Professional Documents
Culture Documents
controller design
problem is considered for time-delay plants with nitely many unstable zeros and
innitely many unstable poles. It is shown that this problem is the dual version of
the same problem for the plants with nitely many unstable poles and innitely many
unstable zeros, that is solved by the so-called Skew-Toeplitz approach. The optimal
H
controller design problem is given for plants with nitely many unstable
poles or unstable zeros by using duality and the Skew-Toeplitz approach. Necessary
and sucient conditions on time-delay systems are determined for applicability of
the Skew-Toeplitz method to nd optimal H
controller by another H
controller
in the feedback loop. In another approach, when the optimal controller is unstable
(with innitely or nitely many unstable poles), two methods are given based on a
search algorithm to nd a stable suboptimal controller. In this approach, the main
ii
idea is to search for a free parameter which comes from the parameterization of sub-
optimal H
ish.
I am thankful to my early-Ph.D.friends, Tankut Acarman, Veysel Gazi, Mehmet
Onder Efe,
Umit O gras, O guz Da gc, Peng Yan and late-Ph.D friends, Cosku Kas-
nako glu, Alvaro E. Gil, Nicanor Quijano, Jorge Finke. Since it is dicult to mention
all the names, I am also thankful to CRL graduate students and faculty.
I want to thank my non-departmental friends Cezmi
Unal, Yelda Serina gao glu,
Gokhan Korkmaz, Sahika Vatan, Tansu Demirbilek, Y ucel and Derya Demirer for
spending their time with me.
v
Although a graduate study in abroad causes to lost many friendship in your home
country, I am thankful to K ursad Erg ut, Hatice Din c, Ulas Bars Sarsoy,
Ilke and
Controller Design . . . . . . . . . 4
1.3 Dissertation Outline . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2. On the Mixed Sensitivity Minimization for Systems with Innitely Many
Unstable Modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1 Main Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Summary of Results . . . . . . . . . . . . . . . . . . . . . . . . . . 14
ix
3. H
Controller Design . . . . . . . . . . . . . . . . 29
3.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.1 IF Plant Example . . . . . . . . . . . . . . . . . . . . . . . 33
3.3.2 FF Plant Example . . . . . . . . . . . . . . . . . . . . . . . 36
3.4 Summary of Results . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4. Stable H
Controllers . . . . . . . . . . . . . . . . . . . . . . 50
5.2 Stable suboptimal H
Controller Design . . . 70
6.1 Strong stabilization of MIMO systems . . . . . . . . . . . . . . . . 71
6.2 Stable H
controllers . . . . . . . . . . . . . . . . . . . . . 76
6.4 Summary of Results . . . . . . . . . . . . . . . . . . . . . . . . . . 79
x
7. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Appendices:
A. Skew-Toeplitz Approach for Mixed Sensitivity Problem . . . . . . . . . . 85
B. IF and FI Plant Calculations . . . . . . . . . . . . . . . . . . . . . . . . 88
B.1 Example: IF Plant Case . . . . . . . . . . . . . . . . . . . . . . . . 88
B.2 Example: FF Plant Case . . . . . . . . . . . . . . . . . . . . . . . . 89
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
xi
LIST OF FIGURES
Figure Page
3.1 Optimal H
. . . . . . . . . . . . . . . . . . . . . . . . . 64
5.2 Z(s) plot for right half plane . . . . . . . . . . . . . . . . . . . . . . . 64
5.3
min
versus n
2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.4 Magnitude plot of U(j) . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.5 u
controller . . . . . . . . . . . . . . . . 69
6.1 Standard Feedback System . . . . . . . . . . . . . . . . . . . . . . . . 71
6.2 Comparison for plant G
1
. . . . . . . . . . . . . . . . . . . . . . . . . 76
6.3 Comparison for plant G
2
. . . . . . . . . . . . . . . . . . . . . . . . . 76
xii
LIST OF TABLES
Table Page
6.1 Stable H
controllers are designed to achieve the closed loop stability (Robust Stabilization
Problem) and pre-specied performance level (Robust Performance Problem) [2].
Stable controller design for a given plant is called strong stabilization. It is de-
sirable to have a stable controller for practical and theoretical reasons, ([3, 4]). For
1
example it is well known that the unstable poles of the controller degrades the track-
ing performance and/or disturbance rejection capabilities of the closed loop system
([5, 6, 7]). Moreover, unstable controllers are highly sensitive and their response to
sensor-faults and plant uncertainties/nonlinearities are unpredictable ([8]). It is also
well-known, [3], that the strong stabilization problem has a close connection with
simultaneous stabilization of two or more plants by a single controller. Also, stable
controllers can be tested o-line to check some faults in the implementation and to
compare with the theoretical design specications. Therefore, it is desirable to use
strongly stabilizing controllers in the feedback loop whenever possible.
In this thesis, a fundamental tool in robust control, H
controllers for
time-delay and multi-input multi-output (MIMO) systems. Unless otherwise speci-
ed, the discussion will be primarily be conned to single-input single-output (SISO)
systems. For the rest of this Chapter, a brief overview of the literature on optimal
H
control
2
problems are solved by operator theoretic methods. In this thesis we will refer to one
of these methods known as the Skew-Teoplitz approach. In [2], a mixed sensitivity
minimization problem is converted into two-block, and then to an one-block H
problem after a series of factorizations and transformations, and the resulting Nehari
problem is solved using the Commutant Lifting Theorem. In [15] the overall algorithm
is signicantly simplied. The optimal H
controllers are
shown (i.e., innite dimensional part of the controller can be represented in nite
impulse response lter (FIR)). Optimal H
Controller Design
A necessary and sucient condition for the existence of a stable controller is the
parity interlacing property (p.i.p.), [25]. A plant satises the p.i.p. if the number of
poles of the plant (counted according to their McMillan degrees) between any pair
of real right half plane blocking zeros is even. There are several design methods for
constructing strongly stabilizing controllers, see e.g. [3, 25]. In general, the procedures
involve construction of a unit in H
performance cost minimization with stable controller has also been stud-
ied in the literature. For nite dimensional SISO plants, optimal stable H
con-
troller for weighted sensitivity minimization is given in [33]. Using the interpolation
conditions of sensitivity function and Nevanlinna-Pick approach, optimal stable H
controller is obtained. An explicit formula for this optimal controller is given in [34]
4
where certain parameters of the controller are found from the solution of a set of non-
linear equations. The SISO discrete-time version of the weighted sensitivity problem
is solved in [35] by using a convex integer programming approach. For the mixed
sensitivity problem, certain conditions on the weighting functions and the plant are
determined in [36] for stability of the optimal controller. In [37] a sucient condi-
tion is obtained for the synthesis of SISO nite dimensional suboptimal stable H
controller design methodologies will be given for time-delay systems as well as MIMO
systems.
1.3 Dissertation Outline
In Chapter 2, optimal H
controller is given for this plant. In Chapter 3, using this result and the Skew-Toeplitz
method, the optimal H
perfor-
mance level for dead time systems is given in Chapter 4. This approach is based on
stabilization of H
controller by another H
cost,
opt
, dened as
opt
= inf
C stabilizing P
_
_
_
_
_
W
1
(1 +PC)
1
W
2
PC(1 + PC)
1
__
_
_
_
, (2.2)
where W
1
and W
2
are given nite dimensional weights. Note that in the above
formulation, the plant has nitely many unstable modes, because M
d
(s) is nite
8
dimensional, whereas it may have innitely many zeros in M
n
(s). In this section, by
using duality, the mixed sensitivity minimization problem will be solved for plants
with nitely many right half plane zeros and innitely many unstable modes.
2.1 Main Result
Assume that the plant to be controlled has innitely many unstable modes, nitely
many right half plane zeros and no direct transmission delay. Then, its transfer func-
tion is in the form P =
N
M
, where M is inner and innite dimensional (it has innitely
many zeros in C
+
, that are unstable poles of P), N = N
i
N
o
with N
i
being inner nite
dimensional, and N
o
is the outer part of the plant, possibly innite dimensional. For
simplicity of the presentation we further assume that N
o
, N
1
o
H
.
To use the controller parameterization of Smith, [44], we rst solve for X, Y H
satisfying
NX + MY = 1 i.e. X(s) =
_
1 M(s)Y (s)
N
i
(s)
_
N
1
o
(s). (2.3)
Let z
1
, ..., z
n
be the zeros of N
i
(s) in C
+
, and again for simplicity assume that they
are distinct. Then, there are nitely many interpolation conditions on Y (s) for X(s)
to be stable, i.e.
Y (z
i
) =
1
M(z
i
)
i = 1, . . . , n.
Thus by Lagrange interpolation, we can nd a nite dimensional Y H
and innite
dimensional X H
= inf
QH
_
_
_
_
_
W
1
(Y NQ)
W
2
N(X + MQ)
__
_
_
_
(2.6)
where Y (s) N(s)Q(s) = 0, W
1
and W
2
are given nite dimensional (rational)
weights. From (2.3) equation, we have
_
_
_
_
_
W
1
Y W
1
NQ
W
2
N
_
1MY
N
_
+ W
2
MNQ
__
_
_
_
=
_
_
_
_
_
W
1
(Y N
i
(N
o
Q))
W
2
(1 M(Y N
i
(N
o
Q)))
__
_
_
_
.(2.7)
Thus, the H
opt
= inf
Q
1
H
and Y N
i
Q
1
=0
_
_
_
_
_
W
1
(Y N
i
Q
1
)
W
2
(1 M(Y N
i
Q
1
))
__
_
_
_
(2.8)
where Q
1
= N
o
Q, and note that W
1
(s), W
2
(s), N
i
(s), Y (s) are rational functions, and
M(s) is inner innite dimensional.
The problem dened in (2.8) has the same structure as the problem dealt in
Chapter 5 of the book by Foias,
Ozbay and Tannenbaum [2], where Skew-Toeplitz
approach has been used for computing H
, because
sb
k(s+a)
controller.
1) Dene the functions,
F
(s) =
_
s
a
+ b
s
_
,
1
2
2
for > 0
where a
=
_
1 +
2
2
, and b
=
_
(1
2
2
)
2
+
2
.
2) Calculate the minimum singular value of the matrix,
M
=
_
_
1 j
MF
(j
) j
MF
(j
)
1 a MF
(a) aMF
(a)
MF
(j
) j
MF
(j
) 1 j
MF
(a) aMF
(a) 1 a
_
_
for all values of (max{,
1+
2
2
},
1
) and MF
(s) = M(s)F
(s). The
optimal gamma value,
opt
, is the largest gamma which makes the matrix M
singular.
3) Find the eigenvector l = [l
10
, l
11
, l
20
, l
21
]
T
such that M
opt
l = 0.
4) The optimal H
opt
( + s)
,
k
f
=
_
kb
opt
l
11
opt
l
21
2
opt
2
_
,
K
2,FIR
(s) = A(s) + B(s)e
hs
,
k
f
+ A(s) =
k(s + a)c(s) + (s + b)d(s)
((1
2
opt
2
) + (
2
opt
2
)s
2
)(s a)
,
B(s) =
(s b)c(s) + k(s a)d(s)
((1
2
opt
2
) + (
2
opt
2
)s
2
)(s a)
where c(s) = (a
opt
+ b
opt
s)(l
11
s + l
10
) and d(s) =
opt
( s)(l
21
s + l
20
).
As a numerical example, if we choose the plant as P(s) =
2
(
s3
s+1
)
1+2
(
s3
s+1
)
e
0.5s
and
the weight functions as W
1
(s) = 0.5, W
2
(s) =
1+0.1s
0.4+s
, then the optimal H
cost is
opt
= 0.5584, and the corresponding controller is
C
opt
(s) =
_
0.558s + 0.223
2s + 3.725
_
(1.477 + K
2,FIR
(s))
where K
2,FIR
(s) is equal to
(2.081s
2
6.302s 0.826) (0.615s
3
0.768s
2
5.269s + 1.587)e
0.5s
(0.302s
3
0.905s
2
+ 0.950s 2.850)
and its impulse response is of nite duration, k
2,FIR
(t) = L
1
(K
2,FIR
(s)) can be
written as
_
0.27e
3t
+ 7.16 cos(1.77t) + 0.36 sin(1.77t) 2.037(t 0.5) 0 t 0.5
0 t > 0.5
.
2.3 Summary of Results
In this Chapter, we have considered H
control
techniques [2] for innite dimensional systems with nitely many right half plane
poles. An example from delay systems is given to illustrate the computational tech-
nique. This result is used in the next Chapter to design optimal H
controller for
general time-delay systems, including dead-time systems.
15
CHAPTER 3
H
n
i=1
R
i
(s)e
h
i
s
m
j=1
T
j
(s)e
j
s
.
In this Chapter, optimal H
P(s) =
m
n
(s)
N
o
(s)
m
d
(s)
(3.1)
where m
n
(s) is inner, innite dimensional and m
d
(s) is inner, nite dimensional and
N
o
(s) is outer, possibly innite dimensional. The optimal H
controller,
C
opt
, stabi-
lizes the feedback system and achieves the minimum H
cost,
opt
:
opt
=
_
_
_
_
_
W
1
(1 +
P
C
opt
)
1
W
2
P
C
opt
(1 +
P
C
opt
)
1
__
_
_
_
(3.2)
where
W
1
and
W
2
are nite dimensional weights for the mixed sensitivity minimization
problem.
17
In Chapter 2, the optimal H
P(s) =
m
d
(s)
N
o
(s)
m
n
(s)
(3.3)
where m
n
is inner, innite dimensional, m
d
(s) is nite dimensional, inner, and
N
o
(s)
is outer, possibly innite dimensional. For this dual problem, the optimal controller,
C
opt
, and minimum H
cost,
opt
, are found for the mixed sensitivity minimization
problem which can be dened as,
opt
=
_
_
_
_
_
W
1
(1 +
P
C
opt
)
1
W
2
P
C
opt
(1 +
P
C
opt
)
1
__
_
_
_
. (3.4)
Optimal controller designs for problems (3.2) and (3.4) are outlined in Appendix A.
In this Chapter, we will design optimal H
n
i=1
R
i
(s)e
h
i
s
m
j=1
T
j
(s)e
j
s
.
The assumptions on the plant is given in Section 3.1 as A.1 A.7. We apply the
design methods in [15] and Chapter 2 for this plant. Note that P is the generalized
version of time-delay systems for SISO case. We assume that P has nitely many
right half plane zeros or poles (assumption A.7 in Section 3.1). We give the necessary
and sucient conditions when this assumption is valid in Section 3.1.1. The special
structure of H
controllers is given in this section. Note that single delay plant with
nite dimensional plant is a special case of the plant P. By showing nite impulse
response lter structure of neutral and retarded time-delay systems, we eliminate
right half plane pole-zero cancellation in the controller. The resulting controller in
this structure is numerically stable and easy for implementation.
18
3.1 Preliminary Work
We consider the plant, P, which has a transfer function,
P(s) =
R(s)
T(s)
=
n
i=1
R
i
(s)e
h
i
s
m
j=1
T
j
(s)e
j
s
. (3.5)
satisfying the following assumptions,
A.1 R
i
and T
j
are nite dimensional, stable, proper transfer functions,
A.2 Time delays are nonnegative rational numbers shown as h
i
and
j
with h
i
< h
i+1
for i = 1, . . . , n 1 and
j
<
j+1
for j = 1, . . . , m1,
A.3 h
1
1
,
A.4 Relative degree of R
1
is smaller than or equal to that of R
i
, j = 2, . . . , n and
the relative degree of T
1
is smaller than or equal to that of T
i
, i = 2, . . . , m,
A.5 Relative degree of T
1
is smaller than or equal to that of R
1
,
A.6 P has no imaginary axis zeros or poles,
A.7 R or T has nitely many zeros in the right half plane.
These assumptions are not restrictive. It is always possible to obtain a representa-
tion such that A.1 A.2 are satised. The rst assumption results in stable transfer
functions in the controller which is desired for implementation purposes. The assump-
tion A.3 guarantees that the plant is causal. The numerator and denominator of the
plant is neutral or retarded time delay system by the assumption A.4. The plant is
proper transfer function if the assumption A.5 is valid. Note that the assumptions
A.1A.5 are necessary by denition of neutral and retarded time delay systems. The
19
condition A.6 is a technical assumption which can be removed, but it is assumed for
simplicity in derivation. The last condition comes from the limitation of standard the
skew-toeplitz approach [2] and its dual approach given in Chapter 2.
For example, consider the plant whose dynamical behavior is described in time
domain by the following set of equations:
x
1
(t) = x
1
(t 0.2) x
2
(t) + u(t) + 2u(t 0.4),
x
2
(t) = 5x
1
(t 0.5) 3u(t) + 2u(t 0.4),
y(t) = x
1
(t). (3.6)
Its transfer function is
P(s) =
s + 3 + 2(s 1)e
0.4s
s
2
+ se
0.2s
+ 5e
0.5s
, (3.7)
and it can be re-written in form of (3.5) as,
P =
R
T
=
R
1
e
h
1
s
+ R
2
e
h
2
s
T
1
+ T
2
e
2
s
+ T
3
e
3
s
(3.8)
where
R
1
=
s+3
(s+1)
2
, R
2
=
2(s1)
(s+1)
2
,
T
1
=
s
2
(s+1)
2
, T
2
=
s
(s+1)
2
, T
3
=
5
(s+1)
2
,
are stable proper nite dimensional transfer functions and the delays are
h
1
= 0, h
2
= 0.4,
1
= 0,
2
= 0.2,
3
= 0.5.
It is clear that A.1 A.3 are satised. The relative degree of R
1
is equal to that of
R
2
which is 1 and the relative degree of T
1
, which is 0, is smaller than that of T
2
and
T
3
, which are 1 and 2, respectively. Also, the relative degree of T
1
is smaller than
that of R
1
. Therefore assumptions A.4 A.5 hold. The plant P has no imaginary
20
axis zeros. By a simple computation, T has nitely many right half plane zeros at
0.4672 1.8890j, whereas R has innitely many unstable zeros converging to the
asymptote 1.7329 (5k + 2.5)j as k . In conclusion, this system satises all
the assumptions, A.1 A.7.
We dene the conjugate of R(s) =
n
i=1
R
i
(s)e
h
i
s
as
R(s) := e
h
n
s
R(s)M
C
(s)
where M
C
is inner, nite dimensional whose poles are poles of R. For example, we
can calculate the conjugate of R(s) in the previous example,
R(s) =
s + 3
(s + 1)
2
+
2(s 1)
(s + 1)
2
e
0.4s
where h
2
= 0.4 and M
C
(s) =
_
s1
s+1
_
2
. The conjugate of R(s) can be written as,
R(s) = e
h
2
s
R(s)M
C
(s),
=
_
2
(s + 1)
+
(s 3)
(s + 1)
2
e
0.4s
_
.
We will design an optimal H
n
i=1
R
i
(s)e
h
i
s
has (innitely) nitely many right half plane zeros if and
only if
R
G
(s) = 1 +
R
2
(s)
R
1
(s)
e
h
1
s
+ . . . +
R
n
(s)
R
1
(s)
e
h
n
s
21
has (innitely) nitely many unstable zeros respectively.
The following fact will be used later in this Chapter and in Chapter 5.
Claim: R
G
(s) has no unstable zeros with real part extending to innity.
Proof: Assume that s =
0
+j. For suciently large
o
, we can write the following
inequalities,
R
i
(
0
+ j)
R
1
(
0
+ j)
+
i
i = 2, . . . , n
because all R
i
(s) terms are nite dimensional. Following holds by the triangle in-
equality,
1
2
e
h
2
0
. . .
n
e
h
n
0
|R
G
(
0
+ j)| 1 +
+
2
e
h
2
0
+ . . . +
+
n
e
h
n
0
This inequality tells us that R
G
(s) has no unstable zeros with real part extending to
innity because lim
|R
G
( + j)| = 1. Therefore, R(s) has no unstable zeros
with real part extending to innity.
Lemma 3.1.1. Assume that R(s) =
n
i=1
R
i
(s)e
h
i
s
is a neutral or retarded time
delay system with no imaginary axis zeros and poles, then the system R has nitely
many right half plane zeros if and only if all the roots of the polynomial, 1+
2
r
h
2
h
1
+
+
n
r
h
n
h
1
has magnitude greater than 1 where
i
= lim
R
i
(j)R
1
1
(j) i = 2, . . . , n,
h
i
=
h
i
N
, N,
h
i
Z
+
, i = 1, . . . , n.
If the system R satises this condition, it is called as F-system where F is an abbre-
viation for nitely many unstable zeros.
Proof. Since the system is neutral or retarded time delay system, it may have innitely
many right half plane zeros extending to innity in imaginary part with xed positive
22
real part [46]. R has innitely many right half plane zeros if and only if for xed
o
> 0, lim
R(
o
+j) has innitely many unstable zeros. Dene r = e
(
o
+j)/N
,
then
lim
R(
o
+ j)
R
1
(
o
+ j)
= r
h
1
+
2
r
h
2
+ +
n
r
h
n
. (3.9)
Let r
0
is the root of (3.9) which satises |r
o
| > 1. Then the system R do not have
innitely many right half plane zeros, since
|r
o
| = e
o
/N
,
o
= N ln |r
o
| < 0
which is a contradiction. Therefore, if all the roots of the polynomial (3.9) has
magnitude less than 1, then the system, R, does not have innitely many unstable
zeros. Assume that there exist a root, r
o
, of (3.9) with magnitude smaller than 1.
Then, there are innitely many right half plane zeros of R converging to the asymptote
r
o,k
=
ln |r
o
|
N
jN(r
o
+ 2k) as k where k Z and r
o
is the phase of the
complex number r
o
. The lemma is also valid when delays are real numbers. If the
delays are rational numbers, the resulting function is polynomial which is easy to
check its roots. When the delays are real numbers, it will be dicult to nd the
roots of the function. Therefore, the delays are assumed as rational numbers for
convenience.
We dene R(s) as I-system (where I is an abbreviation for innitely many unstable
zeros) if
R is F-system. In order to nd whether R is an I-system, we can check the
23
magnitude of all the roots of the polynomial, 1 +
2
r
h
2
h
1
+ +
n
r
h
n
h
1
where
i
= lim
R
i
(j)R
1
1
(j) i = 2, . . . , n,
h
i
=
h
i
N
, N,
h
i
Z
+
, i = 1, . . . , n.
If the magnitude of all the roots is smaller than 1, then R is an I-system.
In this section, the plant in (3.5) has nitely many unstable zeros or poles. This
is guaranteed by assumption A.7. By Lemma 3.1.1, it is easier to check that the
assumption A.7 is valid by nding the magnitude of all roots of the given polynomial.
We can conclude that the assumption A.7 is satised if and only if either R or T is a
F-system (or both).
3.1.2 FIR Structure of Neutral and Retarded Time Delay
Systems
In this section, we will show special structure of neutral and retarded time delay
systems. This is a key lemma to be used in the next section for the proofs of main
results.
Lemma 3.1.2. Let R be as in Lemma 3.1.1 and M
R
be a nite dimensional system.
Dene S
+
z
be the set of common C
+
zeros of R and M
R
. Then
R
M
R
, can always be
decomposed as,
R(s)
M
R
(s)
= H
R
(s) +F
R
(s) (3.10)
where H
R
is an innite dimensional system which does not have any poles in S
+
z
and
F
R
is a nite impulse lter.
Proof. For clarity of the presentation we give the proof for the case where the entries
z
1
, z
2
, . . . , z
n
z
of S
+
z
are distinct. For the general case main idea is the same, except
24
that some of the expressions below may become more complicated (notation needs to
be modied to include the multiplicity of each common zero, and the partial fraction
expansion and inverse Laplace transformation needs to consider these multiplicities).
So, we assume R(z
k
) = M
R
(z
k
) = 0, k = 1, . . . , n
z
. We can rewrite the expression
R
M
R
as
R
M
R
=
n
i=1
R
i
M
R
e
h
i
s
.
We can decompose each term by partial fraction,
R
i
M
R
= H
i
+F
i
where the poles of F
i
are elements of S
+
z
and dene the terms H
R
and F
R
as
H
R
(s) =
n
i=1
H
i
(s)e
h
i
s
,
F
R
(s) =
n
i=1
F
i
(s)e
h
i
s
.
Note that each term, F
k
is strictly proper and F
R
(z
k
) is nite k = 1, . . . , n
z
. The
lemma ends if we can show that F
R
is a FIR lter. Inverse Laplace transform of F
R
can be written as
f
R
(t) =
n
z
k=1
_
n
i=1
Res{F
i
(s)}
s=z
k
e
z
k
(th
i
)
u
h
i
(t)
_
where Res(.) is the residue of the function and u
h
i
(t) is the unit step function delayed
by h
i
. The impulse response is equivalent to
f
R
(t) =
n
z
k=1
e
z
k
t
_
n
i=1
Res{F
i
(s)}
s=z
k
e
h
i
z
k
_
25
for t > h
n
. It is clear from partial fraction that Res{F
i
(s)}
s=z
k
= R
i
(z
k
) which
results f
R
(t) is FIR lter,
f
R
(t) =
n
z
k=1
e
z
k
t
_
n
i=1
R
i
(z
k
) e
h
i
z
k
_
,
=
n
z
k=1
e
z
k
t
R(z
k
) 0 for t > h
n
,
since z
k
is the zero of R, i.e., R(z
k
) = 0. Therefore, we can conclude that F
R
is a
FIR lter with support [0, h
n
].
Note that this decomposition eliminates right half plane pole-zero cancellation in
R
M
R
and brings into form which is easy for numerical implementation. Lemma 3.1.2
explains the FIR structure of H
controllers for
time-delay systems.
Assume that P is in the form of P(s) =
n
k=1
P
k
(s)e
h
k
s
where P
k
are stable,
proper, nite dimensional transfer functions. Also, P
0
is a bi-proper, nite dimen-
sional system. By partial fraction, we can write
P
k
P
0
as
P
k
(s)
P
0
(s)
= P
k,p
(s) + P
k,0
k = 1, . . . , n
where the P
k,0
is a proper transfer function whose poles are same as the zeros of P
0
.
Then, the decomposition operator, , can be dened as,
(P(s), P
0
(s)) = H
P
(s) +F
P
(s)
where H
P
=
n
k=1
P
k,p
(s)e
h
k
s
and F
P
=
n
k=1
P
k,0
(s)e
h
k
s
are innite dimensional
systems. Note that if the zeros of P
0
are also unstable zeros of P, then F
P
is an FIR
lter shown in Lemma 3.1.2.
26
3.2 Main Results
In this section, we will construct the optimal H
1
)s
M
R
(s)
{e
h
1
s
R(s)}
R(s)
,
m
d
= M
T
(s),
N
o
=
R(s)
M
R
(s)
{e
1
s
T(s)}
M
T
(s)
(3.11)
where M
R
is an inner function whose zeros are the right half plane zeros of
R(s).
Since R is an I-system, the conjugate of R has nitely many right half plane zeros,
27
therefore M
R
is well dened. Similarly, zeros of M
T
are right half plane zeros of T.
Note that m
n
and m
d
are inner functions, nite and innite dimensional respectively.
N
o
is an outer term. Although there are right half plane pole-zero cancellations in
m
d
and
N
o
, this problem will be solved in Section (3.2.2) by using the technique in
Section 3.1.2.
FI Plant Factorization
The plant satises the assumptions A.1 A.7, R is F-system and T is I-system.
Since the design method in Chapter 2 requires the plant and its inverse to be causal
proper transfer functions, we restrict P in (3.5) to satisfy the following additional
assumptions:
B.3 h
1
=
1
= 0,
B.5 Relative degree of T
1
is equal to that of R
1
.
Then the plant P can be factorized as in (3.3),
m
n
= M
T
(s)
T(s)
T(s)
,
m
d
= M
R
(s),
N
o
=
R(s)
M
R
(s)
T(s)
M
T
(s)
. (3.12)
The zeros of inner function M
R
are right half plane zeros of R. The unstable zeros
of
T(s) are the same as the zeros of inner function M
T
. Similar to previous section,
conjugate of T has nitely many right half plane zeros since T is a I-system. The right
half plane pole-zero cancellations in m
n
and
N
o
will be eliminated in Section 3.2.2.
28
FF Plant Factorization
For FF plants, assumptions for IF plants are valid (i.e., assumptions A.1 A.7),
but R and T are F-systems. The plant P has a factorization as in (3.1),
m
n
= e
(h
1
1
)s
M
R
(s),
m
d
= M
T
(s),
N
o
=
{e
h
1
s
R(s)}
M
R
(s)
{e
1
s
T(s)}
M
T
(s)
(3.13)
where M
R
and M
T
are inner functions whose zeros are right half plane zeros of R and
T respectively. Note that when h
1
=
1
, m
n
is nite dimensional term. Then exact
internal pole-zero cancellations are possible for this case (except the ones in
N
o
).
3.2.2 Optimal H
Controller Design
By using the methods in [15] and Chapter 2, given the appropriately chosen weight
functions W
1
and W
2
, the optimal H
cost,
opt
can be found. After
opt
is calculated,
it is easy to obtain E
opt
, F
opt
and L. We will give the structure of optimal H
C
opt
=
K
opt
(s)
_
e
1
s
T(s)
M
T
(s)
_
R(s)
M
R
(s)
+ e
1
s
R(s)F
opt
(s)L(s)
(3.14)
where K
opt
(s) = E
opt
(s)F
opt
(s)M
T
(s)L(s). In order to obtain the structure of
controller:
1. Do the necessary cancellations in K
opt
,
29
(H
opt
1) +F
opt
H
T
+F
T
- - - d
6
+
1
s
T
T
, M
T
),
H
R
1
+F
R
1
= (
R, M
R
opt
),
H
R
2
+F
R
2
= (e
1
s
RF
opt
L,
opt
),
as explained in Section 3.1.2.
Then, the optimal controller has the form
C
opt
=
H
T
+F
T
H
opt
+F
opt
where H
T
, H
opt
= H
R
1
+ H
R
2
are neutral/retarded time-delay systems and F
T
,
F
opt
= F
R
1
+ F
R
2
are FIR lters. The controller has no right half plane pole-zero
cancellations and its structure is shown in Figure 3.1.
30
(H
R
1) +F
R
H
opt
+F
opt
- - - d
6
+
C
1
opt
=
K
opt
(s)
_
R(s)
M
R
(s)
_
T(s)
M
T
(s)
+ T(s)F
opt
(s)L(s)
(3.15)
where K
opt
(s) = E
opt
(s)F
opt
(s)M
R
(s)L(s). We can obtain the optimal controller
following exactly same steps for IF plant case (note that R and T are interchanged,
and h
1
=
1
= 0 from assumptions):
C
opt
=
H
opt
+F
opt
H
R
+F
R
.
The optimal controller is the dual of the controller for the IF plants, as expected.
There is no internal right half plane pole-zero cancellations and its structure is shown
in Figure 3.2.
31
Controller Structure of FF Plants
Structure of FF plants is similar to that of IF plants. We can calculate
opt
, E
opt
,
F
opt
, L by the method in [15], Chapter 2 and then write the optimal controller as:
C
opt
=
K
opt
(s)
{e
1
T(s)}
M
T
(s)
{e
h
1
s
R(s)}
M
R
(s)
+ e
1
s
R(s)F
opt
(s)L(s)
(3.16)
where K
opt
(s) = E
opt
(s)F
opt
(s)M
T
(s)L(s). We can nd controller structure follow-
ing similar steps as in IF plants case:
1. Do the cancellations in K
opt
,
2. Factorize K
opt
and nd
opt
,
T
,
3. Obtain (H
T
, F
T
), (H
R
1
, F
R
1
)and (H
R
2
, F
R
2
) by
H
T
+F
T
= (e
1
s
T
T
, M
T
),
H
R
1
+F
R
1
= (e
h
1
s
R, M
R
opt
),
H
R
2
+F
R
2
= (e
1
s
RF
opt
L,
opt
).
The controller structure can be written as:
C
opt
=
H
T
+F
T
H
opt
+F
opt
where H
opt
= H
R
1
+ H
R
2
and F
opt
= F
R
1
+ F
R
2
. Note that the resulting structure
is same as IF plant case. It is possible to cancel the zeros of
opt
with denominator
when h
1
=
1
= 0 which is considered in Example 3.3.2.
32
3.3 Examples
We will show the optimal H
R(s)
,
m
d
= M
T
(s),
N
o
=
R(s)
M
R
(s)
T(s)
M
T
(s)
33
where
M
R
=
s 0.2470
s + 0.2470
,
M
T
=
s
2
0.9344s + 3.7866
s
2
+ 0.9344s + 3.7866
,
R =
2(s + 1) + (s 3)e
0.4s
(s + 1)
2
.
For the given plant and weight functions, the optimal H
cost (3.2) is
opt
= 0.7203
and the corresponding optimal controller is [2] (see A.3 at Appendix A),
C
opt
= E
opt
(s) m
d
(s)
N
1
o
(s)F
opt
(s)L(s)
1 + m
n
(s)F
opt
(s)L(s)
where
E
opt
=
3.4812 + 47.8803s
2
0.5188(1 100s
2
)
,
F
opt
=
0.5188(1 10s)
1.384s
2
+ 2.842s + 1.381
,
L =
0.589s
2
0.3564s + 0.1616
0.589s
2
+ 0.3564s + 0.1616
.
The optimal controller,
C
opt
, has internal unstable pole-zero cancellations (i.e., the
zeros of E
opt
and m
d
are cancelled by the denominator of the controller). In order
to eliminate this, we will follow the procedure given in Section 3.2.2:
steps 1-2 After cancellation, K
opt
= E
opt
F
opt
M
T
L can be factorized as
K
opt
=
opt
(s)
T
(s),
where
opt
=
3.4812 + 47.8803s
2
(1.384s
2
+ 2.842s + 1.381)
M
T
(s),
T
=
L(s)
(1 + 10s)
.
34
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
Impulse response of F
T
(s)
time
Figure 3.3: f
T
(t) for IF plant
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.005
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
Impulse response of F
(s)
time
Figure 3.4: f
opt
(t) for IF plant
step 3 We can nd (H
T
, F
T
), (H
R,1
,F
R
1
) and (H
R
2
, F
R
2
) by the partitioning opera-
tor, . The expressions can be found in Appendix B (see equations (B.1-B.6)).
Then, the optimal controller can be found from,
C
opt
=
H
T
+F
T
H
opt
+F
opt
where H
opt
= H
R
1
+H
R
2
and F
opt
= F
R
1
+F
R
2
. The impulse responses of F
opt
and
F
T
are shown in Figures 3.3 and 3.4.
35
3.3.2 FF Plant Example
We will nd the optimal H
N
o
=
R(s)
M
R
(s)
T(s)
M
T
(s)
(3.18)
where
M
R
=
s
2
2.757s + 4.455
s
2
+ 2.757s + 4.455
,
M
T
=
s
3
4.09s
2
+ 8.185s 9.128
s
3
+ 4.09s
2
+ 8.185s + 9.128
.
36
Given the weight functions of the previous example, the optimal H
cost is computed
as
opt
= 0.7031. The optimal controller is as in (A.3) where
E
opt
=
3.5056 + 45.4406s
2
0.4944(1 100s
2
)
,
F
opt
=
0.4944(1 10s)
1.3482s
2
+ 2.7681s + 1.3446
,
L =
0.1364s
3
+ 0.3852s
2
+ 0.5654s + 0.1154
0.1364s
3
0.3852s
2
+ 0.5654s 0.1154
.
Since m
n
is nite dimensional, 1+ m
n
(s)F
opt
(s)L(s) is nite dimensional. Therefore,
the exact cancellation between
opt
and 1 + m
n
(s)F
opt
(s)L(s) is possible. We will
follow the procedure of Section 3.2.2 with a slight modication:
steps 1-2 After the cancellation, K
opt
= E
opt
F
opt
M
T
L can be factorized as
K
opt
=
opt
(s)
T
(s),
where
opt
=
3.5056 + 45.4406s
2
1.3482s
2
+ 2.7681s + 1.3446
M
T
(s),
T
=
L(s)
(1 + 10s)
.
Then, the controller can be written as,
C
opt
=
_
T
T
M
T
_
R
M
R
_
1+M
R
F
opt
L
opt
_
step 3 We perform the cancellations in
K
=
_
1+M
R
F
opt
L
opt
_
as shown in Appendix B
(see equations (B.7-B.8)),
step 4 We can nd the controller parameters (H
T
, F
T
) from
H
T
+F
T
= (T
T
, M
T
)
37
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
Impulse response of F
T
(s)
time
Figure 3.5: f
T
(t) for FF plant
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
Impulse response of F
(s)
time
Figure 3.6: f
opt
(t) for FF plant
and (H
opt
,F
opt
) from
H
opt
+F
opt
= (R
K
, M
R
).
See Appendix B for these partitions (equations (B.9-B.12)). In Figure 3.5 and
3.6, impulse responses of F
T
and F
opt
are given.
Note that the structure of controller is same, however exact cancellation simplies
the controller expression.
3.4 Summary of Results
In this Chapter, we determined a necessary and sucient condition for general
time-delay systems for applicability of the Skew-Toeplitz approach [2]. Note that
the neutral and retarded delay systems are the generalized versions of the time-
delay systems. The optimal H
controller design
for systems with time delays is given. It is known that for a certain class of time
delay systems the optimal H
controller C
opt
is determined. Then, F will
be designed to yield a stable K, such that the feedback system remains stable, and
WS is relatively close to the optimal weighted sensitivity WS
opt
:= W(1+PC
opt
)
1
.
When the plant P contains a time delay, and the sensitivity weight W is bi-proper,
the indirect approach outlined above requires internal stabilization of C (which con-
tains innitely many unstable modes) by F. In Chapter 2, a solution to the two-block
H
control problem involving a plant with innitely many poles in the open right
40
r
z
y
+ +
C P
F
K
W
Figure 4.1: (i) F = 0: the weighted sensitivity is H
+ tan
1
+ tan
1
= (4.1)
where
=
_
1
2
2
, and < <
1
. Once
opt
is computed as above, the corre-
sponding optimal controller is
C
opt
(s) =
(1
2
opt
2
) + (
2
opt
2
)s
2
opt
( + s)(1 + s)
N
1
p
(s)
1 +
opt
_
s
1+s
_
e
hs
. (4.2)
Also, dene the optimal sensitivity function as S
opt
(s) = (1 +P(s)C
opt
(s))
1
, then,
S
opt
(j) =
1 +
_
opt
(j)
1+j
_
e
jh
1 +
_
1j
opt
(+j)
_
e
jh
. (4.3)
In [47], it was mentioned that H
opt
e
jh
1
opt
e
jh
and |S
1
opt
(j)|
opt
. Since <
opt
<
1
opt
_
+j
1+j
_
. Then,
W(j)S
opt
(j) =
opt
_
g
1
(j) + m
1
(j)
1 + g
1
(j)m
2
(j)
_
=
opt
_
1 + g(j)m
1
(j)
g(j) + m
2
(j)
_
and hence |W(j)S
opt
(j)| =
opt
as expected.
4.2 Sensitivity Deviation Problem
Recall that the H
0
:=
opt
= inf
C stab. P
W(1 + PC)
1
where W(s) =
1+s
s+
, with > 0, > 0, < 1, and P(s) = N
p
(s)M
p
(s), with
N
p
, N
1
p
H
, and M
p
is inner and innite dimensional, e.g. M
p
(s) = e
hs
. We
have obtained the optimal controller for the sensitivity minimization problem in (4.2).
Claim: The optimal H
.
It is easy to verify this claim by comparing (4.2) with (4.4): we see that
N
c
(s) =
2
opt
W
2
(s)m
2
(s) m
1
(s)
1 +
1
opt
W(s)m
2
(s)
,
=
1
2
opt
( + s)
2
(1
2
opt
2
) + (
2
opt
2
)s
2
1 +
1
opt
_
1s
+s
_
e
hs
(4.5)
D
c
(s) =
1
opt
W(s) + m
1
(s)
1 +
1
opt
W(s)m
2
(s)
=
1
opt
W(s)
_
1 + m
1
(s)
opt
W
1
(s)
1 +m
2
(s)
1
opt
W(s)
_
,
=
1
opt
W(s)(1 + P(s)C
opt
(s))
1
,
=
_
s
+s
_
e
hs
+
1
0
_
1+s
+s
_
1 +
1
0
_
1s
+s
_
e
hs
(4.6)
43
where m
1
(s) =
_
s
+s
_
e
hs
, m
2
(s) =
_
1s
1+s
_
e
hs
. Note that N
c
(s) has no right half
poles or zeros (it has only two imaginary axis poles that are cancelled by the zeros
at the same locations). Therefore N
c
, N
1
c
H
1
0
W
D
c
D
c
+ M
p
N
c
_
.
Our goal is to have a stable controller K, by an appropriate selection of F:
K(s) =
C
opt
(s)
1 + F(s)C
opt
(s)
.
At the same time we would like to have the resulting sensitivity function,
S(s) = (1 +P(s)K(s))
1
=
_
1 + M
p
(s)N
p
(s)
N
1
p
(s)
N
c
(s)
D
c
(s)
1 + F(s)N
1
p
(s)
N
c
(s)
D
c
(s)
_
1
, (4.7)
to be close to the optimal sensitivity, S
opt
= (1 +PC
opt
)
1
. By the parameterization
of the set of all stabilizing controllers for C
opt
[44], F can be written as,
F(s) =
X(s) + D
c
(s)Q(s)
Y (s) N
1
p
(s)N
c
(s)Q(s)
with N
1
p
(s)N
c
(s)X(s) + D
c
(s)Y (s) = 1 which can be solved as Y = 0 and X =
N
1
c
N
p
where Q H
,
Q(s) = 0 is equivalent to
1,opt
= inf
QH
_
_
_
_
W
_
S
0
S
S
__
_
_
_
,
= inf
QH
_
_
_
_
W
_
(M
p
N
c
)(1 + D
c
N
p
N
1
c
Q)
D
c
+ M
p
N
c
__
_
_
_
. (4.11)
Note that, |D
c
(j) + M
p
(j)N
c
(j)| = |
1
0
W(j)| as shown before. Then,
1,opt
= inf
QH
0
N
c
(1 + D
c
Q)
(4.12)
where
Q = N
p
N
1
c
Q. For stability of the feedback system formed by the resulting
controller K and the original plant P, we also want the sensitivity function S =
(1 M
p
N
p
Q)
1
to be stable. Once the optimal Q is determined from (4.12), a
sucient condition for stability of S (and hence the original feedback system) can be
determined as
|N
p
(j)| < |Q(j)|
1
(4.13)
Note that the problem dened in (4.12) is equivalent to a sensitivity minimization
with an innite dimensional weight
0
N
c
for a stable innite dimensional plant
D
c
. For the case where both the plant and the weight are innite dimensional,
sensitivity minimization problem is dicult to solve. So, we propose to approximate
the weight by a nite dimensional upper bound function: nd a stable rational weight
W
1
such that |
0
N
c
(j)| |W
1
(j)|. We suggest an envelope which is in the form,
W
1
(s) =
0
K
s +
1
s +
1
,
45
where
K = 1 +
1
opt
1
= (
opt
+ )(1
opt
)
1
1
and
1
is determined in some optimal fashion.
Then, we can solve the one following block problem as in Chapter 2
1,opt
2,opt
= inf
QH
W
1
(1 + D
c
Q)
. (4.14)
Note that
2,opt
is the smallest value of
2
, in the range
0
K <
2
< (
0
K)
1
,
satisfying
= tan
1
_
1
_
+ h + tan
1
_
1
0
cos(h) + (
1
0
sin(h))
( +
1
0
cos(h)) +
1
0
sin(h)
_
+tan
1
_
1
_
tan
1
_
1
0
cos(h) (
1
0
sin(h))
( +
1
0
cos(h))
1
0
sin(h)
_
(4.15)
where =
_
(
o
K)
2
2
1
2
2
2
1
2
2
(
0
K)
2
. After nding
2,opt
, we can write the C
2,opt
as,
C
2,opt
(s) = A(s)
1
1 D
c
(s)B(s)
(4.16)
where,
A(s) =
(
2
0
K
2
2
1
2
2,opt
2
1
) + (
2
2,opt
2
0
K
2
)s
2
0
K
2,opt
(
1
+ s)(
1
+ s)
B(s) =
_
2,opt
0
K
__
1
s
1
+ s
_
.
In order to calculate
Q
2,opt
(s) corresponding to C
2,opt
(s), we will use the transforma-
tion
Q
2,opt
(s) =
C
2,opt
(s)
1 + P(s)C
2,opt
(s)
.
46
That gives
Q
2,opt
(s) = A(s)
1
1 D
c
(s)B
1
(s)
=
(
2
1
2
K
2
1
) + (
2
K
1)s
2
(
1
+ s)(
K
(
1
+ s) D
c
(s)(
1
s))
(4.17)
where
K
=
2,opt
0
K
. After nding
Q
2,opt
(s), F(s) can be calculated via (4.8),
F(s) = (
Q
1
2,opt
(s) + C
1
opt
(s))
where
Q
2,opt
(s) and C
opt
(s) are found in (4.17) and (4.2) respectively.
Similarly, the resulting controller K(s) is determined as
K(s) =
Q
2,opt
(s)
which is shown in (4.9).
4.3 Summary of Results
An indirect approach to design stable controller achieving a desired H
perfor-
mance level for time delay systems is studied. This approach is based on stabilization
of H
controller by another H
performance level. When the optimal controller is unstable (with innitely or nitely
many unstable poles), two methods are given based on a search algorithm to nd
a stable suboptimal controller. However, both methods are conservative. In other
words, there may be a stable suboptimal controller achieving a smaller performance
level, but the designed controller satises the desired overall H
controllers.
It is known that a H
controller, C
opt
,
stabilizes the feedback system and achieves the minimum H
cost,
opt
:
opt
=
_
_
_
_
_
W
1
(1 +PC
opt
)
1
W
2
PC
opt
(1 + PC
opt
)
1
__
_
_
_
, (5.2)
= inf
C stabilizing P
_
_
_
_
_
W
1
(1 + PC)
1
W
2
PC(1 +PC)
1
__
_
_
_
where W
1
and W
2
are nite dimensional weights for the mixed sensitivity minimization
problem.
In the next section, the structure of optimal and suboptimal H
controllers will
be summarized. The optimal controller with innitely many unstable poles case is
considered in Section 5.2. The conditions and a design method for stable suboptimal
H
controller is given in the same section. Similar work is done in Section 5.3 for
the optimal controller with nitely many unstable poles. Examples related for these
design methods are presented in Section 5.4, and concluding remarks can be found in
Section 5.5.
5.1 Structure of H
Controllers
Assume that the problem (5.2) satises (W
2
N
o
), (W
2
N
o
)
1
H
, then optimal
H
controller can be written as [2] (for controller calculations, see Appendix A),
C
opt
(s) = E
opt
(s)m
d
(s)
N
1
o
(s)F
opt
(s)L(s)
1 + m
n
(s)F
opt
(s)L(s)
. (5.3)
Similarly, the suboptimal controller achieving the performance level can be de-
ned as
C
subopt
(s) = E
(s)m
d
(s)
N
1
o
(s)F
(s)L
U
(s)
1 + m
n
(s)F
(s)L
U
(s)
. (5.4)
50
Note that the unstable zeros of E
opt
and m
d
are always cancelled by the denomi-
nator in (5.3). Therefore, C
opt
is stable if and only if the denominator in (5.3) has no
unstable zeros except the unstable zeros of E
opt
and m
d
(multiplicities considered).
Same conclusions are valid for the suboptimal case, C
subopt
is stable provided that
the denominator in (5.4) has unstable zeros only at the unstable zeros of E
and m
d
(again, multiplicities considered).
It is clear that the optimal, respectively suboptimal, controllers have innitely
many unstable poles if and only if there exists
o
> 0 such that the following inequality
holds
lim
|F
opt
(
o
+ j)L
opt
(
o
+ j)| > 1, (5.5)
respectively,
lim
|F
(
o
+ j)L
U
(
o
+ j)| > 1. (5.6)
The controller may have innitely many poles because of the delay term in the de-
nominator. All the other terms are nite dimensional.
Even when the optimal controller has innitely many unstable poles, a stable
suboptimal controller may be found by proper selection of the free parameter U(s).
In Section 5.2, this case is discussed.
Note that the previous case covers one and two block cases (i.e., W
2
= 0 and
W
2
=0 respectively). When F
opt
is strictly proper, then the optimal and suboptimal
controllers may have only nitely many unstable poles. Existence of stable suboptimal
H
controllers and their design will be discussed in Section 5.3 for this case.
51
5.2 Stable suboptimal H
|F
(j)L
U
(j)| 1. (5.7)
Proof Assume that the suboptimal controller has innitely many unstable poles,
then the equation
1 + e
h(+j)
M( + j)F
( + j)L
U
( + j) = 0
has innitely many zeros in the right half plane, i.e. (recall Section 3.1.1) there exists
=
o
> 0 and for suciently large ,
1 + e
h(
o
+j)
lim
(F
(
o
+ j)L
U
(
o
+ j)) = 0 (5.8)
will have innitely many zeros. Since F
and L
U
are nite dimensional,
lim
(j) = lim
( + j)
lim
L
U
(j) = lim
L
U
( + j) > 0.
By using this fact, we can rewrite (5.8) as,
1 + e
h(
o
+j)
lim
(F
(j)L
U
(j)) = 0 (5.9)
52
which implies that in order to have innitely many zeros, the condition in lemma
should be satised. Conversely, a similar idea can be used to show that (5.7) implies
nitely many unstable poles.
Note that this lemma is valid not only for only nite dimensional U(s) term, but
also for any U H
, U
1 provided that
lim
U(j) = lim
U( + j) = u
, > 0. (5.10)
is satised where u
= lim
|F
(j)| > 1,
u
= lim
U(j),
k = lim
L
2
(j)
L
1
(j)
.
Lemma 5.2.2. The suboptimal controller has nitely many unstable poles if and only
if the following inequalities hold:
|k|
1
f
, |u
|
1 f
|k|
f
|k|
(5.11)
when (n
1
+ l) is odd (even) and ku
< 0, (ku
|k| 1
f
|k|
< |u
| <
f
|k| + 1
f
+|k|
(5.12)
when (n
1
+ l) is odd (even) and ku
> 0, (ku
< 0).
Proof By using Lemma 5.2.1, when (n
1
+ l) is odd (even) and ku
< 0, (ku
> 0),
we can re-write (5.7) as
f
|k| +|u
|
1 +|k||u
|
1.
53
After algebraic manipulations and using f
> 0, (ku
|k| |u
|
1 |k||u
1,
and (5.12) satises this condition.
Note that u
and k.
Theorem 5.2.1. Assume that the optimal and central suboptimal controller (when
U = 0) has innitely many unstable poles, if there exists U H
, U
< 1 such
that L
1U
has no C
+
zeros and |L
U
(j)F
(s
o
)L
U
(s
o
)| > 1 e
h
|F
(j)L
U
(j)| > 0,
therefore, there is no unstable zero, s
o
= +j with > 0. Since, all imaginary axis
zeros are cancelled by E
max
= max
|L
U
(j)F
(j)|=1
,
max
= max
[0,)
|L
U
(j)F
(j)|.
54
It is important to design
max
and
max
as small as possible by the choice of U. Oth-
erwise, at high frequencies the delay term will generate unstable zeros when
max
is
large. Similarly, when
max
is large, although
max
is small, it may cause unstable
zeros. The design method given below searches for a rst order U, and it is based on
the above ideas. An example will be given in Section 5.4.1.
Algorithm
Dene U(s) = u
_
u
z
+s
u
p
+s
_
such that u
, u
p
, u
z
R, |u
| < 1, u
p
> 0 and u
p
u
|u
z
|,
1) Fix >
opt
,
2) Obtain f
, u
p
, u
z
) such that L
1U
(s) is stable,
5) Find the minimum
max
and
max
for all admissible (u
, u
p
, u
z
).
6) Check in the region D = {s = +j, 0 : |e
hs
M(s)F
(s)L
U
(s)| > 1} whether
1 + e
hs
M(s)F
(s)L
U
(s) has no C
+
zeros except unstable zeros of E
and m
d
.
When the central suboptimal controller has innitely many unstable poles, it is
not possible to obtain a stable suboptimal controller by a choice of U as strictly proper
or inner function. Once we nd a U from the above algorithm, the resulting stable
suboptimal H
opt
L
opt
and F
L
U
has magnitude greater than one as . It is not dicult to
see that controllers will have nitely many unstable poles if F
opt
and F
are strictly
proper. Since, these terms decrease as and delay term decays as increases,
only nitely many unstable poles may appear. Clearly, there may be H
controllers
(depending on parameter values) with nitely many poles while F
opt
and F
are bi-
proper. However, it is important to nd the sucient conditions when they are strictly
proper, which results in controllers with nitely many unstable poles regardless of
parameters.
Lemma 5.3.1. The H
opt
is strictly proper,
i.e. (F
opt
(s)) > 0. To show this, we can write by using denition of F
opt
and (A.4),
(F
opt
(s)) = (G
opt
(s)),
=
1
2
(
_
G
W
1
(s) + G
W
2
(s)
2
opt
G
W
1
(s)G
W
2
(s)
_
1
),
=
1
2
(
_
G
W
1
(s) + G
W
2
(s)
2
opt
G
W
1
(s)G
W
2
(s)
_
),
=
1
2
min {(G
W
1
(s)), (G
W
2
(s)), (G
W
1
(s)G
W
2
(s))},
= min {(W
1
(s)), (W
2
(s)), (W
1
(s)) + (W
2
(s))}.
Strictly properness of F
opt
implies,
min {(W
1
(s)), (W
2
(s)), (W
1
(s)) + (W
2
(s))} < 0. (5.13)
We know that (W
1
(s)) 0 and (W
2
(s)) 0, [2]. Therefore, the inequality (5.13)
is satised if and only if (W
1
(s)) 0 and (W
2
(s)) < 0 are valid which means that
W
1
(s) is proper and W
2
(s) is improper. Since we have (W
2
N
o
)
1
RH
[2], we
can conclude that the plant is strictly proper. Same proof is valid for the suboptimal
case.
We know that the suboptimal controllers are written as (5.4),
C
subopt
(s) = E
(s)m
d
(s)
N
1
o
(s)F
(s)L
U
(s)
1 + m
n
(s)F
(s)L
U
(s)
57
we can rewrite the suboptimal controllers as,
C
subopt
(s) =
_
N
1
o
(s)F
(s)
dE
(s)dm
d
(s)
__
L
2
(s) + L
1
(s)m
n
(s)F
(s)
P
1
(s) + P
2
(s)U(s)
_
where
P
1
(s) =
L
1
(s) + L
2
(s)m
n
(s)F
(s)
dE
(s)dm
d
(s)
,
P
2
(s) =
L
2
(s) + L
1
(s)m
n
(s)F
(s)
nE
(s)nm
d
(s)
,
and nE
, dE
and nm
d
, dm
d
are numerator and denominator of E
and m
d
respec-
tively. Denominators of P
1
and P
2
are cancelled by numerators.
Note that unstable poles of C
subopt
are the zeros of P
1
+ P
2
U. If there exists a
U RH
with U
P =
M
d
N
o
where
M and
M
d
are inner, nite dimensional and
N
o
is outer and innite dimensional.
Finding stable S
U
with U H
opt
= inf
UH
S
U
= inf
UH
(1 +
PU)
1
.
58
If we x as >
opt
, then there exists a free parameter Q (Q H
and Q
1)
which parameterizes all functions stabilizing S
U
and achieving performance level .
We will show that the sensitivity function achieving performance level as S
U,
(Q).
Lemma 5.3.2. Assume that W
1
and W
2
are proper and improper respectively. If
there exists
o
>
opt
and Q
o
with Q
o
H
and Q
o
1 satisfying
_
1 S
U,
o
(Q
o
(j))
S
U,
o
(Q
o
(j))
__
P
1
(j)
P
2
(j)
_
1, (5.14)
then the suboptimal controller, C
subopt
, achieves the performance level by selecting
the parameter U as,
U(s) =
_
1 S
U,
o
(Q
o
(s))((s)
S
U,
o
(Q
o
(s))
__
P
1
(s)
P
2
(s)
_
(5.15)
Proof The result of theorem is immediate. Since Q
o
satises the norm condition of
U and makes S
U,
(Q
o
) stable, the suboptimal controller has no right half plane poles
by selection of U as shown in theorem.
A stable suboptimal controller can be designed by nding Q
o
for
o
. By using
a search algorithm, we can nd Q
o
satisfying the norm condition for U. Instead of
nding U resulting in a stable suboptimal controller, the problem is converted to nd
Q
o
satisfying the norm condition. First problem needs to check whether a quasi-
polynomial has unstable zeros. However, by using the theorem, this problem reduced
into searching stable function with innity norm less than one and satisfying norm
condition for U. Conservatively, the search algorithm for Q
o
can be done for rst
order bi-proper functions such that Q
o
(s) = u
_
s+z
u
s+p
u
_
where p
u
> 0, z
u
R, and
|u
| max {1,
p
u
|z
u
|
}. The algorithm for this approach is explained below.
59
Algorithm
Assume that the optimal and central suboptimal controllers have nitely many un-
stable poles. We can design a stable suboptimal H
M
d
(s
i
)
and z
i
=
s
i
a
s+a
where a > 0.
4) Search for minimum which makes the Pick matrix positive semi-denite,
Q
P
{}
(i,k)
=
_
ln
w
i
ln
w
k
+ j2(n
k
n
i
)
1 z
i
z
k
_
(5.16)
where n
[.]
is integer. Note that most of the integers will not result in posi-
tive semi-denite Pick matrix. Therefore, for each integer set, we can nd the
smallest and
opt
will be the minimum of these vales. For details, see [33].
5) After the integer set and
opt
is found, the function g(z) H
can be obtained
satisfying interpolation conditions,
g(z
i
) = ln
w
i
opt
j2n
i
(5.17)
60
by Nevanlinna-Pick interpolation approach [2, 49]. Then, we can write S
U
(s) =
opt
M
d
(s)e
G(s)
where G(s) = g(
sa
s+a
) and obtain U(s). Check the norm con-
dition U
with
interpolation conditions,
g(z
i
) = ln
w
i
j2n
i
. (5.18)
Note that since g(z) has a free parameter q(z) (q H
and q
1), we can
write the function as g(z, q). Then, search for parameters (u
,z
u
,p
u
) satisfying
_
1
M
d
(j)e
G(j,Q)
M
d
(j)e
G(j,Q)
_
_
P
1
(j)
P
2
(j)
_
1, [0, ) (5.19)
where G(s, Q(s)) = g(
sa
s+a
, q(
sa
s+a
)) and Q(s) = u
_
s+z
u
s+p
u
_
as dened before. If
one of the parameter set satises the inequality, then Q
o
= u
,o
_
s+z
u,o
s+p
u,o
_
and
corresponding U results in a stable suboptimal H
controller. If no parameter
set satises the inequality, go to step 6, and repeat the procedure for suciently
high , until a pre-specied maximum is reached, in which case go next step.
7) Increase , go to step 2, if a maximum pre-specied is reached, stop. This
method fails to provide a stable H
controller.
An illustrative example is presented in Section 5.4.2.
5.4 Examples
Two examples will be given in this section. In the rst example, the optimal and
central suboptimal controllers have innitely many unstable poles; by using the design
61
method, we show that there exists a stable suboptimal controller even the magnitude
condition (|L
U
(j)F
con-
troller whose central controller is unstable with nitely many unstable poles and
implements the algorithm step by step as mentioned in Section 5.3.
5.4.1 Example 1
Let P(s) = e
0.1s
_
s1
s+1
_
and choose W
1
(s) =
1+0.6s
s+1
and W
2
= 0 (one-block prob-
lem). Using Skew-Teoplitz approach in [2], the minimum H
value,
opt
, is 0.8108.
The optimal controller has innitely many unstable poles converging to the asymptote
s = 3.0109 j
(2k+1)
h
as k . If central suboptimal controller (U = 0) is calcu-
lated for = 0.814, it has innitely many unstable poles converging to the asymptote
s = 2.445 j
(2k+1)
h
as k . The suboptimal controllers can be represented as,
C
subopt
(s) = E
(s)
F
(s)L
U
(s)
1 + m
n
(s)F
(s)L
U
(s)
(5.20)
where
m
n
(s) = e
0.1s
_
s 1
s + 1
_
,
E
(s) =
0.3374 + 0.3026s
2
0.6626(1 s
2
)
,
F
(s) = 0.814
_
1 s
1 + 0.6s
_
,
L
U
(s) =
L
2U
(s)
L
1U
(s)
=
L
2
(s) + L
1
(s)U(s)
L
1
(s) + L
2
(s)U(s)
,
L
2
(s) = (0.9413s + 1.8716),
L
1
(s) = (s + 1.8373).
62
We will use the design method of the Section 3 to nd a stable suboptimal con-
troller by search for U. The central suboptimal controller (U = 0) has innitely many
unstable poles as mentioned before. The algorithm is tried for u
z
= u
p
= 0 case, i.e.,
U(s) = u
.
1) Fix = 0.814 >
opt
= 0.8108,
2) k = 0.9413 and f
< 0.6668.
4) L
1U
(s) is stable for u
[1, 0.98].
5) Overall admissible values for U are u
value as
max
= 19.458 at u
= 0.813.
6) Figure 5.2 shows the plot of Z(s) = |1 + e
hs
M(s)F
(s)L
U
(s)| in the right half
plane. The function has only right half plane zero at s = 1.056j, which is
right half zeros of E
norm = 0.814.
63
1 0.95 0.9 0.85 0.8 0.75 0.7 0.65 0.6
0
20
40
60
80
100
120
140
160
max
and
max
versus u
max
max
Figure 5.1: w
max
and
max
versus u
0
10
20
30
40
0
1
2
3
4
5
0
0.5
1
1.5
2
2.5
j
Graph of Z(+j)
opt
(s)
F
opt
(s)L
opt
(s)
1 + m
n
(s)F
opt
(s)L
opt
(s)
(5.21)
where
m
n
(s) = e
3s
,
E
opt
(s) =
1.2162 + 2.7838s
2
3.7838(1 s
2
)
,
F
(s) = 5.5119
(1 s)
(2.24 + s)
2
,
L
opt
(s) = 1.
The optimal controller has unstable poles at s = 0.0292 2.2354j. Note that
since W
1
and W
2
are proper and improper respectively, all H
(s)
F
(s)L
U
(s)
1 + m
n
(s)F
(s)L
U
(s)
(5.22)
where
m
n
(s) = e
3s
,
E
(s) =
1.2154 + 2.7846s
2
3.7846(1 s
2
)
,
F
(s) = 5.5115
(1 s)
(2.24 + s)
2
,
L
U
(s) =
L
2U
(s)
L
1U
(s)
=
L
2
(s) + L
1
(s)U(s)
L
1
(s) + L
2
(s)U(s)
,
L
2
(s) = (2.9837 + 0.9946s),
L
1
(s) = (2.9829 +s),
and U is free parameter such that U H
, U
1. We can write P
1
and
P
2
as,
P
1
(s) =
L
1
(s) + m
n
(s)F
(s)L
2
(s)
nE
(s)
,
=
(2.9829 +s)(2.24 + s)
2
+ 5.5115(1 s)(2.9837 + 0.9946s)e
3s
(1.2154 + 2.7846s
2
)(2.24 +s)
2
,
P
2
(s) =
L
2
(s) + m
n
(s)F
(s)L
1
(s)
nE
(s)
,
=
(2.9837 0.9946s)(2.24 + s)
2
+ 5.5115(1 s)(2.9829 s)e
3s
(1.2154 + 2.7846s
2
)(2.24 + s)
2
.
65
Note that P
1
and P
2
has unstable zeros at 0.02872.2346j and 0.02972.2346j
respectively. Therefore, the central controller (U = 0) for the chosen perfor-
mance level, = 1.9458, is unstable.
3) Dene the following variables and functions as,
p
i
= 0.0287 2.2346j, i = 1, 2,
s
i
= 0.0297 2.2346j, i = 1, 2,
M
d
(s) =
(s p
1
)(s p
2
)
(s + p
1
)(s + p
2
)
=
s
2
0.0574s + 4.9943
s
2
+ 0.0574s + 4.9943
,
w
i
=
1
M
d
(s
i
)
= 58.4002 0.7501j, i = 1, 2,
z
i
=
s
i
1
s
i
+ 1
= 0.6598 0.7383i
where conformal mapping parameter, a, is chosen as a = 1.
4) In order to nd the minimum resulting in positive semi-denite Pick matrix,
Q
P
{} =
_
q
11
+2 ln
q
21
(q
11
+q
12
)+2 ln +j2(n
2
n
1
)
q
22
(q
11
q
12
)+2 ln +j2(n
1
n
2
)
q
22
q
11
+2 ln
q
21
_
(5.23)
where q
11
= 8.1348, q
21
= 0.0196, q
12
= 0.0257j and q
22
= 1.1097 0.9742j,
we will nd the minimum for all possible integer pairs (n
1
, n
2
). It is not
dicult to do this search since many integer pairs do not result in positive
semi-denite Pick matrix. For each integer pair, we can nd the minimum ,
min
, and then
opt
will be smallest of all
min
. Note that since Pick matrix
depends on dierence of integers, we can normalize the search by taking n
1
= 0.
In Figure 5.3, we can see the minimum values for integers, n
2
. The minimum
of all
min
values is
opt
= 58.4167.
66
20 15 10 5 0 5 10 15 20
50
60
70
80
90
100
110
120
130
140
m
in
n
2
values
The minimum for n
2
values
Figure 5.3:
min
versus n
2
0 10 20 30 40 50 60 70 80 90 100
0.88
0.9
0.92
0.94
0.96
0.98
1
frequency
|
Q
(
j
)
|
Magnitude plot of U(j)
Figure 5.4: Magnitude plot of U(j)
5) The calculation of U(s) for
opt
is omitted. It does not satisfy the norm condition
U
1.
6) Fix = 64 and n
1
= n
2
= 0. The interpolation conditions for g(z) can be written
as,
g(z
i
) = 0.0915 0.0128j, i = 1, 2. (5.24)
By Nevanlinna-Pick approach, (see e.g.[2]),
g(z, q) =
(1.0878z
2
1.3782z + 0.9804)q(z) + (0.0724z 0.1054)
(0.9804z
2
1.3782z + 1.0878) + (0.0724z 0.1054z
2
)q(z)
(5.25)
where q(z) is a parameterization term such that q H
and q
1. The
search algorithm tries to nd q
o
satisfying the norm condition
_
1
M
d
(j)e
G(j,Q)
M
d
(j)e
G(j,Q)
_
_
P
1
(j)
P
2
(j)
_
1, [0, ) (5.26)
67
where G(s, Q(s)) is equal to
g
_
s 1
s + 1
, q
_
s 1
s + 1
__
,
=
(0.69s
2
0.2148s + 3.4464)Q(s) (0.0330s
2
+ 0.3556s + 0.0330)
(0.69s
2
+ 0.2148s + 3.4464) (0.0330s
2
0.3556s + 0.0330)Q(s)
and Q H
, Q
with |u
. The various u
controller can
be seen in Figure 5.5. We can observe that as is increased, the range of u
in the u
controllers
whose structure is given in [2, 15]. We considered the controllers in two subsections
68
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
0.97
0.975
0.98
0.985
0.99
0.995
1
1.005
u
in
f
in
it
y
n
o
r
m
o
f
U
(
s
)
=70
=100
=150
=200
=300
=1000
Figure 5.5: u
controller
according to their number of poles (nite, innite). For each case, necessary condi-
tions and design methods based on simple sucient condition are given to nd stable
suboptimal H
controllers.
69
CHAPTER 6
REMARKS ON STRONG STABILIZATION AND STABLE
H
CONTROLLER DESIGN
In this Chapter, a state space based design method is given to nd strongly stabi-
lizing controllers for multi-input-multi-output plants (MIMO). A sucient condition
is derived for the existence of suboptimal stable H
K
G
-
z
y
w
u
Figure 6.1: Standard Feedback System
Notation
The notation is fairly standard. A state space realization of a transfer function,
G(s) = C(sI A)
1
B + D, is shown by G(s) =
_
A B
C D
_
and the linear frac-
tional transformation of G by K is denoted by F
l
(G, K) which is equivalent to
G
11
+ G
12
K(I G
22
K)
1
G
21
where G is partitioned as G =
_
G
11
G
12
G
21
G
22
_
. As
a shorthand notation for LMI expressions, we will dene (A, B) := B
T
A
T
+ AB
where A, B are matrices with compatible dimensions.
6.1 Strong stabilization of MIMO systems
Consider the standard feedback system shown in Figure 6.1 with generalized plant,
G, which has state space realization,
G(s) =
_
_
A B
1
B
C
1
D
11
D
12
C D
21
0
_
_
. (6.1)
where A R
nn
, D
12
R
p
1
m
2
, D
21
R
p
2
m
1
and other matrices are compatible
with each other. We suppose the plant satises the standard assumptions,
A.1 (A, B) is stabilizable and (C, A) is detectable,
71
A.2
_
A I B
C
1
D
12
_
has full column rank for all Re{} 0,
A.3
_
A I B
1
C D
21
_
has full row rank for all Re{} 0,
A.4 A has no eigenvalues on the imaginary axis.
Let the controller has state space realization, K
G
(s) =
_
A
K
B
K
C
K
0
_
where A
K
R
nn
, B
K
R
np
2
and C
K
R
m
2
n
. Dene the matrix X R
nn
, X = X
T
0 as
the stabilizing solution of
A
T
X + XA XBB
T
X = 0 (6.2)
(i.e., A BB
T
X is stable) and the A-matrix of the closed loop system as A
CL
=
_
A BC
K
B
K
C A
K
_
. Note that since (A, B) is stabilizable, X is unique and A
X
:=
(A BB
T
X) is stable. Also, the closed loop stability is equivalent to whether A
CL
is stable or not.
Lemma 6.1.1. Assume that the plant (6.1) satises the assumptions A.1 A.4.
There exists a stable stabilizing controller K
G
RH
if there exists X
K
R
nn
,
X
K
= X
T
K
> 0 and Z R
np
2
for some
K
> 0 satisfying the LMIs
(X
K
, A) + (Z, C) < 0, (6.3)
_
_
(X
K
, A
X
) + (Z, C) Z XB
Z
T
K
I 0
B
T
X 0
K
I
_
_
< 0, (6.4)
where X is the stabilizing solution of (6.2) and A
X
is as dened previously. Moreover,
a stable controller has state space realization K
G
(s) =
_
A
X
+ X
1
K
ZC X
1
K
Z
B
T
X 0
_
and this controller satises K
G
<
K
.
72
Proof. By using similarity transformation, one can show that A
CL
is stable if and
only if A
X
and A
Z
:= A + X
1
K
ZC are stable. Since X is a stabilizing solution, A
X
is stable. If we rewrite the LMI (6.3) as
(A + X
1
K
ZC)
T
X
K
+ X
K
(A + X
1
K
ZC) < 0,
it can be seen that A
Z
is stable since X
K
> 0. The second LMI (6.4) comes from
KYP lemma and guarantees that K
G
<
K
.
Remark 1 If the design only requires the stability of closed loop system, it is enough
to satisfy the LMI (6.3), (1, 1) block of (6.4), i.e.,
A
T
X
X
K
+ X
K
A
X
+ C
T
Z
T
+ ZC < 0 (6.5)
and the controller has same structure as above.
Remark 2 The Lemma (6.1.1) is generalization of Theorem 2.1 in [41]. If the al-
gebraic riccati equation (ARE) (7) in [41] has a stabilizing solution, Y = Y
T
0,
then there exists a stable controller in the form,
_
A
X
2
K
Y C
T
C
2
K
Y C
T
B
T
X 0
_
. This
structure is the special case of the LMIs (6.3) and (6.4) when X
K
= (
K
Y )
1
and
Z =
K
C
T
. Note that our formulation does not assume special structure on Z. Also
in [41], the stability of A
Z
is guaranteed by the same riccati equation, we satisfy the
stability condition of A
Z
with another LMI (6.3) which is less restrictive. Therefore,
the Lemma (6.1.1) is less conservative as will be demonstrated in examples.
Corollary 6.1.1. Assume that the sucient condition (6.3) and (6.5) holds. Then
all controllers in the set
K
G,ss
:= {K = F
l
(K
0
G,ss
, Q) : Q RH
, Q
<
Q
}
73
are strongly stabilizing where
K
0
G,ss
(s) =
_
_
A
X
+ X
1
K
ZC X
1
K
Z B
B
T
X 0 I
C I 0
_
_
. (6.6)
and
Q
=
_
C(sI (A
X
+ X
1
K
ZC))
1
B
_
1
.
Proof. The result is direct consequence of parameterization of all stabilizing con-
trollers [51].
6.2 Stable H
problem
is solvable. All suboptimal H
, Q)
where the central controller is in the form
M
(s) =
_
_
A
c
B
c1
B
c2
C
c1
D
c11
D
c12
C
c2
D
c21
0
_
_
and Q is free parameter satisfying Q RH
and Q
as plant and =
K
, by using Lemma (6.1.1), we can nd
a strictly proper stable K
M
stabilizing M
controller,
C
= F
l
(M
, K
M
) where K
M
<
K
. If sucient conditions (6.3) and (6.4) are
satised, then K
M
(s) =
_
A
c
B
c2
B
T
c2
X
c
+ X
1
Kc
Z
c
C
c2
X
1
Kc
Z
c
B
T
c2
X
c
0
_
74
and by similarity transformation, we can obtain the state space realization of C
as,
C
(s) =
_
A
C
B
C
C
C
D
c11
_
where X
c
is the stabilizing solution of
A
T
c
X
c
+ X
c
A
c
X
c
B
c2
B
T
c2
X
c
= 0
as in (6.2) and X
Kc
, Z
c
are the solution of (6.3), (6.4) respectively and the matrices,
A
C
=
_
A
c
B
c2
B
T
c2
X
c
B
c2
B
T
c2
X
c
0 A
c
+ X
1
Kc
Z
c
C
c2
_
B
C
=
_
B
c1
B
c1
X
1
Kc
Z
c
D
c21
_
C
C
=
_
C
c1
D
c12
B
T
c2
X
c
D
c12
B
T
c2
X
c
Note that C
< .
6.3 Numerical examples and comparisons
6.3.1 Strong stabilization
The numerical example is chosen from [41]. In order to see the performance of our
method, we calculated the minimum
K
satisfying the sucient conditions in Lemma
(6.1.1) for the following plants:
G
1
(s) =
_
_
(s+5)(s1)(s5)
(s+2+j)(s+2j)(s)(s20)
(s+1)(s1)(s5)
(s+2+j)(s+2j)(s)(s20)
_
_
G
2
(s) =
_
_
(s+1)(s2j)(s2+j)
(s+2+j)(s+2j)(s1)(s5)
(s+5)(s2j)(s2+j)
(s+2+j)(s+2j)(s1)(s5)
_
_
For various values, the minimum
K
is found. Figure 6.2 and 6.3 illustrates the
conservatism of [41] mentioned in Remark 2 (where
min
is the minimum value of the
free parameter
K
corresponding to the method of [41]).
75
10 20 30 40 50 60 70 80 90 100
40
60
80
100
120
140
160
180
200
Comparison for G
1
min
K
Figure 6.2: Comparison for plant G
1
12 13 14 15 16 17 18 19 20 21 22
0
5
10
15
20
25
30
35
Comparison for G
2
min
K
Figure 6.3: Comparison for plant G
2
6.3.2 Stable H
controllers
We applied our method to stable H
=
_
0.1 0.1 1
0.5 0.5 0
_
,
_
C
1
C
2
_
=
_
_
0.2 1
0 0
10 11.5470
_
_
,
_
D
11
D
12
D
21
0
_
=
_
_
0 0 0
0 0 1
0.7071 0.7071 0
_
_
.
76
The optimal value for standard H
problem is
opt
= 1.2929. Using the syn-
thesis in [42], a stable H
controller is found at
min
= 1.36994. When our method
applied, we reached stable H
controller for
K,min
= 1.36957. Although it seems
slight improvement, our method is much more simpler with help of LMI problem
formulation. Apart from standard problem solution (nding M
), the algorithm in
[42] nds the stable H
problem.
Another common benchmark example is from suppression of thermoacoustical
instabilities in a combustion chamber and this example is used to nd stable H
controller in [40, 8]. The generalized plant is described by the following equations:
z
1
=
0.03s
7
+ 0.008s
6
+ 0.19s
5
+ 0.037s
4
+ 0.36s
3
+ 0.05s
2
+ 0.18s + 0.015
s
8
+ 0.161s
7
+ 6s
6
+ 0.582s
5
+ 9.984s
4
+ 0.407s
3
+ 3.9822s
2
(w
1
+ u),
z
2
= u,
y = w
2
+
0.0064s
5
+ 0.0024s
4
+ 0.071s
3
+ s
2
+ 0.1045s + 1
s
8
+ 0.161s
7
+ 6s
6
+ 0.582s
5
+ 9.984s
4
+ 0.407s
3
+ 3.9822s
2
(w
1
+ u).
In [40], it is noted that for this problem, the sucient condition in [37] is not
satised for even large values of and the method is not applicable. As we can
see from Table 6.1, the performance of our method is better than the method in
[40] except the last case. For all cases, the result of [8] is superior from all other
Gumussoy-Ozbay(GO) [40] [8]
opt
GO
ZO
CZ
0.1 0.232 0.241 0.245 0.237
0.01 0.142 0.176 0.178 0.151
0.001 0.122 0.170 0.170 0.132
Table 6.1: Stable H
problem which
is complicated compared to our simple LMI formulation. Also if the algorithm in
[8] fails, selection of a new parameter Q is suggested which is an ad-hoc procedure.
Although the performance of the controller suggested in this Chapter is slightly worse,
it is numerically stable and easily formulated.
The following example is taken from [1]. Stable H
min
32.557 37.551 43.167 21.787
Order 2n 2n n 3n
Table 6.2: Stable H
problem is dened as
opt
= inf
KstabilizingP
_
_
_
_
_
W
1
(1 + PK)
1
W
2
K(1 + PK)
1
__
_
_
_
controller
can be found for = 42.51 with the method in [41], whereas our method, which can be
seen as a generalization of [41], gives = 35.29. In numerical simulations, we observed
that when approaches to the minimum value satisfying sucient conditions, the
solutions of algebraic riccati equations of [41] become numerically ill-posed. However,
the LMI based solution proposed here does not have such a problem.
6.4 Summary of Results
In this Chapter, sucient conditions for strong stabilization of MIMO systems are
obtained and applied to stable H
controller design for time-delay systems and MIMO systems are studied.
In Chapter 2, the optimal H
controller without
cancellations due to its special structure. Therefore, the result in [23] is extended
from dead-time systems to general time-delay systems.
81
An indirect approach to design stable controller achieving a desired H
perfor-
mance level for time delay systems is given in Chapter 4. This approach is based
on stabilization of H
controller by another H
performance
level. When the optimal controller is unstable with innitely many unstable poles,
a method is given to choose the free parameter of suboptimal controller such that
the controller itself is stable. When the optimal controller is unstable with nitely
many unstable poles, a search algorithm is given to nd stable H
controller using
Nevanlinna-Pick interpolation approach. However, both methods are conservative.
In other words, there may be a stable suboptimal controller achieving a smaller per-
formance level, but the designed controller satises the desired overall H
norm. The
stability of optimal and suboptimal controller is discussed and necessity conditions
are given.
Strong stabilization problem for MIMO systems is studied in Chapter 6. A state
space based design method is given to nd strongly stabilizing controllers for nite
dimensional MIMO plants. A sucient condition is derived for the existence of sub-
optimal stable H
controller design is
considered for time-delay and MIMO systems.
In Chapter 2, the plant is assumed to be bi-proper in order to guarantee the
well-posedness of the plant and its inverse. This assumption is needed since the
method is dual version of Skew-Toeplitz approach. Although, it is not shown in this
thesis, strictly proper plant case with innitely many unstable poles and nitely many
unstable zeros can be solved as well.
It may be possible to generalize H
controller can
also be designed without the factorization. Therefore, the work in Chapter 3 can be
generalized according to this development.
In Chapter 5, the stable H
controller can be
designed and an upper bound on input signal can be calculated such that the feedback
system operates in the linear region with no saturation in signals. Second, state space
representation of the H
controller, C
opt
, stabi-
lizes the feedback system and achieves the minimum H
cost,
opt
:
opt
=
_
_
_
_
_
W
1
(1 +PC
opt
)
1
W
2
PC
opt
(1 + PC
opt
)
1
__
_
_
_
= inf
Cstab.P
_
_
_
_
_
W
1
(1 + PC)
1
W
2
PC(1 + PC)
1
__
_
_
_
(A.2)
where W
1
and W
2
are nite dimensional weights for the mixed sensitivity minimization
problem.
Assume that the problem (A.2) satises (W
2
N
o
), (W
2
N
o
)
1
H
, then optimal
H
opt
(s)m
d
(s)
N
1
o
(s)F
opt
(s)L(s)
1 + m
n
(s)F
opt
(s)L(s)
(A.3)
85
where E
=
_
W
1
(s)W
1
(s)
2
1
_
, and for the denition of the other terms, let the right
half plane zeros of E
(s) be
i
, i = 1, . . . , n
1
, the right half plane poles of P(s) be
i
,
i = 1, . . . , l and that of W
1
(s) be
i
i = 1, . . . , n
1
. Then, F
(s) = G
(s)
n
1
i=1
s
i
s+
i
where
G
(s)G
(s) =
_
1
_
W
2
(s)W
2
(s)
2
1
_
E
_
1
(A.4)
and G
, G
1
, and L(s) =
L
2
(s)
L
1
(s)
, L
1
(s) and L
2
(s) are polynomials with degrees
less than or equal to (n
1
+l1) and they are determined by the following interpolation
conditions,
0 = L
1
(
i
) + m
n
(
i
)F
(
i
)L
2
(
i
) i = 1, . . . , n
1
(A.5)
0 = L
1
(
i
) + m
n
(
i
)F
(
i
)L
2
(
i
) i = 1, . . . , l
0 = L
2
(
i
) + m
n
(
i
)F
(
i
)L
1
(
i
) i = 1, . . . , n
1
0 = L
2
(
i
) + m
n
(
i
)F
(
i
)L
1
(
i
) i = 1, . . . , l.
The optimal performance level,
opt
, is the largest value such that spectral factor-
ization (A.4) exists and interpolation conditions (A.5) are satised.
Similarly, the suboptimal controller achieving the performance level, , can be
dened as,
C
subopt
(s) = E
(s)m
d
(s)
N
1
o
(s)F
(s)L
U
(s)
1 + m
n
(s)F
(s)L
U
(s)
(A.6)
where >
opt
and L
U
(s) =
L
2U
(s)
L
1U
(s)
=
L
2
(s)+L
1
(s)U(s)
L
1
(s)+L
2
(s)U(s)
with U H
, U
1. The
polynomials, L
1
(s) and L
2
(s), have degrees less than or equal to n
1
+ l. Same inter-
polation conditions are valid with instead of . Moreover, there are two additional
86
interpolation conditions for L
1
(s) and L
2
(s):
0 = L
2
(a) + (E
(a) + 1)F
(a)m
n
(a)L
1
(a) (A.7)
0 = L
1
(a) (A.8)
where a R
+
is arbitrary. The above terms and notations are the same as in [2, 15].
87
APPENDIX B
IF AND FI PLANT CALCULATIONS
In this section, we compute certain terms appearing in the examples at Chapter
3.3.1 and 3.3.2.
B.1 Example: IF Plant Case
The optimal controller for the plant (3.6) can be written as,
C
opt
=
H
T
+F
T
H
opt
+F
opt
where
H
T
= H
T,1
(s) + H
T,2
(s)e
0.2s
+ H
T,3
(s)e
0.5s
(B.1)
and
H
T,1
=
0.04538s
4
0.05342s
3
+ 0.02919s
2
+ 0.00159s + 0.0001374
s
5
+ 2.705s
4
+ 2.745s
3
+ 1.402s
2
+ 0.3897s + 0.02744
,
H
T,2
=
0.005009s
4
+ 0.03183s
3
0.06717s
2
+ 0.02217s 0.0003618
s
5
+ 2.705s
4
+ 2.745s
3
+ 1.402s
2
+ 0.3897s + 0.02744
,
H
T,3
=
0.06594s
5
+ 1.065s
4
+ 2.713s
3
+ 2.039s
2
+ 1.156s + 0.262
s
6
+ 2.805s
5
+ 3.015s
4
+ 1.677s
3
+ 0.5299s
2
+ 0.06641s + 0.002744
,
F
T
= F
T,1
(s) +F
T,2
(s)e
0.2s
+F
T,3
(s)e
0.5s
(B.2)
and
F
T,1
=
0.05462s 0.01897
s
2
0.9344s + 3.787
,
F
T,2
=
0.005009s + 0.04994
s
2
0.9344s + 3.787
,
F
T,3
=
0.06594s + 0.08666
s
2
0.9344s + 3.787
.
88
Note that H
opt
= H
R
1
+ H
R
2
and F
opt
= F
R
1
+F
R
2
where
H
R
1
= H
R
1
,1
(s) + H
R
1
,2
(s)e
0.4s
(B.3)
and
H
R
1
,1
=
0.001218
s + 1
,
H
R
1
,2
=
0.0008883s + 0.003325
s
2
+ 2s + 1
,
F
R
1
= F
R
1
,1
(s) +F
R
1
,2
(s)e
0.4s
(B.4)
and
F
R
1
,1
=
0.05903s
4
+ 0.1265s
3
+ 0.3086s
2
+ 0.2892s + 0.05386
s
5
1.181s
4
+ 4.09s
3
1.021s
2
+ 0.2921s 0.068
,
F
R
1
,2
=
0.02802s
4
0.05152s
3
+ 0.009902s
2
0.3267s 0.08069
s
5
1.181s
4
+ 4.09s
3
1.021s
2
+ 0.2921s 0.068
,
H
R
2
= H
R
2
,1
(s) + H
R
2
,2
(s)e
0.4s
(B.5)
and
H
R
2
,1
=
0.04701s
3
+ 0.1765s
2
+ 0.2979s + 0.4496
s
4
+ 2.605s
3
+ 2.484s
2
+ 1.154s + 0.2744
,
H
R
2
,2
=
0.06055s
3
0.009855s
2
+ 0.1892s 0.3028
s
4
+ 2.605s
3
+ 2.484s
2
+ 1.154s + 0.2744
,
F
R
2
= F
R
2
,1
(s) +F
R
2
,2
(s)e
0.4s
(B.6)
and
F
R
2
,1
=
0.04701s
3
0.1185s
2
0.2389s 0.3281
s
4
0.9344s
3
+ 3.859s
2
0.06794s + 0.2753
,
F
R
2
,2
=
0.06055s
3
+ 0.007461s
2
0.1341s + 0.2218
s
4
0.9344s
3
+ 3.859s
2
0.06794s + 0.2753
.
B.2 Example: FF Plant Case
Since h
1
=
1
= 0, m
n
= M
R
is nite dimensional term, it is possible to do exact
cancellation in the term
K
as
89
K
=
_
1 + M
R
F
opt
L
_
opt
, (B.7)
=
_
(s2.255)(s
2
+0.077)(s
2
+2.408s+3.377)(s
2
1.835s+4.048)
(s+1.264)(s+0.789)(s0.240)(s
2
2.584s+3.524)(s
2
+2.757s+4.455)
_
_
33.705(s2.255)(s
2
+0.077)(s
2
1.835s+4.048)
(s+2.255)(s+1.264)(s+0.789)(s
2
+1.835s+4.048)
_ .
After cancellations, we obtain
K
=
0.030s
5
+ 0.193s
4
+ 0.635s
3
+ 1.265s
2
+ 1.472s + 0.915
s
5
0.067s
4
+ 0.813s
3
2.002s
2
+ 16.130s 3.771
. (B.8)
The optimal controller for the plant (3.17) can be written as,
C
opt
=
H
T
+F
T
H
opt
+F
opt
where
H
T
= H
T,1
(s) + H
T,2
(s)e
0.8s
(B.9)
and
H
T,1
=
0.393s
6
+ 1.399s
5
+ 1.657s
4
+ 0.533s
3
0.174s
2
+ 0.232s + 0.076
s
7
+ 0.275s
6
1.311s
5
+ 3.983s
4
+ 7.483s
3
+ 2.313s
2
0.686s 0.085
,
H
T,2
=
0.424s
6
+ 3.658s
5
+ 10.890s
4
+ 16.410s
3
+ 14.200s
2
+ 6.767s + 1.011
s
7
+ 0.275s
6
1.311s
5
+ 3.983s
4
+ 7.483s
3
+ 2.313s
2
0.686s 0.085
,
F
T
= F
T,1
(s) +F
T,2
(s)e
0.8s
(B.10)
and
F
T,1
=
0.293s
2
+ 0.680s + 0.893
s
3
4.090s
2
+ 8.185s 9.128
,
F
T,2
=
0.424s
2
1.808s + 0.548
s
3
4.090s
2
+ 8.185s 9.128
,
H
opt
= H
opt
,1
(s) + H
opt
,2
(s)e
0.5s
+ H
opt
,3
(s)e
0.7s
(B.11)
and
H
opt
,1
=
0.329s
5
0.070s
4
1.983s
3
2.187s
2
0.293s + 0.426
s
6
+ 0.175s
5
1.328s
4
+ 4.115s
3
+ 7.072s
2
+ 1.606s 0.846
,
H
opt
,2
=
0.640s
5
+ 1.411s
4
+ 0.420s
3
0.909s
2
0.754s 0.310
s
6
+ 0.175s
5
1.328s
4
+ 4.115s
3
+ 7.072s
2
+ 1.606s 0.846
,
H
opt
,3
=
0.127s
5
+ 1.773s
4
+ 6.175s
3
+ 6.913s
2
+ 3.290s + 0.824
s
6
+ 0.175s
5
1.328s
4
+ 4.115s
3
+ 7.072s
2
+ 1.606s 0.846
,
90
F
opt
= F
opt
,1
(s) +F
opt
,2
(s)e
0.5s
+F
opt
,3
(s)e
0.7s
(B.12)
and
F
opt
,1
=
0.299s + 1.164
s
2
2.757s + 4.455
,
F
opt
,2
=
0.641s + 0.526
s
2
2.757s + 4.455
,
F
opt
,3
=
0.127s 2.145
s
2
2.757s + 4.455
.
91
BIBLIOGRAPHY
[1] Y. Choi and W. Chung, On the stable H
optimization, System and Control Letters, vol. 32, pp. 111120, 1997.
[8] D. U. Campos-Delgado and K. Zhou, H
: bounds of norm
and degree of interpolants, Systems and Control Letters, vol. 17, pp. 251256,
1991.
[30] O. Toker, On the order of simultaneously stabilizing controllers, IEEE Trans-
actions on Automatic Control, vol. 41, pp. 430433, 1996.
[31] J.L.Abedor and K.Poolla, On the strong stabilization of delay systems, Proc.
Decision and Control Conference, pp. 23172318, 1989.
[32] K. Suyama, Strong stabilization of systems with time-delays, Proc. IEEE In-
dustrial Electronics Society Conference, pp. 17581763, 1991.
[33] C. Ganesh and J. Pearson, Design of optimal control systems with stable feed-
back, Proc. American Control Conf., pp. 19691973, 1986.
[34] A. Barabanov, Design of H
controllers, IEEE
Transactions on Automatic Control, vol. 44, pp. 431435, 1999.
[41] M. Zeren and H.
Ozbay, On the strong stabilization and stable H
-controller
design problems for MIMO systems, Automatica, vol. 36, pp. 16751684, 2000.
[42] P. Lee and Y. Soh, Synthesis of stable H
strong-
stabilization problems, Proc. Conference on Decision and Control, pp. 5155
5160, 2003.
96