is a positive integer, and substituting s into (24) results in

1 e Tα ε (cos (T β ε ) j sin (T β ε ))

= e

2N α ε cos 2N β ε j sin




2N β ε



j + α ε + β ε j + 1


Setting k to be large enough, and then neglecting the higher order infinitesimal, we have

1 e Tα ε cos (T β ε ) e

2N α ε 1 sin


2N β ε 2






e Tα ε sin (T β ε ) ≈ − e

T 2N α ε




Furthermore, we can obtain

α ε

T (N + 1)

2N (2) 2 .

Therefore, we can conclude α ε > 0. This implies the RC system (21) with an input delay τ ε > 0 is unstable no matter how small τ ε is (or how large N is).


B. Proof of Lemma 1

Let (4) be the original system. To apply Additive Decom- position[16], choose the primary system as follows:

A 0, x˙ p (t)


x p (t) + A 1, x p (t T) + δ (t)

x p (θ)

= v (T + θ) , θ [T, 0]


where A 0, , A 1, R n×n , δ (t) R n . From the original system (4) and the primary system (25), the secondary system is determined by Additive Decomposition[16]:

A 0, x˙ p (t) =

x s (t) + x s (t T)

+ (I A 1, ) x p (t T) δ (t)

x s (θ) = 0, θ [T, 0] .


By Additive Decomposition[16], we have

x (t) = x