Optimal Control of Networked Control Systems

with Uncertain Time-Varying Transmission Delay

Michal Iz ak, Daniel G orges and Steven Liu (izak|goerges|sliu@eit.uni-kl.de)
Institute of Control Systems, University of Kaiserslautern, Germany
Introduction
Motivation
Control systems are increasingly realized as
networked control systems (NCS)
Communication leads to uncertain time-
varying transmission delay
k
=
CA
k
+
SC
k
Discrete-time models of NCS depend
nonlinearly on the uncertainty in
k
Standard robust control theory not applicable
Polytopic overapproximation
Previous work primarily focuses on stability
Network
u(t -
k
)
SC

k
P ZOH
h
x
c
(t)
K
J
Controller
Cost
Plant Actuator Sensor
CA

k
Contributions
Polytopic overapproximation of both the plant model and the cost function
Optimal control design based on a parameter-dependent Lyapunov function
Modeling
System Model and Cost Function
Continuous-time system
xc(t ) = Acxc(t ) + Bcu(t
k
)
Time-varying input delay
k
represents the transmission delay
Discretized for constant sampling period h / uncertain time-varying time delay
k
I =[, ]
with = nom

, = nom +

0 and h using ZOH

Discrete-time system
_
xc(k + 1)
u(k)
_
=
_
(h) (h) (h
k
)
0 0
__
xc(k)
u(k 1)
_
. .
x(k)
+
_
(h
k
)
I
_
u(k) (1)
with
(t ) = e
Act
, (t ) =
_
t
0
(t )dsBc
Continuous- and discrete-time cost function
J =
_

0
_
xc(t )
u(t (t ))
_
T
_
Q
1c
0
0 Q
2c
__
xc(t )
u(t (t ))
_
dt =

k=0
_
_
_
xc(k)
u(k 1)
u(k)
_
_
_
T
Q(
k
)
_
_
_
xc(k)
u(k 1)
u(k)
_
_
_ (2)
with
Q(
k
) =
_
_
_
_
_
h
0

T
(t )Q
1c
(t )dt
_

k
0

T
(t )Q
1c
(t )dt
_
h

T
(t )Q
1c
(t
k
)dt

k
0

T
(t )Q
1c
(t )+Q
2c
dt 0

_
h

T
(t
k
)Q
1c
(t
k
)+Q
2c
dt
_
_
_
_
Problem 1. For the discrete-time system (1) nd a control sequence u(k) such that the
discrete-time cost function (2) is robustly minimized for all sequences
k
I, i.e.
min
u(k)
max

k
I
J subject to (1).
Uncertain
k
I affects (1) and (2) in a nonlinear manner Use polytopic overapproximation
Polytopic Formulation
Taylor series expansion
(t ) =

q=0
A
q
c
q!
t
q
=
M

q=0
A
q
c
q!
t
q
. .
TE(t ,M)
+, (t ) =
_
t
0

q=0
A
q
c
q!
s
q
dsBc =
M+1

q=1
A
q1
c
q!
t
q
Bc
. .
TE(t ,M)
+
Q(k) =
_
_
_
_
Q11
_
k
0

T
TE
(t , M
a
12
)Q1cTE(t , M
b
12
)dt
_
h
k

T
TE
(t , M
a
13
)Q1cTE(t k, M
b
13
)dt

_
k
0

T
TE
(t , M
a
22
)Q1cTE(t , M
b
22
)+Q2cdt 0

_
h
k

T
TE
(t k, M
a
33
)Q1cTE(t k, M
b
33
)+Q2cdt
_
_
_
_
. .
QTE(k,M
a
12
,M
b
12
,...)
+Q
Truncation parts are matrix polynomials
Polynomial orders
multiplication addition of polynomial orders
integration increase by one
Truncation orders can be chosen differently
but polynomial orders must be identical
Matrix Polynomial Order Truncation Orders
M + 1 M = 2M + 2
Q12 M
a
12
+ M
b
12
+ 2 M
a
12
= M + 1, M
b
12
= M
Q13 M
a
13
+ M
b
13
+ 2 M
a
13
= M + 1, M
b
13
= M
Q22 M
a
22
+ M
b
22
+ 3 M
a
22
= M
b
22
= M
Q33 M
a
33
+ M
b
33
+ 3 M
a
33
= M
b
33
= M
Polynomial order

M = 2M + 3
The matrix polynomials can be enveloped by polytopes

TE
(
k
, M

) =

i =0

i
(
k
)
i
Q
TE
(
k
, M
a
12
, M
b
12
, . . .) =

i =0

i
(
k
)Q
i
Q(
k
)
Q( ) _

k Q
2
Q
1
Q
3
Q
TE
(
k
,...)
Q
Q( )
_

_
_
Discrete-time system with polytopic and additive norm-bounded uncertainty
x(k + 1) =
_
_
_
_
_
_

i =0

i
(
k
)
_
(h) (h)
i
0 0
_
. .
A
i
+
_
0
0 0
_
. .
A
_
_
_
_
_
_
x(k) +
_
_
_
_
_
_

i =0

i
(
k
)
_

i
I
_
. .
B
i
+
_

0
_
. .
B
_
_
_
_
_
_
u(k)
Discrete-time cost function with polytopic and additive norm-bounded uncertainty
J =

k=0
_
x(k)
u(k)
_
T
_
_

i =0

i
(
k
)Q
i
+ Q
_
_
_
x(k)
u(k)
_
Control
Closed-Loop Formulation
Full state feedback control law
u(k) =
_
Kx Ku
_
. .
K
_
xc(k)
u(k 1)
_
Discrete-time closed-loop system
x(k + 1) =
_
_
_
_
_
_
_

i =0

i
(
k
)
_
(h) +
i
Kx (h) +
i
(Ku I )
Kx Ku
_
. .

A
i
+
_
Kx (Ku I )
0 0
_
. .

A
_
_
_
_
_
_
_
x(k) (3)
Discrete-time closed-loop cost function
J =

k=0
x
T
(k)
_
_
_
_
_
_
_

i =0

i
(
k
)
_
I
K
_
T
Q
i
_
I
K
_
. .

Q(
k
)
+
_
I
K
_
T
Q
_
I
K
_
. .

Q
_
_
_
_
_
_
_
x(k) (4)
The additive norm-bounded uncertainty can be addressed by different approaches and will not
be considered in the following.
Problem 2. For the discrete-time closed-loop system (3) nd a feedback gain K such that the
discrete-time closed-loop cost function (4) is robustly minimized for all sequences
k
I, i.e.
min
K
max

k
I
J subject to (3).
Problem 2 is computationally intractable Minimize an upper bound instead
Upper Bound on the Cost Function
For a bounded value of the cost function (4) the quadratic function
V(x(k), k) = x
T
(k)P(
k
)x(k) with P(
k
) =

i =0

i
(
k
)P
i
, P
i
0
satisfying
V(x(k + 1), k + 1) V(x(k), k) < x
T
(k)

Q(
k
)x(k) (5)
for all
k
I and x(k) = 0 upper bounds the cost function, i.e.
max

k
I
J < V(x(0), 0) max

k
I
J < x
T
(0)P(
0
)x(0),
which follows by summing (5) over k = 0, . . . ,
The maximum cost degradation w.r.t. the nominal case is given by
P(
0
) Pnom (6)
where Pnom is the symmetric and positive denite solution of the ARE underlying the LQR
problem related to a constant sampling period h and a nominal transmission delay nom
Problem 2 can be redened to: min
K,P
i
subject to (5) and (6)
Control Design
Theorem 1. An upper bound on is obtained from the LMI optimization problem
max subject to > 0, Z
i
P
1
nom
,
_
_
_
_
_
_
G
T
+ G Z
i

Q
1
2
i
_
G
W
_
I
A
i
G + B
i
W 0 Z
j
_
_
_
_
_
_
0
for all i , j = 1, . . . ,

M in the variables G,W, Z
i
and . The corresponding robust feedback
matrix is K = WG
1
and the minimum upper bound results from = 1/ .
Example
Networked control of an active suspension system
xc(t )=
_
0 1
ks/ms bs/ms
_
xc(t ) +
_
0
1/ms
_
u(t )=
_
zs(t )
zs(t ) zr (t )
_
y(t )=
_
ks/ms bs/ms
1 0
_
xc(t ) +
_
1/ms
0
_
u(t )=
_
zs(t )
zs(t )zr (t )
_
with zr (0) = 0, ms = 315kg, ks = 29.5
kN
m
, bs = 1.5
kNs
m
Cost function
J =
_

0
q
1
zs(t )
2
+q
2
(zs(t ) zr (t ))
2
+rF(t )
2
dt
with q
1
= 1000, q
2
= 10, r = 0.02
Timing and design parameters
Constant sampling period h = 25ms
Nominal transmission delay nom = 12.5ms
Truncation order M = 4
Cost degradation for LQR follows from Theorem 1
by setting W = GKnom
0 0.1 0.2 0.3 0.4 0.5
1
1.5
2
2.5
3
3.5

/ h

Robust Controller
LQR K
nom
designed for
nom
Conclusions
Modeling and optimal control design method allowing to assess and handle the effect of an
uncertain time-varying transmission delay on the control performance in a quantitative way
Formulation for uncertain time-varying sampling period / switched interval uncertainty possible
2nd IFAC Workshop on Distributed Estimation and Control in Networked Systems (NecSys10) Annecy (France), 13-14 September 2010