Beruflich Dokumente
Kultur Dokumente
P
`
P
`
P
``
and C"[C
C
`
],
S
M
"(I!P
`
(I!C
`
P
``
)C
M), (2)
G
"
"(P
#P
`
(I!C
`
P
``
)C
`
P
`
).
Using the formula (I!:[)"I#:([!:),
S
M
can be rewritten as
S
M
"I#P
`
(I!C
`
P
``
!C
MP
`
)C
M.
Note that G
"
would be the transfer function from w to e if
M were zero.
The remainder of this section is devoted to the design
of M. The design of C needs a lot more preparation and it
will be discussed in Section 6. The internal model should
be capable of generating signals very similar to w (precise
equality is not necessary, and in fact it is not possible in
our case). We assume that w is periodic with period t, so
that the Fourier expansion
w(t)"
I8
w
I
e'IJR (3)
1186 G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199
Fig. 2. (a) The structure of the internal model M, which consists of
q copies of a feedback connection of a delay line and a "lter W. Here
A(s)"exp(!t
, M will have poles very close to ikv for those k39 for
which kv3[c
'
, c
"
]. Thus, G will have zeros very close
to these points ikv, causing the steady-state part of e to be
small (as will be shown in Section 5).
Often c
'
"0, and then the simplest choice for W is
W(s)"1/(1#s). (6)
Denoting c
, (7)
where means &&much smaller''. But we must be careful
not to choose c
. If
W is a band-pass "lter, then the best choice we can think
of is t
(s), where W
(ic)
should be as close as possible to 1 for c4c
"
. This is in
order to get the poles of M as close as possible to ikv, as
explained earlier in this section. Since the Taylor expan-
sions of e'OO'Q and of W(s) are well known, we can
compute the "rst terms of the Taylor expansion of W
:
W
(s)"1#(t!t
!)s#O(s`).
Thus, a good way of getting W
"t!. (8)
G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199 1187
The same argument remains valid for any "lter W whose
Taylor expansion is W(s)"1!s#O(s`). Since by
Eq. (7) t, formula (8) can be understood as a small
correction subtracted from the value t used in previous
references. We shall see in the design example in Sec-
tion 7 that this correction actually helps to reduce the
steady-state error signi"cantly. The poles of M will be
located close to the imaginary axis, slightly to the left.
For c(c
"
, they will be very close to ikv. The location
of the poles can be imagined as in Fig. 2b, where a cross
denotes the actual placement of the poles of M and a dot
indicates the ideal placement corresponding to t
"t
and W"1.
3. Some background
In this section we recall some facts about regular linear
systems and their dynamic stabilization. For a more
detailed discussion on well-posedness and regularity we
refer to Weiss (1994a, b) and to Weiss and Curtain (1997).
We make the following convention: If a meromorphic
function is de"ned on some right half-plane and can be
extended meromorphically to a greater right half-plane,
we will not make any distinction between the initial
function and its extension. This will not lead to con-
fusions, since the extension, if it exists, is unique. Indeed,
an analytic function on a connected domain is uniquely
de"ned by its series expansion at one point.
For each :31, H`
?
denotes the space of bounded
and analytic functions on the right half-plane "
?
"
s3" Re(s)':. With the norm
]G]
&
`
?
"sup
Q
"
?
G(s), (9)
H`
?
is a Banach space. With our earlier convention, we
have that
H`
?
LH`
@
if :4[.
For :"0 we use the notation H` instead of H`
"
. By the
maximum modulus principle, the H` norm of a transfer
function G3H`, denoted ]G]
`
, can be interpreted as
the peak value in the Bode amplitude plot of G. The
space H`
?
has natural generalizations for vector-valued
and for matrix-valued functions. We shall use the nota-
tion H`
?
also for spaces of vector- and matrix-valued
functions, and we shall often indicate the range space in
parentheses. In formula (9), absolute values should now
be replaced by the natural vector and matrix norms
(Euclidean norm and greatest singular value).
De5nition 1. A "N
;
K-valued well-posed transfer function
is an element of one of the spaces H`
?
("N
;
K), for some
:31. Such a function G is called regular if the limit
D" lim
H>`
G(z) (z31)
exists. In this case, D3"N
;
K is called the feedthrough
matrix of G.
For example, any well-posed transfer function obtain-
able from rational functions and delays by "nitely many
algebraic operations is regular (this includes all the trans-
fer functions which arise in this work).
Let l be an LTI system with input space "K, state
space X and output space "N. We assume that X is
a Hilbert space. For a vector-valued function u de"ned
on [0, R), we denote by P
O
u its restriction to [0, t]. The
system l is called well-posed, if on any "nite time interval
[0, t], the operator from the initial state x(0) and the
input function P
O
u to the "nal state x(t) and the output
function P
O
y is bounded. Thus, we can write
x(t)
P
O
y
"
3
O
u
O
+
O
%
O
x(0)
P
O
u
,
where x(0)3X, P
O
u3`([0, t], "K), x(t)3X,
P
O
y3`([0, t], "N) and the 2;2 matrix consists of
bounded operators between the appropriate spaces. The
precise de"nition of a well-posed linear system is more
complicated, since the families of operators 3, u, +, %
must satisfy functional equations expressing causality
and time-invariance (one of these is Eq. (10) below). For
the details we refer to Salamon (1989) or Weiss (1994a).
We write l"(3, u, +, %).
We recall some facts about well-posed linear systems
which will be used later. First of all, 3 is a strongly
continuous semigroup of operators on X. The systeml is
called exponentially stable if the growth bound of 3 is
negative. For any t50, we denote by S
O
the right-shift
operator by t, acting on `
'
([0, R), "K). Similarly,
S*
O
will denote the left-shift operator on the same space. If
u, v3`
'
([0, R), "K), their t-concatenation is de"ned by
u
O
v"P
O
u#S
O
v.
The following functional equation holds for all t50:
u
O>R
(u
O
v)"3
R
u
O
u#u
R
v. (10)
There exist two linear operators +
`
: XP`
'
([0, R), "N) and %
`
: `
'
([0, R), "K)P`
'
([0, R), "N)
such that for every t50,
+
O
"P
O
+
`
, %
O
"P
O
%
`
.
For every [31, we denote by `
@
([0, R), "K) the
space of all functions of the form v(t)"e@Ru(t), where
u3`([0, R), "K). By de"nition, ]v]
*@
"]u]
*
. Let us
denote by : the growth bound of 3. Then for every [':,
we have that +
`
3L(X, `
@
([0, R), "K)) and
%
`
3L(`
@
([0, R), "K), `
@
([0, R), "N)). The operator
%
`
is shift-invariant, i.e., %
`
S
O
"S
O
%
`
for all t50.
This implies that %
`
can be represented by a well-
posed transfer function G in the following sense: If
1188 G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199
u3`
@
([0, R), "K) and y"%
`
u, then the Laplace trans-
forms of u and y are related by
yL(s)"G(s)uL(s) for Re(s) 'max:, [.
G is called the transfer function of l. We have
G3H`
@
("N
;
K) for all [':.
The system l is called regular if G is regular. Such
systems have a simple description via their generating
operators A, B, C, D, which are the analogues of the
matrices appearing in the usual representation of "nite-
dimensional linear systems. The operators A, B and C are
unbounded in general. A is the generator of the semi-
group 3. The following result has been proved by
Rebarber (1993): If a regular linear system is stabilizable,
detectable and its transfer function is in H`("N
;
K), then
the system is exponentially stable. For a precise de"ni-
tion of stabilizability and detectability in this framework
we refer to Rebarber (1993) and to Weiss and Curtain
(1997).
The simplest example of a regular linear system with
unbounded A, B and C is a delay line. Its transfer func-
tion is eOQ, with t
] x(t
)"0,
where H[0, t
]
consisting of absolutely continuous functions whose
derivative is in `[0, t
I
!C
P
I
"
(I#PC) !P(I#CP)
C(I#PC) (I#CP)
.
Fig. 3. The standard feedback connection of two well-posed linear
systems. We may think of P as the plant, C as the compensator, r as the
reference signal, d as the disturbance signal and e as the tracking error.
De5nition 2. With P, C and L as above,
(i) C is an admissible feedback transfer function for P if
L is well-posed,
(ii) C stabilizes P if L3H`,
(iii) C exponentially stabilizes P if L3H`
?
for some :(0.
We now recall a result from "nite-dimensional systems
theory. For matrix-valued rational functions, &&well-
posed'' is equivalent to &&proper''. Proper rational func-
tions are always regular. For the concept of pole-zero
cancellation and other details we refer to the survey
paper of Logemann (1993).
Proposition 3. Suppose that P and C are proper rational
matrix-valued transfer functions. hen C is an admissible
feedback transfer function for P if and only if
det(I#P(R)C(R))O0. C stabilizes P (equivalently,
C exponentially stabilizes P) if and only if
(I#PC)3H` and there are no unstable pole-zero can-
cellations in PC.
Now let l
P
and l
C
be well-posed linear systems with
transfer functions P and C. We assume that C is an
admissible feedback transfer function for P. Then the
feedback connection in Fig. 3 de"nes a new well-posed
linear system called the closed-loop systemcorresponding
to l
P
and l
C
.
The results of Rebarber mentioned earlier imply the
following theorem. For the proof, the reader may consult
Weiss and Curtain (1997, Section 4).
Theorem 4. et l
P
and l
C
be stabilizable and detectable
regular linear systems, with transfer functions P and C,
respectively. =e assume that C is an admissible feedback
transfer function for P. hen the corresponding closed-loop
system is exponentially stable if and only if C stabilizes P.
If l
P
and l
C
are exponentially stable, then the state-
ment of the above theorem becomes very simple. Indeed,
there is no need to mention stabilizability and detectabil-
ity, since they follow from stability. Moreover, C stabil-
izes P if and only if (I#PC)3H`, or equivalently
(I#CP)3H`.
G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199 1189
4. Stability and robustness
We consider two "nite-dimensional LTI systems
l
P
and l
C
with transfer functions P and C, partitioned as
in Section 2:
P"
P
`
P
`
P
``
and C"[C
C
`
].
The dimensions of P and C are (q#p);(r#m) and
m;(q#p). The regular linear system l
M
is obtained by
connecting q delay lines (realized as explained in Sec-
tion 3) into a feedback loop, as shown in Fig. 2a, where
W is the transfer function of a stable "nite-dimensional
LTI system l
W
and ]W]
`
41. The transfer function
M of l
M
is given in Eq. (4). We interconnect the systems
l
P
, l
C
and l
M
as shown in Fig. 1, obtaining a repetitive
control system with the general structure studied here.
We now state our main result about its stability and
w-stability.
Theorem 5. =ith l
P
, l
C
and l
M
as described above,
assume that the following conditions are satis,ed:
(i) he interconnection of l
P
and l
C
shown in Fig. 4a is
stable.
(ii) =e have ]S]
`
(1, where
S"W(I!P
`
(I!C
`
P
``
)C
). (11)
hen the repetitive control system shown in Fig. 1 is
exponentially stable.
If moreover the conditions
W(R)"0 and
P
`
(R)
P
``
(R)
) ]C(R)](1 (12)
hold, then the feedback system in Fig. 1 is w-stable, in
particular, it is robustly stable with respect to multi-delays.
For the de"nition of w-stability we refer to Georgiou
and Smith (1989, 1993). Roughly, it means that the feed-
back system remains stable if the high frequency behav-
ior of P, M and C is changed. For the de"nition to be
applicable, the blocks P and M have to be combined
into a single block. Alternatively, we could combine
C and M into a single block, and this view would lead
to the same result, as stated in Theorem 5. The concept
of w-stability is stronger than robustness with respect
to multi-delays, as de"ned in Logemann et al. (1996,
Section 6).
Note that the systemin Fig. 4a is "nite dimensional. Its
stability is of course independent of the "lter W. Note
also that S from Eq. (11) is the transfer function from a to
b in Fig. 4a. Like S
`
after Eq. (2), S can be rewritten as
S"W(I#P
`
(I!C
`
P
``
!C
P
`
)C
).
Fig. 4. (a) The interconnection of the plant l
"
(with transfer function
P) and the controller l
'
(with transfer function C) as in Fig. 1, but
without the internal model (M"1) and with a "lter W. The transfer
function from a to b is denoted by S. (b) A feedback system equivalent
to the one in Fig. 1 when no inputs are present, i.e., w"0. The
subsystem with transfer function S contains everything except the delay
lines, and it is shown in detail in Fig. 4a. Recall that the diagonal matrix
A(s)"ePQI represents the q delay lines.
The above theorem does not make any mention of the
signal w in Fig. 1 (its possible periodicity), nor of the error
e. It is concerned only with stability and w-stability. The
w-stability conditions (12) can sometimes be relaxed, see
Example 7. Note that all the conditions are independent
of the delay t
.
Proof. First we prove the exponential stability of the
feedback system. For this, we transform the control sys-
tem of Fig. 1 into the one shown in Fig. 4b, with no input
and no output. Here, the system l
S
with transfer function
S contains the plant, the compensator and the "lter, but
not the delay lines. As in Section 2, A(s)"eOQI. It is not
di$cult to see that l
S
is exactly the system shown in
Fig. 4a. By assumption (i) of Theorem 5 we have that
l
S
is stable. Its transfer function from a to b is S.
According to Theorem4 and the comments after it, the
system in Fig. 4b is exponentially stable if and only if
(I!AS)3H`. Since ]A]
`
"1, it follows from as-
sumption (ii) of Theorem 5 that this is indeed the case.
Now we assume that Eq. (12) holds as well. Then,
according to Georgiou and Smith (1993, Section 8), the
feedback system in Fig. 1 is w-stable if
M(R) 0
0 I
P
`
(R)
P
``
(R)
) ]C(R)](1.
Indeed, since W(R)"0, P, C and M all have limits as
sPRin "
"
, uniformly with respect to the argument of
s. Since M(R)"I, the above condition for w-stability
reduces to the second part of the conditions (12). )
Remark 6. As promised in Section 2, we now comment
on the di$culties which would result from choosing the
1190 G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199
"lter W"1. If W were allowed to be 1, then for most
plants the condition ]S]
`
(1 would become impossible.
Indeed, if W"1 and rank P
`
(s)(q for some s3"
"
, or
for s"R, then we see from Eq. (11) that ]S]
`
51.
Moreover, with W"1 the repetitive control system
would not be robustly stable with respect to delays, in the
sense of Logemann et al. (1996). This means that arbitrar-
ily small delays in the feedback loop (e.g., on the signal u)
would destabilize the system. This follows from the main
result of Logemann et al. (1996). In particular, the feed-
back system would not be w-stable. The lack of robust-
ness for in"nitely many poles of M on the imaginary axis
can be understood also in the context of the analysis of
Lee and Smith (1996, 1998), who use the gap metric
robustness margin to evaluate the system.
The problem now is to "nd a compensator C such
that the conditions (i) and (ii) of Theorem 5 are satis"ed.
We shall formulate this as a standard H` problem. This
class of problems has been extensively studied in the
recent robust control literature and good algorithms (and
programs) for their solution are available, see for
example Green and Limebeer (1995) or Zhou et al. (1996).
It may happen that the problem has no solution. In our
speci"c case, we may overcome this by choosing a less
restrictive weight function W, which means that the
frequency band [c
'
, c
"
] in which W(ic) is very close
to 1 gets narrower, and thus the steady-state error
may grow.
Example 7 (SISO control system). Now we restrict our-
selves to a plant with transfer function P
"
which is SISO,
and hence to a compensator with transfer function
C which is SISO as well. Thus, there is no additional
measurement information y, so that the closed-loop sys-
tem of Fig. 1 reduces to the one in Fig. 5. Apart from this,
our analysis is quite general, so that this discussion is not
really an example, it concerns a class of repetitive control
systems. Repetitive control systems as in Fig. 5 have
been analyzed in several earlier papers, and our notation
follows Weiss (1997).
To simplify the exposition, we assume that the plant
and the compensator are minimal (and hence we may
represent them by their transfer functions). We have
w"[r d]', where r is the reference and d is the dis-
turbance, so that here P"[I !P
"
!P
"
]. Thus,
P
"[I !P
"
], P
`
"!P
"
and P
`
and P
``
do not
exist. We have C
"C and C
`
does not exist. The condi-
tions (i) and (ii) of Theorem 5 become simpler. Indeed, the
stability of the feedback system from Fig. 4a required in
(i) is equivalent to the fact that C stabilizes P
"
. The
transfer function S from (11) is the weighted sensitivity:
S"W(1#P
"
C), and condition (ii) has now the
simpler form
]W(1#P
"
C)]
`
(1. (13)
Fig. 5. SISO repetitive control system, the classical structure. M is the
internal model, C is the stabilizing compensator and P
"
is the plant. The
reference r and the disturbance d are periodic with period t, and the
error e should be kept small.
The problem of "nding a stabilizing compensator C that
satis"es the condition (13) is called the weighted sensitiv-
ity H` problem.
The su$cient conditions on w-stability in Eq. (12) may
be relaxed to
W(R)(1 and P
"
(R) C(R)(1!W(R),
and they appear in this form in Weiss (1997). These
results have a partial MIMO extension in which P
"
and
C become MIMOsystems, but we still use the (somewhat
restrictive) block diagram of Fig. 5. The stability part of
Theorem 5 in this context (without w-stability) is due to
Hara et al. (1988) (see their Corollary 1).
5. Estimates for the error e
In this section we give some estimates of the size of the
error e in the repetitive control system of Fig. 1. We
decompose the error e into two parts, a steady-state and
a transient part, e"e
''
#e
''
, where e
''
is periodic and
e
''
3`
@
([0, R), "O) for some [(0. The reader should
recall the notation and the material on well-posed linear
systems from Section 3.
Lemma 8. If l"(3, u, +, %) is an exponentially stable
well-posed linear system and u3` ([0, R), "K), then
lim
R`
u
R
u"0.
Proof. For each t50, we split u into two parts,
u"u
R
v
R
,
where v
R
"S*
R
u. Then by the functional equation (10),
u
`R
u"u
`R
(u
R
v
R
)"3
R
u
R
u#u
R
v
R
.
Using the fact that if 3 is exponentially stable, then
]u
R
]4M for all t50, as proved, e.g., in Weiss (1989), we
get
]u
`R
u]4M]3
R
]]u]#M]v
R
].
Since ]3
R
]P0 and ]v
R
]P0, we obtain the claim to be
proved. )
G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199 1191
Lemma 9. et G3H`
?
("
p;m
) for some :(0. Assume
that the input function u3`
'
([0, R), "K) has the
decomposition
u"u
''
#u
''
, (14)
where u
''
is periodic with period t and u
''
3`
@
([0, R), "K)
with :4[40. If y is the corresponding output function,
yL"GuL, then y3`
'
([0, R), "N) has a similar decomposi-
tion:
y"y
''
#y
''
, (15)
where y
''
is periodic with period t and y
''
3`
@
([0, R), "N).
Moreover, if the vector-valued sequences (u
I
) and (y
I
) are
the Fourier coe.cients of u
''
and y
''
, i.e.,
u
''
"
I
9
u
I
e'IJR and y
''
"
I
9
y
I
e'IJR
with v"2/t (v is the fundamental frequency), then
y
I
"G(ikv)u
I
. (16)
If l is an exponentially stable well-posed linear system
with transfer function G, and if u from Eq. (14) is its input
function, then for any initial state, the state trajectory of
l approaches a limit cycle.
We remark that (u
I
) and (y
I
) are vector-valued l` se-
quences and the two Fourier series appearing in the
lemma converge in the `-sense on any "nite time inter-
val. Note that if u
''
"0, we can take [":, so that the
transient output signal y
''
is in `
?
([0, R), "N). This case
will be used later.
Proof. We know from Salamon (1989) (see also Sta!ans,
1999) that G has a realization l"(3, u, +, %) such that
the growth bound of the semigroup 3 is : and
+
`
3L(X, `
?
([0, R), "N)), (17)
where X is the state space of l. We show that the state
trajectory x corresponding to any initial state x(0) and
the input u from Eq. (14),
x(t)"3
R
x(0)#u
R
u (18)
approaches a limit cycle x
''
. More precisely, there exists
a continuous X-valued periodic function with period t,
denoted by x
''
, such that
lim
R`
]x(t)!x
''
(t)]"0. (19)
To prove Eq. (19) (i.e., the last part of the lemma), we do
not need Eq. (17), only the exponential stability of 3. The
function x
''
must satisfy
x
''
(t)"3
R
x
''
(0)#u
R
u
''
. (20)
Because of the periodicity (x
''
(t)"x
''
(0)), we must have
(I!3
O
) x
''
(0)"u
O
u
''
. (21)
Since 3 is exponentially stable, the spectral radius of 3
O
is
less than one. Thus, I!3
O
is invertible, and we can
de"ne the initial state of x
''
by
x
''
(0)"(I!3
O
)u
O
u
''
.
For all other t, x
''
is now de"ned by Eq. (20), and it is
easy to see that this function is indeed periodic. Using
Eqs. (18) and (20), we get
x(t)!x
''
(t)"3
R
x(0)#u
R
u!3
R
x
''
(0)!u
R
u
''
.
Since 3 is exponentially stable, Eq. (19) holds if
lim
R`
]u
R
u!u
R
u
''
]"lim
R`
]u
R
u
''
(t)]"0
and according to Lemma 8 this is the case.
Now we have to prove that also y can be decomposed
into a periodic part y
''
and a transient part y
''
. The
output function y is given by
y"+
`
x(0)#%
`
u.
We de"ne
y
''
"+
`
x
''
(0)#%
`
u
''
,
y
''
"+
`
[x(0)!x
''
(0)]#%
`
u
''
, (22)
so that clearly Eq. (15) holds. First we show that y
''
is
indeed periodic with period t. Recall that by S*
O
we
denote the left-shift operator by t. Then the functional
equations for + and % (see Weiss, 1994a) can be com-
bined into
S*
O
y"+
`
x(t)#%
`
S*
O
u.
This equation applied to x
''
and u
''
yields
S*
O
y
''
"+
`
x
''
(t)#%
`
S*
O
u
''
"+
`
x
''
(0)#%
`
u
''
"y
''
.
The fact that S*
O
y
''
"y
''
means that y
''
is periodic with
period t.
Now we turn to y
''
from Eq. (22). The "rst summand of
y
''
is in `
?
([0, R), "N), because of Eq. (17). Since :4[,
this implies that it is also in `
@
([0, R), "N). Since
G3H`
?
("N
;
K), we know that for every [5:, %
`
is
a bounded operator from `
@
([0, R), "K) to
`
@
([0, R), "N). Since u
''
3`
@
([0, R), "K), the second
summand of y
''
in Eq. (22) also in `
@
([0, R), "N). Thus
y
''
3`
@
([0, R), "N) and we have proved the "rst part of
the lemma.
According to the Fourier series expansion of u
''
, its
Laplace transform is
uL
''
(s)"
I
9
u
I
1
s!ikv
.
1192 G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199
Since (]u
I
])3l`, this series converges absolutely for all
s3" which are not of the form ikv. Thus, uL
''
has poles at
ikv for every k39, and u
I
is the corresponding residue,
i.e., u
I
"Res(uL
''
, ikv). Since uL
''
is analytic in "
?
(because
u
''
3`
?
([0, R), "K)), the residues of uL"uL
''
#uL
''
can be
computed: u
I
"Res(uL, ikv). Similarly, using Eq. (15),
y
I
"Res(yL, ikv). Since yL"GuL,
y
I
"Res(GuL, ikv)
"G(ikv) Res(uL, ikv)"G(ikv)u
I
. )
We remark that the Laplace transforms of y
''
and
y
''
can be computed using the last lemma:
yL
''
(s)"
I
9
G(ikv)
s!ikv
u
I
,
yL
''
(s)"
I
9
G(s)!G(ikv)
s!ikv
u
I
#G(s)uL
''
(s).
A decomposition similar to Eq. (15), although in a
di!erent context, has appeared in Langari and Francis
(1995, 1996).
We apply this lemma to the repetitive control
system in Fig. 1. We assume that the external signal
w3`
'
([0, R), "N) is periodic with period t, as in (3).
We show that the error signal e can be decomposed into
e
''
and e
''
and we give the formula for the computation of
the Fourier coe$cients of e
''
.
Proposition 10. =ith the notation of heorem 5, assume
that the conditions (i), (ii) and (12) are satis,ed. hen there
exists an :(0 such that the following properties hold:
(I) he transfer function G from w to e in the system from
Fig. 1, which was given in Eq. (1), satis,es
G3H`
?
("O
;
P).
(II) Suppose that w3`
'
([0, R), "P) is periodic with
period t, as in Eq. (3). hen the error e in the re-
petitive control system from Fig. 1, for any initial
state, can be decomposed as e"e
''
#e
''
, where
e
''
3`
?
([0, R), "O) and e
''
is periodic with period t.
Moreover, let (e
I
) be the sequence of Fourier coe.cients of
e
''
, similarly as in Eq. (3), and let v"2/t. hen for all
k39,
e
I
"G(ikv) w
I
. (23)
Proof. According to Theorem 5, the feedback system in
Fig. 1 is exponentially stable. Thus, the growth bound of
its semigroup 3 is negative: c
"
(3)(0. The statements in
the above proposition hold for any :'c
"
(3). Indeed,
property (I) holds since the growth bound of G is less
than the growth bound of 3.
Now we prove property (II). Since w3`
'
([0, R), "P)
is periodic with period t, it is clear that the transient part
of this input signal is zero. If the initial state of the system
is zero, then eL"GwL . According to Lemma 9 with :"[,
e can be decomposed as claimed in the proposition. If the
initial state is not zero, then it generates a signal which
belongs to `
?
([0, R), "O), so that it can be absorbed into
e
''
. Applying formula (16) to this particular situation, we
obtain Eq. (23). )
We want to emphasize the strength of the conclusion
e
''
3`
?
([0, R), "O) in the last proposition. This implies
that there is an M'0 such that
R>O
R
]e
''
(o)]` do4Me`?R for all t50.
If w is su$ciently smooth, then the above conclusion can
be strengthened:
Proposition 11. =ith the assumptions and the notation
of Proposition 10, if the derivative wR belongs to
`
'
([0, R), "P), then
lim
R`
e'?'Re
''
(t)"0.
Proof. We know from Proposition 10 that e"e
''
#e
''
,
with e
''
periodic and e
''
3`
?
([0, R), "O). If we
apply Proposition 10 to the derivatives, it follows that
also eR
''
3`
?
([0, R), "O). De"ne v(t)"e'?'Re
''
(t), so that
v(t) 3`([0, R), "O). Then from
vR(t)":e'?'Re
''
(t)#e'?'ReR
''
(t)
we see that vR 3`([0, R), "O). By Barbalat's Lemma,
v3` and vR 3` imply that lim
R`
v(t)"0, which is
what we had to prove. )
Lemma 12. he transfer function S
M
from Eq. (2) can be
written as
S
M
"(I!AW)(I!S
AW)S
, (24)
where
A(s)"eOQI and S
"(I!P
`
(I!C
`
P
``
)C
)
(this is like the transfer function S
M
but with the internal
model M"I).
Proof. Using that M"I!AW, we have the follow-
ing chain of equalities:
S
M
"M[M!P
`
(I!C
`
P
``
)C
]
"M[S
!AW]
"M[S
(I!S
AW)]
"M(I!S
AW)S
. )
G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199 1193
Proposition 10 and Lemma 12 enable us to derive an
estimate for the `-norm of e
''
over a period. Assume
that the conditions of Theorem 5 hold and w is limited to
a frequency band [c
'
, c
"
] (as in Section 2). We denote
"]WS
]
`
and
"
"]S
G
"
]
`
,
so that (1 by condition (ii) of Theorem 5, since
S"WS
. Note that S
G
"
is the transfer function from
w to e in Fig. 4a. We introduce the function W
(which is
usually not in H`) by
eOQW(s)"eOQW
(s),
as in Section 2. Our main result in this section is the
following:
Theorem 13. =ith the above assumptions and notation,
]e
''
]
*'" O'
4
"
1!
max
'IJ''S' S"'
1!W
(ikv) ) ]w]
*
`
'" O'
. (25)
This shows clearly the importance of having W
(ikv)
close to 1. Formula (8) for t
(s)) (I!eOQW(s) S
(s))S
(s) G
"
(s).
From the last two formulas we can easily obtain the
estimate (25), using that eO'IJ"1 and eO'IJ"1. )
Example 14. We now return to our example in Fig. 5,
and apply Proposition 10. The Fourier coe$cients of the
steady-state error e
''
can be computed using Eqs. (23)
and (24),
e
I
"
(1!W
(ikv))S
(ikv)
1!W
(ikv)S
(ikv)
[r
I
!P
"
(ikv)d
I
],
with S
"(1#P
"
C). Again, we see that e
I
is small if
1!W
and
j
`
are small numbers and they are not always needed, depending on
certain rank conditions.
We embed the plant l
P
into an augmented plant m l
P
,
which has an exogenous input wJ and an external output
zJ. The input wJ contains the input w of l
P
and two
additional inputs, v
and v
`
, so that wJ "[v
v
`
w]'.
The external output zJ contains the weighted output
b with bK"W(eL#vL
y#j
v
`
]'. The
transfer functions of l
P
and of l
P
are denoted by P and
by P , respectively.
The parameters j
and j
`
are small numbers, while
could be any number. The standard H` problem for
l
P
is to "nd a stabilizing compensator l
C
, to be con-
nected from yJ to u (as shown in Fig. 6), such that the
H` norm of T
XU
, the transfer function from wJ to zJ, is
smaller than a given bound. This problem might be
impossible to solve, or it might have in"nitely many
solutions, whose transfer function C can be parametrized
after solving two Riccati equations, see Green and
Limebeer (1995) or Zhou et al. (1996). By an iteration on
the bound used for the H` norm, we can obtain a com-
pensator for which ]T
X
`
U
`
]
`
is very close to the in"mum of
its possible values.
If (A
5
, B
5
, C
5
, D
5
) are the matrices of the "lter
l
W
and if (A, B
, B
`
, C
, C
`
, D
, D
`
, D
`
, D
``
) are the
matrices of the plant l
P
, then the augmented plant l
P
is
described by
xR "AIx#[BI
BI
`
]
wJ
u
,
zJ
yJ
"
CI
CI
`
x#
DI
DI
`
DI
`
DI
``
wJ
u
,
1194 G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199
where x consists of the states of l
P
and l
W
and
AI BI
BI
`
CI
DI
DI
`
CI
`
DI
`
DI
``
"
A 0 0 0 B
B
`
B
5
C
A
5
B
5
0 B
5
D
B
5
D
`
D
5
C
C
5
D
5
0 D
5
D
D
5
D
`
0 0 0 0 0 j
`
C
0 0 D
D
`
C
`
0 0 j
D
`
D
``
. (26)
The algorithm hinfopt of MATLAB ( The Math-
Works) which solves the standard H` problem needs
DI
`
to have full column rank, respectively DI
`
to have
full row rank. To ensure this, the additional blocks
j
and j
`
have been added. If the rank conditions for
DI
`
and DI
`
are already satis"ed without those blocks,
then of course we can omit them.
Minimizing the transfer function from wJ to zJ is not
the problem we really want to have solved. We want
a stabilizing compensator l
C
for which ]S]
`
(1 (these
are conditions (i) and (ii) of Theorem 5) and we want the
ratio
"
/(1!) to be as small as possible (because of the
estimate (25)). Note that S"WS
G
"
is the transfer function from w to e,
all in Fig. 6. The arti"cial parameter has been introduc-
ed in order to search for the best ratio
"
/(1!) by
varying . By solving the H` problem for various values
of , j
and j
`
, we are looking for the solution for which
"
/(1!) is positive and as small as possible.
In our design examples (one of which is presented
in the next section) we have usually considered j
and
j
`
temporarily "xed and made iterations over . Occa-
sionally, we have changed the parameters j
and j
`
by
a factor of 10 or 0.1. We had to watch not only the ratio
"
/(1!) but also the Bode plots of the compensator C,
to avoid compensators with too high gains at certain
frequencies, since these would be di$cult to implement.
7. A design example
In this section we apply the results of the previous
sections to the control of a DC-motor with "xed stator
"eld (e.g., a permanent magnet), whose movements (its
rotation angle) should track a periodic signal, without
being signi"cantly in#uenced by the periodic load torque.
For example, this motor might be driving (via some
gears) a joint in a robot arm, which has to perform
a periodic task. The equations which model the motor
have been taken from Ogata (1992, Chapter 5). By : we
denote the angle of the motor shaft, by D":R the angular
velocity, by M the load torque, by u the input voltage and
by i the armature current. Then
d
dt
:
D
i
"
0 1 0
0 0 I
(
0 !I
*
!0
*
:
D
i
#
0
!
(
0
M#
0
0
*
u.
We denote by r the reference angle and by e"r!:
the tracking error. For the sake of comparison, "rst we
consider that there is no additional measurement avail-
able. Then the plant P from Fig. 1 (with the signal y not
present) has the realization
xR "
0 1 0
0 0 I
(
0 !I
*
!0
*
x#
0 0
0 !
(
0 0
w#
0
0
*
u,
e"[!1 0 0]x#[1 0]w, (27)
where x"[: D i]' and w"[r M]'.
Assuming that we have the additional measurement
information D, the set of ordinary di!erential equations
remains the same, but we have a second output y of the
plant. Now the output of P is given by
e
y
"
!1 0 0
0 1 0
x#
1 0
0 0
w#
0
0
u. (28)
We consider the following numerical values: the resist-
ance R"0.2 D, the inductance "20 mH, the moment
of inertia of the rotor, with gears and joint
J"1.5;10` kg m` and the motor torque constant
(same as the back-electromagnetic force constant)
k"9;10` Vs/rad. For computing the compensator,
we choose the following low-pass "lter:
W(s)"
200
s#200
. (29)
First we have considered the plant with the output as
in Eq. (27) (no additional measurement information).
After some trial and error we have chosen j
`
"10`,
while j
"10`, j
`
"10` and "40. Now "0.48 and
"
/(1!)"3.16. The transfer function C"[C
C
`
] of
the corresponding compensator is given by
C
(s)"1.21;10"
(s#10)(s#7.27;10`)
(s#3.87;10")(s#200)(s#5.01;10`)
,
C
`
(s)"!1.85;10`
(s#990)(s#10)
(s#6.67;10`)(s#3.87;10")(s#5.01;10`)
.
Again we have very large factors in the expressions of
C
and C
`
. The Bode plot of C
are practically
equal to zero and so we eliminate the corresponding
factors from the expression of C
. Additionally, we divide
the large pole of C
by 500, so
that for relatively low frequencies, it remains practically
unchanged. Thus we obtain the approximation C'
for the
component C
C'
`
] de"ned by
C'
(s)"2.42;10`
(s#10)
(s#7.74;10`)(s#200)
,
C'
`
(s)"!7.4;10"
(s#10)(s#990)
s(s#1.33;10`)(s#7.74;10`)
.
The gain of C'
`
is obviously unbounded for very low
frequencies, because of the pole at zero, but for frequen-
cies above 2 rad/s it remains less than 35 dB. Thus the
reduced compensator is realistic for implementation. The
simulations have been carried out both with and without
compensator approximation, and the results have been
indistinguishable. Thus, using more sophisticated model-
reduction techniques for C would not make any notice-
able di!erence.
For the simulation, we assume that the reference signal
r is obtained by passing a rectangular signal of frequency
0.25 Hz and amplitude 20 rad through the low-pass "lter
F(s)"(1#0.25s)(1#0.04s). The disturbance signal
M is the superposition of three sinusoids:
M(t)"1.5sint#1sin5t#0.7sin6.5t.
Both r and M are plotted in Fig. 7. The fundamental
frequency of these signals is 0.25 Hz, and so t"4 s. To
choose t
"t"4 s,
we get the graphs from Fig. 10. We can see that at the
beginning, the result is similar to the simulation result
with the same compensator and with the correction in
the time-delay. But when we look at the peaks of the
steady-state error, we see that they are approximately 12
times higher than previously. In the ` norm their ratio
would be even larger.
8. Conclusions and directions for future research
We have presented a fairly general design procedure of
repetitive controllers for MIMO linear plants in continu-
ous time. Our framework allows the use of additional
measurement information from the plant to improve
performance. Another improvement is achieved by a
G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199 1197
correction of the delay time, making it slightly less than
the period of the external signals. The closed-loop system
is in"nite-dimensional, and we employ the theory of
regular linear systems. We can guarantee exponential
stability and w-stability of the feedback system under
fairly simple, checkable conditions, which do not depend
on the amount of delay. We have introduced the de-
composition of the error signal into a steady-state and
a transient part, and have derived estimates for the for-
mer. These estimates have been used to build the design
procedure.
An important issue for further research is to allow
external signals of slowly varying shape, including vari-
ations of the period. Variations of the shape with un-
changed period are automatically taken care of by the
repetitive control system, and the transient response dies
down at an exponential rate. On the other hand, a signal
of varying period (strictly speaking, this is a contradic-
tion in terms) necessitates a mechanism for tracking the
period, and various ideas for this are available. The most
widespread approach is to use a phase-locked-loop, and
for some other approaches we refer to HillerstroK m (1996),
Reinke (1994), Russell (1986) and Tsao and Qian (1993).
An entirely di!erent direction for further research is to
allow external signals which are superpositions of peri-
odic signals of various (arbitrary) periods. This area,
called multi-periodic repetitive control, has been explored
in Garimella and Srinivasan (1994) and in Weiss (1997)
and the authors will present a more detailed investigation
of this in the near future (see also the preliminary report
Weiss and HaK fele (1998)).
Another technical question which is worth thinking
about is how to minimize the expression
"
/(1!)
appearing in our estimate (25) of the steady-state error,
over all controllers which satisfy the conditions in
Theorem 5.
Note. This work was carried out at the University of
Exeter, UK, Whilst Martin HaK fele was there as an Eras-
mus eschange student studying towards the MEng (Eur)
degree under the supervision of George Weiss.
References
Amann, N. (1996). Optimal algorithms for iterative learning control.
Ph.D. thesis. School of Engineering, University of Exeter, UK.
Francis, B. A., & Wonham, W. M. (1975). The internal model principle
for linear multivariable regulators. Applied Mathematics and Optim-
ization, 2, 170}194.
Garimella, S. S., & Srinivasan, K. (1994). Application of repetitive
control to eccentricity compensation in rolling. Proceedings
of the American Control Conference, Baltimore, Maryland
(pp. 2904}2908).
Georgiou, T., & Smith, M. C. (1989). w-stability of feedback systems.
Systems and Control etters, 2, 271}277.
Georgiou, T., & Smith, M. C. (1993). Graphs, causality and stabili-
zability: Linear, shift-invariant systems on `[0, R). Mathematics
of Control, Signals, and Systems, 6, 195}223.
Green, M., & Limebeer, D. J. N. (1995). inear robust control. Informa-
tion and System Sciences Series. Englewood Cli!s, NJ: Prentice-
Hall.
Hara, S., Yamamoto, Y., Omata, T., & Nakano, M. (1988). Repeti-
tive control system: A new type servo system for periodic exo-
genous signals. IEEE ransactions on Automatic Control, 33,
659}668.
HillerstroK m, G. (1996). Adaptive suppression of vibrations * a repeti-
tive control approach. IEEE ransactions on Control Systems
echnology, 4, 72}78.
HillerstroK m, G., & Sternby, J. (1996). Robustness properties of repetitive
controllers. International Journal of Control, 65, 939}961.
HillerstroK m, G., & Walgama, K. (1996). Repetitive control theory and
applications * A survey. Proceedings of the IFAC 13th Triennial
World Congress, San Francisco, California.
Inoue, T., Nakano, M., & Iwai, S. (1981a). High accuracy control of
servomechanism for repeated contouring. Proceedings of the 10th
Annual Symposium on Incremental Motion Control, Systems and Devi-
ces (pp. 258}292).
Inoue, T., Nakano, M., Kubo, T., Matsumoto, S., & Baba, H. (1981b).
High accuracy control of a proton synchrotron magnet
power supply. Proceedings of the IFAC 8th World Congress
(pp. 216}221).
Langari, A., & Francis, B. (1995). Design of optimal sampled-data
repetitive control systems. Proceedings of the 33rd Allerton
Conference on Communication, Control and Computing, Monticello,
Illinois.
Langari, A., & Francis, B. (1996). Robustness analysis of sampled-data
repetitive control systems. Proceedings of the IFAC 13th Triennial
World Congress, San Francisco, California (pp. 19}24).
Lee, R. C. H., & Smith, M. C. (1996). Some remarks on repetitive
control systems. Proceedings of the Conference on Decision and
Control, Kobe, Japan.
Lee, R. C. H., & Smith, M. C. (1998). Robustness and trade-o!s in
repetitive control. Automatica, 34, 889}896.
Logemann, H. (1993). Stabilization and regulation of in"nite-dimen-
sional systems using coprime factorizations. In A. Bensoussan, R. F.
Curtain, & J. L. Lions (Eds.), Analysis and optimization of systems:
State and frequency domain approaches for in,nite-dimensional sys-
tems, Lecture Notes in Computer Science (Vol. 185, pp. 102}139).
Berlin: Springer.
Logemann, H., Rebarber, R., & Weiss, G. (1996). Conditions for robust-
ness and nonrobustness of the stability of feedback systems with
respect to small delays in the feedback loop. SIAM Journal of
Control and Optimizations, 34, 572}600.
Moore, K. L. (1993). Iterative learning control for deterministic systems.
Advances in Industrial Control. London: Springer.
Ogata, K. (1992). System dynamics (2nd ed.). Englewood Cli!s, NJ:
Prentice-Hall.
Rebarber, R. (1993). Conditions for the equivalence of internal and
external stability for distributed parameter systems. IEEE ransac-
tions on Automatic Control, 34, 994}998.
Reinke, R. D. (1994). Adaptive Regeln zum ernen und Repro-
duzieren periodischer Signale mit dynamischen Netzwerken. Ph.D.
thesis. Fachbereich Mathematik, UniversitaK t Kaiserslautern,
Germany.
Rogers, E., & Owens, D. H. (1992). Stability analysis for linear repetitive
processes. Lecture Notes in Computer Science (Vol. 175). Berlin:
Springer.
Russell, D. L. (1986). Frequency/period estimation and adaptive
rejection of periodic disturbances. SIAM Journal of Control and
Optimization, 24, 1276}1308.
Salamon, D. (1989). Realization theory in Hilbert space. Mathematical
Systems heory, 21, 147}164.
Sta!ans, O. J. (1999). Admissible factorizations of Hankel operators
induce well-posed linear systems. Systems and Control etters (to
appear).
1198 G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199
Tsao, T.-C., & Qian, Y.-X. (1993). An adaptive repetitive control scheme
for tracking periodic signals with unknown period. Proceedings
of the American Control Conference, San Francisco, California
(pp. 1736}1740).
Weiss, G. (1989). Admissibility of unbounded control operators. SIAM
Journal of Control and Optimization, 27, 527}545.
Weiss, G. (1994a). Transfer functions of regular linear systems, part I:
Characterizations of regularity. ransactions of the American
Mathematical Society, 342, 827}854.
Weiss, G. (1994b). Regular linear systems with feedback. Mathematics of
Control, Signals, and Systems, 7, 23}57.
Weiss, G. (1997). Repetitive control systems: Old and new ideas. In C.
Byrnes, B. Datta, D. Gilliam, & C. Martin (Eds.), Systems and
control in the twenty-,rst century (Vol. 22, pp. 389}404). PSCT.
Boston: BirkhaK user.
Weiss, G., & Curtain, R. F. (1997). Dynamic stabilization of regu-
lar linear systems. IEEE ransactions on Automatic Control, 42,
4}21.
Weiss, G., & HaK fele, M. (1998). The exponential stability of multi-
periodic repetitive control systems, Proceedings of the MNS
Symposium, Padova, Italy.
Yamamoto, Y. (1993). Learning control and related problems in in"-
nite-dimensional systems. In H. Trentelman, & J. Willems (Eds.),
Essays on control: Perspectives in the theory and its applications.
(pp. 191}222) Boston: BirkhaK user.
Zhou, K., Doyle, J. C., & Glover, K. (1996). Robust and optimal control.
Englewoods Cli!s, NJ: Prentice-Hall.
George Weiss received the Control Engin-
eering degree from the Polytechnic Insti-
tute of Bucharest, Romania, in 1981 and
the Ph.D. degree in Applied Mathematics
from the Weizmann Institute, Rehovot,
Israel, in 1989. He has been working at
the Weizmann Institute, at Ben-Gurion
University in Beer Sheva, Israel, at the
University of Exeter, UK, and currently he
is with Imperial College in London, UK.
His research interests are distributed
parameter systems, operator semigroups,
power electronics, repetitive control and sampled-data systems.
Martin HaK fele was born in Tettnang,
Germany in 1971. He received his M.Eng.
(Eur) degree in 1997 from the University
of Exeter. In the same year he received
his Diploma Degree in Mechanical Engin-
eering from the University of Stuttgart.
Since 1998 he is Research Assistant at
the Max Planck Institute for Dynamics of
Complex Mechanical Systems in Mag-
deburg and studying towards the Ph.D.
His research area there is modeling, con-
trol and optimization of polymerization
processes. He is also interested in H` control.
G. Weiss, M. HaKfele/Automatica 35 (1999) 1185}1199 1199