Académique Documents
Professionnel Documents
Culture Documents
synthesis.
Since solving LMIs is a convex optimisation problem, such
formulations can be solved efciently using interior-point
algorithms. Linear matrix inequalities have also been suc-
cessful in formulating and solving multi-objective control
problems in which various performance specications (both
in the frequency and the time domains) are used for various
input-output channels [4].
The fact that the positive real character of a system can
also be expressed as an LMI, allows for the formulation
This work was supported in part by EPSRC.
C. Papageorgiou and M. C. Smith are with the Control Group, Cam-
bridge University Engineering Department, Cambridge CB2 1PZ, England
cp206@eng.cam.ac.uk, mcs@eng.cam.ac.uk
of the positive real synthesis problem as a set of matrix
inequalities. Considerable research has been conducted to-
wards the synthesis of positive real controllers that achieve
a level of H
2
performance for the control of exible
structures. In [5] a sub-optimal version of this problem is
shown to be a convex optimisation problem and expressed
in the form of an LMI. In [6] an iterative LMI procedure is
proposed for the same problem. Both methods require that
the order of the controller is the same as the order of the
generalised plant.
In this paper we will use local optimisation algorithms
in order to solve the H
2
and H
F(t)v(t)dt 0. (1)
As shown in [7], [8] the network is passive if and only if the
following condition is satised: Z(s) is analytic in Re(s) >0
and Z(s) +Z(s)
F
s
z
s
_
_
, (5)
where || ||
represents the H
F
m
s
k
s
m
s
z
s
+
k
s
m
s
z
u
, (6)
z
u
=
F
m
u
+
k
s
m
u
z
s
k
s
m
u
z
u
+
k
t
m
u
z
r
k
t
m
u
z
u
, (7)
F = K(s)(s z
s
s z
u
). (8)
The external input of the generalised plant is w = [F
s
, z
r
]
T
,
while the performance output is given by z = [ z
s
, z
s
]
T
.
Writing the above equations in state-space form with the
state vector given by x = [ z
s
, z
s
, z
u
, z
u
]
T
results in the quarter-
car, generalised plant:
x = Ax +B
_
w
F
_
,
_
z
z
s
z
u
_
=Cx, (9)
where,
A =
_
_
0
k
s
m
s
0
k
s
m
s
1 0 0 0
0
k
s
m
u
0
k
s
+k
t
m
u
0 0 1 0
_
_
, B =
_
_
1
m
s
0
1
m
s
0 0 0
0
k
t
m
u
1
m
u
0 0 0
_
_
,
C =
_
_
1 0 0 0
0 1 0 0
1 0 1 0
_
_
.
Let B
1
, B
2
, B
3
be the respective columns of B and let C
1
, C
2
,
C
3
be the respective rows of C. Since the static stiffness of
the suspension is determined a priori, the admittance K(s)
must have zero static stiffness, hence it may not contain an
integrator. This is in fact ensured by the structure of the
generalised plant. The transfer function from F to z
s
z
u
is
given by,
T
F(s z
s
s z
u
)
=
s
_
(m
u
+m
s
)s
2
+k
t
_
m
u
m
s
s
4
+s
2
(m
u
k
s
+k
s
m
s
+k
t
m
s
) +k
t
k
s
.
Therefore a stabilising controller K(s) cannot have an in-
tegrator because it will cause a RHP pole-zero cancellation.
IV. OPTIMISATION OF J
1
The generalised plant for the J
1
optimisation is formed
by considering z
r
as the external disturbance and z
s
as the
performance output. The objective is to nd a positive real
K(s) so that ||T
z
r
s z
s
||
2
is minimised. From (9) the observ-
able and controllable representation of the J
1
generalised
plant is given by,
x = Ax +B
2
z
r
+B
3
F, (10)
z
s
= C
1
x, z
s
z
u
=C
3
x. (11)
Given a controller K(s) of order n
k
, with state-space rep-
resentation as in (2), let the state-space representation of
the closed-loop system resulting from the interconnection
of the generalised plant and the controller be denoted by,
_
_
x
x
k
z
s
_
_
=
_
A
cl
B
cl
C
cl
0
_
_
_
x
x
k
z
r
_
_
. (12)
Theorem 4.1: There exists a positive real controller K(s)
of order n
k
such that ||T
z
r
s z
s
||
2
< and A
cl
is stable, if and
only if the following problem is feasible for some X
cl
> 0,
X
k
>0, Q (symmetric),
2
and A
k
, B
k
, C
k
, D
k
of compatible
dimensions:
_
A
T
cl
X
cl
+X
cl
A
cl
X
cl
B
cl
B
T
cl
X
cl
I
_
< 0,
_
X
cl
C
T
cl
C
cl
Q
_
> 0, tr(Q) <
2
,
_
A
T
k
X
k
+X
k
A
k
X
k
B
k
C
T
k
B
T
k
X
k
C
k
D
T
k
D
k
_
< 0.
5457
The rst three LMIs are necessary and sufcient con-
ditions for the existence of a stabilising controller that
achieves an upper bound of on the H
2
-norm [12]. The
fourth LMI further restricts the controller to be positive
real. Without the positive real constraint the H
2
-synthesis
problem can be formulated as an LMI problem as shown in
[12]. With the positive real constraint it is not obvious how
to do so, hence an iterative optimisation method is employed
to solve the BMI problem locally. The method, which is
described in [13], is to linearise the BMI using a rst-order
perturbation approximation, and then iteratively compute a
perturbation that slightly improves the controller perfor-
mance by solving an LMI problem. The proposed scheme is
already implemented in YALMIP [14], which is a MATLAB
toolbox for rapid prototyping of optimization problems. A
feasible starting point must be given to the algorithm.
A. J
1
optimisation results
There are two issues to be investigated regarding the
proposed synthesis method. The rst is the reproduction of
the results of the xed-structure optimisation and the second
the further improvement in performance.
An example is presented that demonstrates the success
in reproducing the xed-structure optimisation results. The
quarter-car model is considered with static stiffness k
s
=
610
4
N/m. The xed-structure admittance proposed in [2]
is a damper in series with an inerter, i.e. K(s) =cbs/(c+bs),
which achieves a value of J
1
= 1.5851 for c = 3224 Ns/m
and b = 334 kg. Optimising over rst-order admittances
K(s), the use of the YALMIP procedure gave,
K(s) =
3276s +52
s +10
= 5.2 +(1/3271 +1/327.1s)
1
(13)
and the value J
1
= 1.5855. A comparison of the two
admittances is shown in Fig. 4.
10
2
10
1
10
0
10
1
10
2
10
3
10
0
10
2
10
4
L
o
g
M
a
g
n
i
t
u
d
e
Frequency (radians/sec)
LMI optimisation
fixedstructure
10
2
10
1
10
0
10
1
10
2
10
3
0
20
40
60
80
100
P
h
a
s
e
(
d
e
g
r
e
e
s
)
Frequency (radians/sec)
Fig. 4. Comparison of rst-order admittances K(s) for the quarter-car
model for k
s
= 610
4
N/m
It is also investigated whether J
1
can be reduced further
by optimising over higher order admittances. The highest
order admittance considered in [2] was of order 2. The
best results obtained so far with third-order admittances are
presented in Fig. 5 as a function of the static stiffness of
the suspension, along with previous results related to xed-
structure admittances. The best xed-structure admittance
1 2 3 4 5 6 7 8 9 10 11 12
x 10
4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
2.4
2.6
static stiffness in N/m
J
1
layout S1 (damper)
layout S6 (2nd order fixedstructure)
3rd order (BMI optimisation)
Fig. 5. Improvement in J
1
when using higher-order admittances
found in [2] was given the name layout S6, which is an
inerter-damper series arrangement with unequal centring
springs. The percentage improvement is calculated with
respect to the values of J
1
achieved by the S6 xed-
structure admittance. The results are shown in Fig. 6. The
0 2 4 6 8 10 12
x 10
4
2
4
6
8
10
12
14
16
18
20
static stiffness in N/m
%
c
h
a
n
g
e
Fig. 6. Percentage improvement in J
1
when K(s) is third order
optimisation algorithm was run for values of static stiffness
in the range 1 10
4
N/m to 12 10
4
N/m at a spacing of
0.5 10
4
N/m.
V. OPTIMISATION OF J
5
The generalised plant for the J
5
optimisation is formed
by considering F
s
as the external disturbance and z
s
as the
performance output. The objective is to nd a positive real
K(s) so that J
5
= ||T
F
s
z
s
||
F
s
z
s
_
_
< and A
cl
is stable,
if and only if the following problem is feasible for some
X
cl
> 0, X
k
> 0, > 0 and A
k
, B
k
, C
k
, D
k
of compatible
dimensions:
M < 0,
_
A
T
k
X
k
+X
k
A
k
X
k
B
k
C
T
k
B
T
k
X
k
C
k
D
T
k
D
k
_
< 0, (18)
where X
cl
:=
_
X
11
X
12
X
T
12
X
22
_
and M is a symmetric matrix with,
M
11
= X
11
A+()
T
+X
11
B
3
D
k
C
3
+()
T
+X
12
B
k
C
3
+()
T
,
M
12
= X
11
B
3
C
k
+X
12
A
k
+A
T
X
12
+C
T
3
D
T
k
B
T
3
X
12
+C
T
3
B
T
k
X
22
,
M
13
= X
11
B
1
, M
14
=C
T
2
, M
23
= X
T
12
B
1
,
M
22
= X
22
A
k
+()
T
+X
T
12
B
3
C
k
+()
T
,
M
24
= 0
n
k
1
, M
34
= 0, M
33
= M
44
=I,
where ()
T
denotes the transpose of the preceding matrix.
The matrix inequality M < 0 results by applying the
Bounded Real Lemma [3] on the closed-loop system of
(17), which is a necessary and sufcient condition for the
existence of a stabilising controller that achieves an upper
bound on the H
norm was
formulated as a convex optimisation problem in [15]. With
the additional positive real constraint it is not obvious how
to formulate the problem as a convex problem.
The problem of Theorem 5.1 is a BMI and an iterative
algorithm is proposed to solve the problem locally about
a feasible starting point. The idea is to x a subset of the
decision variables so that the matrix inequality is linear with
respect to the remaining decision variables and thus can
be solved efciently. The stopping criterion is how close
the achieved is to a theoretical minimum of J
5
given by
J
5min
= k
1
s
+k
1
t
which is equal to T
F
s
z
s
(0).
B. Optimisation over non-proper controllers
The non-proper admittance (14) can be represented in
ordinary state-space form by the device of introducing a
second measurement equal to the acceleration signal ( z
s
z
u
) as follows:
x
k
= A
k
x
k
+
_
B
k
0
n
k
1
_
z
s
z
u
z
s
z
u
_
, x
k
R
n
k
, (19)
F =C
k
x
k
+
_
D
k
E
k
_
z
s
z
u
z
s
z
u
_
. (20)
The J
5
generalised plant equations are augmented by the
acceleration measurement to give the following equation,
_
z
s
z
u
z
s
z
u
_
=
_
C
3
C
4
_
x +
_
0
D
1
_
F
s
+
_
0
D
2
_
F, (21)
with C
4
=
_
0
k
s
m
s
k
s
m
u
0
k
s
m
s
+
k
s
+k
t
m
u
_
, D
1
=
1
m
s
and
D
2
=
1
m
s
1
m
u
. With the change of variable
Z
k
= (1 E
k
D
2
)
1
E
k
= (1 Z
1
k
)D
1
2
(22)
the closed-loop equations are given by:
_
x
x
k
_
=
_
A+B
3
Z
k
D
k
C
3
+B
3
(Z
k
1)D
1
2
C
4
B
3
Z
k
C
k
B
k
C
3
A
k
__
x
x
k
_
+
+
_
B
1
+B
3
(Z
k
1)D
1
2
D
1
0
n
k
1
_
F
s
,
z
s
=
_
C
2
0
1n
k
_
x
x
k
_
.
The formulation of the synthesis problem follows a
similar reasoning to that of Theorem 5.1 and results in a
BMI which is also solved locally using a similar iterative
algorithm.
C. J
5
optimisation results
The proposed algorithms for J
5
optimisation will be
assessed by comparing the obtained admittances with the
xed-structure admittances suggested in [2]. Fig. 7 presents
the optimum J
5
versus static stiffness and shows the extent
to which it is achieved by two xed-structure admittances.
The rst is a damper, which achieves the optimum up to
k
s
= 6.8 10
4
N/m and the other is a damper in parallel
with an inerter and achieves the optimum up to 10.210
4
N/m.
The algorithm for proper admittances was tested at k
s
=
510
4
N/m. For this value of k
s
the optimum is 2.6666
10
5
. The algorithm produced a rst-order admittance given
by,
K(s) = 6130 +
_
(762/s)
1
+12591
1
_
1
, (23)
that guarantees an upper bound of J
5
<2.666710
5
. The
obtained admittance is more complicated than the damper
proposed in [2] with optimal setting in the range 6030 Ns/m
5459
0 2 4 6 8 10 12
x 10
4
1
2
3
4
5
6
7
8
9
10
11
x 10
5
Static stiffness N/m
J
5
damper, inerter in parallel
damper
theoretical minimum of J
5
Fig. 7. Optimisation results for J
5
over xed-structure admittances.
c
opt
19940 Ns/m. An interesting observation is that
if we replace the stiffness of the spring in (23) with the
extreme values 0 or we recover approximately the optimal
damper range.
The algorithm for the non-proper admittances was tested
for k
s
= 10 10
4
N/m. For this value of k
s
the optimum
is 1.6666 10
5
. The algorithm produced a rst-order
admittance given by,
K(s) = 24010
0.004
s +0.004
+12008 +292s, (24)
that guarantees an upper bound of J
5
<1.666710
5
. The
rst-order lag is relatively small so it can be neglected
without causing a signicant degradation in J
5
. Thus the
suspension consists of a parallel connection of a spring
k
s
=1010
4
N/m, a damper c =12008 Ns/m and an inerter
b = 292 kg (Fig. 8). The optimal suspension proposed in
c b k
s
Fig. 8. Quarter-car suspension as a parallel connection of a spring, a
damper and an inerter.
[2] for this value of k
s
is in fact a damper in parallel with
an inerter with optimal values in the ranges,
11380 Ns/m c
opt
12287 Ns/m, 269 kg b
opt
337 kg.
The suspension obtained by the LMI optimisation is within
the above range. Moreover the LMI algorithm managed to
nd a second-order suspension that achieves the theoretical
minimum at the top range of k
s
(= 12 10
4
) N/m, which
could not be achieved with the xed-order admittances
considered in [2]. The resulting admittance was given by
K(s) = 51450
s
2
+14.11s +1302
s
2
+67.66s +4468
+423s (25)
which gave a rather complicated network when using the
Bott-Dufn realisation method [16].
VI. CONCLUSIONS
The problem of synthesis of positive real controllers is
formulated using matrix inequalities. Two local optimisa-
tion methods are proposed to solve the bilinear matrix
inequality problem in the context of suspension design
for a quarter-car model. The algorithms are successful in
obtaining previously found solutions when optimising over
xed-structure admittances. In the case of the performance
measure J
1
, which characterises the response of the sprung
mass due to road disturbances, the proposed algorithm
found alternative admittances that improve the performance
measure considerably.
REFERENCES
[1] M. C. Smith. Synthesis of mechanical networks: The inerter. IEEE
Transactions on Automatic Control, 47(10):16481662, 2002.
[2] M. C. Smith and F-C. Wang. Performance benets in passive vehicle
suspensions employing inerters. In 42nd Conference on Decision and
Control, pages 22582263, Hawaii, 2003. IEEE.
[3] S. Boyd, L. El Ghaoui, E. Feron, and B. Balakrishnan. Linear Matrix
Inequalities in System and Control Theory. SIAM, 1994.
[4] J. Wang and D. A. Wilson. Mixed GL
2
/H
2
/GH
2
control with pole
placement and its application to vehicle active suspension systems.
International Journal of Control, 74(13):13531369, 2001.
[5] J. C. Geromel and P. B. Gapski. Synthesis of positive real H
2
controllers. In Proceedings of the Conference on Decision and
Control, pages 28642869, Kobe, Japan, December 1996.
[6] T. Shimomura, Y. Yamasaki, and T. Fujii. LMI-based iterative
synthesis of strictly positive real H
2
controllers. In Proceedings
of American Control Conference, pages 332336, Chicago, Illinois,
USA, 2001.
[7] R. W. Newcomb. Linear Multiport Synthesis. New York:McGraw-
Hill, 1966.
[8] O. Brune. Synthesis of a nite two-terminal network whose driving-
point impedance is a prescribed function of frequency. J. Math. Phys.,
10:191236, 1931.
[9] R. Bott and R. J. Dufn. Impedance synthesis without use of
transformers. J. Appl. Phys., 20:816, 1949.
[10] M.C. Smith. Force-controlling mechanical device. patent pending,
Intl. App. No. PCT/GB02/03056, priority date: 4 July 2001.
[11] B.D.O. Anderson and S. Vongpanitlerd. Network Analysis and
Synthesis. Prentice-Hall, 1973.
[12] C. Scherer, P. Gahinet, and M. Chilali. Multi-objective output-
feedback control via LMI optimisation. IEEE Transactions on
Automatic Control, 42(7):896911, 1997.
[13] A. Hassibi, J. How, and S. Boyd. A path-following method for
solving BMI problems in control. In Proceedings of American
Control Conference, pages 13851389, San Diego, California, USA,
1999.
[14] J L ofberg. YALMIP 3, http://control.ee.ethz.ch/ joloef/yalmip.msql,
2004.
[15] P. Gahinet and P. Apkarian. A linear matrix inequality approach to
H