Академический Документы
Профессиональный Документы
Культура Документы
,
(0, 1]) of tracking error is employed. The basic idea
is explained in some detail together with a frequency
domain design method. For implementation, a direct
discretization technique is introduced. An example is
given to show the benets of the D
E
k
(s) (3)
where, when = 0, (3) is a P-type ILC and, when
= 1, it is a D-type ILC. Detailed literature reviews
on ILC research can be found in [3, 4]. Most of the
existing work has focused on the analysis issue of ILC
schemes. However, the convergence conditions found in
the literature are typically not sucient for actual ILC
applications. Therefore, in recent years increasing ef-
forts have been made on the design issue of ILC. These
can be observed from the latest books [5, 4] and the
dedicated ILC web server [6]. A recent survey on the
ILC design issue [7] has documented various practically
tested design schemes, mainly for robotic manipulators.
The focus of this paper is to go in-between P-type
and D-type ILC schemes, i.e., we consider in this pa-
per that in (3) is a real number between 0 and
1. We name this kind of generalized scheme D
-type
( (0, 1] ) ILC. For example, we can dene a D
0.5
-
type ILC updating law according to
u
k+1
(t) = u
k
(t) +
d
0.5
dt
0.5
e
k
(t). (4)
This kind of fractional order ILC scheme is actually
an application of fractional calculus which is a 300-
years-old topic. The theory of fractional-order deriva-
tive was developed mainly in the 19-th century. Recent
books [8, 9, 10, 11] provide a good source of references
on fractional calculus. However, applying fractional-
order calculus to dynamic systems control is a recent
topic of interest [12, 13, 14, 15, 16]. For example, PID
controllers, which dominate industrial control practice,
have been modied using the notion of fractional-order
integrators and dierentiators. It has been shown that
the extra degree of freedom from the use of fractional-
order integrators and dierentiators make it possible
to further improve the performance of traditional PID
controllers. For fractional-order systems, the fractional
controller CRONE [17] has been developed while in [18]
and [19, 20] PD
controller
were proposed respectively. For pioneering work on
this regard, we cite [21, 22].
In the existing literature, to our best knowledge,
there is no eort in developing fractional order deriv-
ative based ILC schemes. In this paper, we extended
the classical Arimoto D-type iterative learning control
(ILC) updating law to a fractional order ILC updat-
ing law D
-type ILC
scheme the learning convergence can be tuned to be
more monotonic in the presence of strong time-varying
nonlinearities.
2 Problem Formulation
2.1 Fractional Order Dierential Equation
(FODE)
Note: This introductory subsection is similar to that
found in a second paper by the authors in these pro-
ceedings
1
. The material is duplicated here to preserve
continuity in the development of the paper.
The fractional calculus is a generalization of integra-
tion and dierentiation to non-integer order operators
[8, 9, 10, 11]. The idea of fractional calculus has been
known since the development of the normal calculus,
with the rst reference probably being associated with
Leibniz and LHospital in 1695. A fundamental oper-
ator
a
D
t
, a generalization of dierential and integral
operators, is introduced as follows.
a
D
t
=
_
_
_
d
dt
() > 0,
1 () = 0,
_
t
a
(d)
() < 0.
There are two commonly-used denitions for gen-
eral fractional dierentiation and integral, i.e., the
Gr unwald denition and the Riemann-Liouville de-
nition [8, 10, 11]. The Gr unwald denition is that
a
D
t
f(t) = lim
h0
1
h
[
ta
h
]
j=0
(1)
j
_
j
_
f(t jh), (5)
where [] is a ooring-operator, while the Riemann-
Liouville denition is given by
a
D
t
f(t) =
1
(n )
d
n
dt
n
_
t
a
f()
(t )
n+1
d, (6)
for (n 1 < < n) and where (x) is the well known
Eulers gamma function.
The formula for the Laplace transform of the
Riemann-Liouville fractional derivative (6) has the
form [23]:
_
0
e
pt
0
D
t
f(t) dt =
= p
F(p)
n1
k=0
p
k
0
D
k1
t
f(t)
t=0
, (7)
1
Analytical stability bound for a class of delayed fractional-
order dynamic systems, YangQuan Chen and Kevin L. Moore,
in Proceedings of 2001 IEEE Conference on Decision and Con-
trol, Orlando, Florida, December 2001.
for (n 1 < n).
For numerical calculation of fractional-order deriva-
tion we can use the relation (8) derived from the
Gr unwald denition (5). This relation has the follow-
ing form:
(tL)
D
t
f(t) h
N(t)
j=0
b
j
f(t jh), (8)
where L is the memory length, h is the step size of
the calculation,
N(t) = min
__
t
h
_
,
_
L
h
__
, (9)
where b
j
is the binomial coecient given by the follow-
ing recursive formula:
b
0
= 1, b
j
=
_
1
1 +
j
_
b
j1
. (10)
To solve the fractional-order dierential equations
(FODE), the Laplacian transformation of the Mittag-
Leer function in two parameters was proposed as an
eective means [23]. A two-parameter function of the
Mittag-Leer type is dened by the series expansion:
E
,
(z) =
k=0
z
k
(k +)
, (, > 0). (11)
In fact, it is shown [23] that the Mittag-Leer function
is a generalization of exponential function e
z
, i.e.,
E
1,1
(z) =
k=0
z
k
(k + 1)
=
k=0
z
k
k!
= e
z
.
In general, a LTI fractional-order controlled system
can be described by
a
n
D
n
t
y(t) + +a
1
D
1
t
y(t) +a
0
D
0
t
y(t) = (12)
b
m
D
m
t
u(t) + +b
1
D
1
t
u(t) +b
0
D
0
t
u(t),
where u(t) and y(t) are control and controlled signals
respectively;
k
,
k
(k = 0, 1, 2, ) are generally real
numbers,
n
> >
1
>
0
,
m
> >
1
>
0
and a
k
, b
k
(k = 0, 1, ) are arbitrary constants. Note
that here (
0
D
t
D
t
).
Using the Laplacian transformation method [23, 24,
25], the transfer function of (12) can be written as
G
s
(j) =
Y (j)
U(j)
=
m
k=0
b
k
(j)
n
k=0
a
k
(j)
k
. (13)
2.2 D
-type ILC
As discussed in the introduction section, in-between P-
type and D-type ILC schemes, by using the fractional
order derivative notion mentioned in Sec. 2.1, the fol-
lowing learning updating law can be used:
u
k+1
(t) = u
k
(t) +
d
dt
e
k
(t) (14)
4452
with its frequency domain form given in (3).
Here we can regard
d
dt
. In practice,
we need to set the transitional frequency range much
larger than the [
L
,
H
], for example, [0.1
L
, 10
H
].
Using this transitional frequency range and a desired
order of approximation, 2N +1, the following formulae
are the so-called Oustaloup-Recursive-Approximation
(ORA):
lim
N
D
N
(s) = D(s) = s
, (15)
where
D
N
(s) = (
u
H
)
k=N
1 +s/
k
1 +s/
k
, (16)
and
u
=
L
, (17)
k
=
L
(
L
)
(k+N+0.50.5)/(2N+1)
, (18)
k
=
L
(
L
)
(k+N+0.5+0.5)/(2N+1)
. (19)
In the implementation of the fractional order deriva-
tive used in our simulations below, we use these ORA
formulae to approximate s
.
3 Convergence Analysis
In this section, we shall perform a convergence analysis
of the proposed D
(Y
d
(s) Y
k
(s))
= U
k
(s) + s
G
c
(s)(U
d
(s) U
k
(s))
= (1 s
G
c
(s))U
k
(s) + s
G
c
(s)U
d
(s). (21)
Let
= 1 s
G
c
(s). (22)
It is easy to write (21) as follows:
U
k+1
(s) =
k
U
k
(s) +
1
k
1
s
G
c
(s)U
d
(s).(23)
When
| |< 1, (24)
we can conclude that
lim
k
U
k
(s) = U
d
(s). (25)
The convergence condition is that
| 1 (j)
G
c
(j) |< 1, . (26)
Remark 3.1 The convergence condition (26) tells us
why a D-type ILC is normally required. The reason is
that the Bode plot of G
c
(j) normally has a low pass
characteristic. Of course, the ideal G
c
(j) of a feedback
controlled system is a constant 1. But this is not phys-
ically possible in practice. Also ideally, it is desirable
to replace the term (j)
G
c
(i) is better shaped, which in turn
facilitates the proper choice of .
Remark 3.3 The zero-phase low-pass lter plus a
pure phase lead is a key implementation skill in ILC
applications [7]. Following the above remarks, we can
see that the major objective is to dilute G
c
(j). This
idea is not to actively dilute G
c
(j) by using a high-
pass lter but simply cuto the high frequency part. The
phase lead element can play a role in compensating the
low pass characteristic of G
c
(j) but its eect is lim-
ited. Therefore, introducing a general s
to replace the
zero-phase low pass lter plus a pure phase lead makes
sense here.
4 Implementation Issues
In general, there are two discretization methods for s
:
direct discretization and indirect discretization. In in-
direct discretization methods [26, 27], as also simply
introduced in Sec. 2.2, two steps are required, i.e., fre-
quency domain tting in the continuous time domain
rst and then discretizing the t s-transfer function.
Other frequency-domain tting methods can also be
used but without guaranteeing the stable minimum-
phase discretization.
In this Section, we focus on the direct discretization
method using the well known Tustin operator. It is
a straightforward scheme to discretize the fractional-
order derivative. The discrete transfer function is sta-
ble and minimum phase. The recursive formula for
discretization with dierent orders of approximation
simplies the programming eorts. In general, the dis-
cretization of the fractional-order dierentiator s
(
is a real number) can be expressed by the so-called
generating function s = (z
1
). This generating func-
tion and its expansion determine both the form of
the approximation and the coecients [28]. For ex-
ample, when a backward dierence rule is used, i.e.,
(z
1
) = (1 z
1
)/T
s
, where T
s
is the sampling pe-
riod, performing the power series expansion (PSE) of
(1 z
1
)
=
_
2
Ts
1z
1
1+z
1
_
= (
2
T
s
)
_
1 z
1
1 +z
1
_
= (
2
T
s
)
lim
n
A
n
(z
1
, )
A
n
(z
1
, )
(27)
where A
0
(z
1
, ) = 1, and
A
n
(z
1
, ) = A
n1
(z
1
, ) +c
n
z
n
A
n1
(z, ), (28)
and
c
n
=
_
/n ; n is odd;
0 ; n is even.
(29)
Therefore,
s
(
2
T
s
)
A
n
(z
1
, )
A
n
(z
1
, )
.
5 An Illustrative Example
To demonstrate the benets from using the proposed
D
(t) =
1
J
(u(t) F(t)) +
1
J
(
1
2
m+M)gl sin (t) (30)
where (t) is the angular position of the manipulator;
u(t) is the applied joint torque; F(t) is the friction
torque; m, l are the mass and length of the manip-
ulator respectively, M is the mass of the tip load, g is
the gravitational acceleration, and J is the moment of
inertia w.r.t the joint and is given by J = Ml
2
+ml
2
/3.
The friction torque F(t) is
F(t) =
_
f
+
+B
+
,
0,
f
+B
,
< 0,
where the Coulomb friction f
+
= 8.43Nm and f
=
8.26Nm; the viscous friction coecients B
+
=
4.94Nm/rad/s and B
d
(t) =
b
+ (
b
f
)(15
4
6
5
10
3
) (31)
d
(t) = (
b
f
)(60
3
30
4
30
2
) (32)
where = t/(t
f
t
0
). In the simulation,
b
= 0
,
f
= 90
, t
0
= 0, and t
f
= 1. sec.
The RK-4 method is used to numerically inte-
grate the state equation with constant time step
T
s
= 0.01 second. The initial states at each ILC
repetition are all set to 0. The ILC iterations
end when e
b1
1
/sec. where e
b1
= sup
t[0,1]
|
d
(t) (t) |
and e
b2
= sup
t[0,1]
|
d
(t)
(t) |. In the ILC
updating law (14), the tracking error is dened by
e
k
(t) =
d
(t)
k
(t). Here we only assume that the
angular velocity is available. We also assume that
that accurate system dynamics are not known and
that the estimated
J is 50% of the true value in the
simulations. Then, for = 1, the best choice of
is J, which makes = 0 [4]. In the following, we x
4454
the learning gain = J/2 which implies that we have
assumed 50% uncertainties in parameter J around
its nominal value actually used in the simulations.
We consider ve cases, with =
0, 0.25, 0.5, 0.75, 1.0. For = 0, only the state
variables are used for ILC updating and this is thus
called P-type ILC. For = 1, the derivative of state
variable is to be used which involves the angular
acceleration and thus this is called D-type. In our
simulation, the acceleration is obtained by digital
dierentiation, which is a 5-point FIR lter given by
(k) (
(k2)8
(k1)+8
(k+1)
(k+2))/(12T
s
).
Using the Muir-recursive expansion of Tustin operator
in Sec. 4, for T
s
= 0.01 sec., given N=5, we have
s
0.25
3.761z
5
0.9402z
4
+ 0.09402z
3
0.3173z
2
+ 0.04701z 0.188
z
5
+ 0.25z
4
+ 0.025z
3
+ 0.08437z
2
+ 0.0125z + 0.05
s
0.5
14.14z
5
7.071z
4
+ 1.414z
3
2.475z
2
+ 0.7071z 1.414
z
5
+ 0.5z
4
+ 0.1z
3
+ 0.175z
2
+ 0.05z + 0.1
s
0.75
53.18z
5
39.89z
4
+ 11.97z
3
14.79z
2
+ 5.983z 7.977
z
5
+ 0.75z
4
+ 0.225z
3
+ 0.2781z
2
+ 0.1125z + 0.15
10 20 30 40 50 60
0
10
20
30
40
50
60
70
80
90
100
e
b
1
vs. ILC iter. no.
iteration number
e
b
1
(
d
e
g
)
=0
=0.25
=0.5
=0.75
=1
Figure 1: Comparison of D
-type
ILC is explained in the form of three remarks. For
implementation, a recursive direct discretization tech-
nique is introduced. An example is given to show the
benets from the D
-
type ILC design with those using the H
loop shaping
or other robust control techniques.
Along the iteration axis, we can consider the
fractional-order dierence. This may become a sys-
tematic way of designing high-order ILC.
Using the QFT (qualitative feedback control) tech-
niques [31, 32], we may propose frequency templates
for ILC design in frequency domain. We can also make
simultaneous design for feedback (re)design and feed-
forward (ILC) design.
We can also consider the case when is a complex
number.
4455
References
[1] S. Arimoto, S. Kawamura, and F. Miyazaki,
Bettering operation of robots by learning, J. of Ro-
botic Systems, vol. 1, no. 2, pp. 123140, 1984.
[2] T. Sugie and T. Ono, An iterative learning con-
trol law for dynamical systems, Automatica, vol. 27,
no. 4, pp. 729732, 1991.
[3] K. L. Moore, Iterative learning control - an
expository overview, Applied & Computational Con-
trols, Signal Processing, and Circuits, vol. 1, no. 1.
[4] YangQuan Chen and Changyun Wen, Iterative
Learning Control: Convergence, Robustness and Appli-
cations, vol. LNCIS-248 of Lecture Notes series on Con-
trol and Information Science, Springer-Verlag, London,
1999.
[5] Zeungnam Bien and Jian-Xin Xu, Iterative
Learning Control - Analysis, Design, Integration and
Applications, Kluwer Academic Publishers, 1998.
[6] YangQuan Chen, Dedicated web
server for Iterative Learning Control research,
http://cicserver.ee.nus.edu.sg/~ilc, Oct. 1998.
[7] Richard W. Longman, Designing Iterative
Learning and Repetitive Controllers. In Z. Bien and
Xu J.-X. eds, Iterative Learning Control - Analysis,
Design, Integration and Application, Kluwer Acad-
emic Publishers, pp. 107145, 1998.
[8] K. B. Oldham and J. Spanier, The Fractional
Calculus, Academic Press, New York, 1974.
[9] S. G. Samko, A. A. Kilbas, and O. I. Marichev,
Fractional integrals and derivatives and some of their
applications, Nauka i technika, Minsk, 1987.
[10] K. S. Miller and B. Ross, An Introduction to the
Fractional Calculus and Fractional Dierential Equa-
tions, Wiley, New York, 1993.
[11] I. Podlubny, Fractional Dierential Equations,
Academic Press, San Diego, 1999.
[12] Boris J. Lurie, Three-parameter tunable tilt-
integral-derivative (TID) controller, US Patent
US5371670, 1994.
[13] Igor Podlubny, Fractional-order systems and
PI