Networked Control Systems: A Model-Based Approach

Networked Control Systems: A Model-Based
Approach
Luis A. Montestruque, Panos J. Antsaklis
University of Notre Dame, Notre Dame, IN 46556, U.S.A.
{lmontest, pantsakl}@nd.edu
Key words: control over networks, stabilizability, minimum feedback infor-
mation, model-based control
Summary. In this article a class of networked control systems called Model-Based
Networked Control Systems (MB-NCS) is considered. This control architecture uses
an explicit model of the plant in order to reduce the network trac while attempt-
ing to prevent excessive performance degradation. MB-NCS can successfully address
several important control issues in an intuitive and transparent way. In this arti-
cle the main results of this MB approach are described with examples. Specically,
conditions for the stability of state and output feedback continuous and discrete sys-
tems are derived under dierent scenarios that include delay compensation, constant
and time varying update times, non-linear plants and quantization of the feedback
signals. In addition, a performance measure for MB-NCS with noise inputs is intro-
duced.
1 Introduction
A networked control system (NCS) is a control system in which a data net-
work is used as feedback medium. NCS is an important research area; see for
example [16] and [15, 18, 19]. The use of networks as media to interconnect
the dierent components in an industrial control system is rapidly increasing.
The use of networked control systems poses, though, some challenges. One of
the main problems to be addressed when considering a networked control sys-
tem is the size of the bandwidth required by each subsystem. It is clear that
the reduction of bandwidth necessitated by the communication network in a
networked control system is a major concern. This can perhaps be addressed
by two methods: the rst method is to minimize the transfer of information
between the sensor and the controller/actuator; the second method is to com-
press or reduce the size of the data transferred at each transaction. Since
Luis Montestruque, Panos J. Antsaklis, Networked Control Systems: A Model-Based Approach, The
Handbook of Networked and Embedded Systems, Hristu and Levine, Eds., pp. 601-626, Birkhauser, 2005.
2 Luis A. Montestruque, Panos J. Antsaklis
shared characteristics among popular industrial networks are small transport
time and big overhead, using less bits per packet has small impact on the
overall bit rate. So reducing the rate at which packets are transmitted brings
better benets than data compression in terms of bit rate used. In this pa-
per, we consider the problem of reducing the packet rate of a NCS using a
novel approach called Model-Based NCS (MB-NCS). The MB-NCS architec-
ture makes explicit use of knowledge about the plant dynamics to enhance the
performance of the system. Model-Based Networked Control Systems (MB-
NCS) were introduced in [11].
In Section 2 the basic MB-NCS setup is introduced for continuous plants.
Stability of MB-NCS with no quantization and periodic transmissions are
considered. In Section 3 a performance measure is introduced for the previ-
ously presented MB-NCS. The stability of MB-NCS with time-varying and
stochastic transmission times is studied in Section 4. Quantization schemes
are considered in Section 5. Finally, stability for a class of nonlinear systems
is studied in Section 6.
2 Stability of Continuous Linear MB-NCS with
Constant Update Times
2.1 State Feedback MB-NCS
We consider the control of a continuous linear plant where the state sensor
is connected to a linear controller/actuator via a network. In this case, the
controller uses an explicit model of the plant that approximates the plant
dynamics and makes possible the stabilization of the plant even under slow
network conditions.
Fig. 1. Proposed conguration of networked control system.
We will concentrate on characterizing the transfer time between the sensor
and the controller/actuator, which is the time between information exchanges.
Networked Control Systems: A Model-Based Approach 3
The plant model is used at the controller/actuator side to recreate the plant
behavior so that the sensor can delay sending data since the model can provide
an approximation of the plant dynamics. The main idea is to perform the
feedback by updating the models state using the actual state of the plant that
is provided by the sensor. The rest of the time the control action is based on a
plant model that is incorporated in the controller/actuator and is running open
loop for a period of h seconds. The control architecture is shown in Figure 1.
Our approach is novel in that it incorporates a model of the plant, the
state of which is updated at discrete intervals by the plants state. We present
a necessary and sucient condition for stability that results in a maximum
transfer time. We make use of standard stability denition such as the ones
found in [1].
If information for all the states is available, then the sensors can send this
information through the network to update the models vector state. For our
analysis we will assume that the compensated model is stable and that the
transportation delay is negligible. We will assume that the frequency at which
the network updates the state in the controller is constant. The idea is to nd
the smallest frequency at which the network must update the state in the
controller, that is, an upper bound for h, the update time. Usual assumptions
in the literature include requiring a stable plant or in the case of a discrete
controller, a smaller update time than the sampling time. Here we do not
make any of these assumptions and the plant may be unstable.
Consider the control system of Figure 1 where plant is given by x = Ax +
Bu, the plant model by

x =

A x +

Bu, and the controller by u = K x. The
state error is dened as e = x x, and represents the dierence between the
plant state and the model state. The modeling error matrices

A = A

A and
B = B

B represent the dierence between the plant and the model. We will
now express the combined state z(t) =
_
x(t)
T
e(t)
T

T
in terms of the initial
condition x(t
0
). A necessary and sucient condition for stability of the state
feedback MB-NCS will then be presented. Dene =
_
A+BK BK
A+

BK

A

BK
_
.
Proposition 1 [13] The State Feedback MB-NCS with initial conditions
z(t
0
) =
_
x(t
0
)
T
0
T

T
= z
0
, has the following response:
z(t) = e
(tt
k
)
__
I 0
0 0
_
e
h
_
I 0
0 0
__
k
z
0
,
for t [t
k
, t
k+1
), with t
k+1
t
k
= h
Theorem 1 [13] The State Feedback MB-NCS is globally exponentially stable
around the solution z =
_
x
T
e
T

T
= 0 if and only if the eigenvalues of
M =
_
I 0
0 0
_
e
h
_
I 0
0 0
_
are strictly inside the unit circle.
One can gain insight into the closed loop system behavior by noticing that
the nonzero eigenvalues of M =
_
I 0
0 0
_
e
h
_
I 0
0 0
_
are exactly the eigenvalues
of another matrix:
N = e
(

A+

BK)h
+e
Ah
_
h
0
e
A
(

A+

BK)e
(

A+

BK)
d (1)
This can be shown either directly or by using the Laplace transform. The
second method involves replacing the update time variable h by the time vari-
able t, transforming the expression using the Laplace transform. The trans-
formed expression can then be easily manipulated. After isolating the upper
left block matrix, inverse Laplace transform can be used to return to time
domain [10].
First we observe that the eigenvalues of the compensated model ap-
pear in the rst term of N. We can see the term = e
Ah
_
h
0
e
A
(

A +
BK)e
(

A+

BK)
d as a perturbation on the desired eigenvalues, that is, the
eigenvalues of the compensated model. Even if the eigenvalues of the original
plant were unstable the perturbation can be made small enough by having
h and

A+

BK small and thus minimizing their impact over the eigenvalues of
the compensated plant. We also observe that if the update time h is driven to
zero, then = 0. Also it is possible to make = 0 by having a plant model
that is exact.
Example 1 Consider the following unstable plant (double integrator): A =
_
0 1
0 0
_
, B =
_
0
1
_
We will use the state feedback controller given by u = Kx
with K =
_
1 2
. Usually it is assumed that the actuator/controller will

hold the last value received from the sensor until the next time the sensor trans-
mits and a packet is received. Under this assumption the controller/actuators
model acts as a zero order hold when updated. We will rst analyze this situa-
tion. To do so, we will transform the plant model so that it holds the last state
update presented to it by the network. This is given by:

x =

A x +

Bu , with
A =
_
0 0
0 0
_
and

B =
_
0 0
T
. Now we need to search for the biggest h such
that
_
I 0
0 0
_
e
h
_
I 0
0 0
_
has its eigenvalues inside the unit circle. To do so we
plotted the maximum eigenvalue magnitude versus the update time. The plot
is shown in Figure 2.
From Figure 2 we see that the condition for stability is to have h < 1
second. In fact the test matrix M will have one eigenvalue with magnitude
1 for h = 1 second and the system will be marginally stable. If we use the
results by [19] we would have obtained that, in order to stabilize the system,
we would need to have h < 2.1304 10
4
seconds, which is very conservative.
Simulations of the system with update times of 2.1304 10
4
and 0.5 seconds
are shown in Figure 3. Note that the plant was initialized with an initial
condition of
_
1 1
T
.
0 0.5 1 1.5
0.5
1
1.5
2
2.5
m
a
x

e
i
g
e
n
v
a
l
u
e

m
a
g
n
i
t
u
d
e
update times h
Fig. 2. Maximum eigenvalue magnitude of the test matrix M versus the update
time h.
0 5 10 15
1
0.5
0
0.5
1
1.5
time (sec)
p
l
a
n
t

s
t
a
t
e
h=0.0002 sec
0 5 10 15
1
0.5
0
0.5
1
1.5
time (sec)
p
l
a
n
t

s
t
a
t
e
h=0.5 sec
Fig. 3. System response for h = 2.1304 10
4
and h = 0.5 sec.
It is clear that the performance obtained with h = 0.5 seconds is not too
dierent to the one obtained with h < 2.130410
4
seconds, but the dierence
in the amount of bandwidth used is large. If we were to use Ethernet that
has a minimum message size of 72bytes (including preamble bits and start
of delimiter elds) the data rate would be 2.7Mbits/sec for the case of h <
2.1304 10
4
seconds, and 1.2Kbits/sec for the case of h = 0.5 seconds.
Example 2 In real applications uncertainties can frequently be expressed as
tolerances over the dierent measured parameter values of the plant. This
can be mapped into structured or parametric uncertainties on the state space
matrices. Next an example is given on how the Theorem 1 can be applied if
two entries on the A matrix of the model can vary within a certain interval.
model:

A =
_
0 1
0 0
_
,

B =
_
0
1
_
;
plant: A =
_
0 1 + a
12
0 + a
21
0
_
, B =
_
0
1
_
;
with a
12
= [0.5, 0.5], a
21
= [0.5, 0.5]
controller: K =
_
1 2
.
The system will now be tested for an update time of h = 2.5 seconds. The
following contour plot in Figure 4 represents the magnitude of the maximum
eigenvalue of the test matrix M as a function of the (1,2) and (2,1) entries
of A parameters a
12
and a
21
. Here the contour of height equal to one is the
relevant one to stability. It is easy to isolate the stable and unstable regions in
the uncertainty parameter plane. The stable region is between the lines labeled
as 1.
0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5
0.5
0.4
0.3
0.2
0.1
0
0.1
0.2
0.3
0.4
0.5
a
12
a
2
1
0
.2
0
.2
0
.2
0
.2
0
.2
0
.2
0
.4
0
.4
0
.4
0
.4
0
.4
0
.4
0
.6
0
.6
0
.6
0
.6
0
.6
0
.8
0
.8
0
.8
0
.8
0
.8
1
1
1
1
1
.2
1
.2
1
.2
1
.4
1
.4
1
.4
1
.6
1
.6
1
.6
1
.8
1
.8
1
.8
2
2
Fig. 4. Contour Plot Maximum Eigenvalue Magnitude vs Model Error.
2.2 Output Feedback MB-NCS
We have been considering plants where the full state vector is available at the
output. We now extend our approach to include plants where the state is not
directly measurable. In this case, in order to obtain an estimate of the plant
state vector, a state observer is used. It is assumed that the state observer is
collocated with the sensor. Again, we use the plant model,

x =

A x +

Bu, to
design the state observer. The observer uses the plant output and generates
a copy of the plant input applied by the controller. See Figure 5.
Having the sensor carry the computational load of an observer is justied
by the fact that typically sensors that can be connected to a network have an
embedded processor inside. This processor is usually in charge of performing
the sampling, ltering and implementing the network layer services required
to connect to the network. Ishii and Francis give a similar justication in
Fig. 5. Proposed conguration of an output feedback networked control system.
[17]. In their approach an observer is placed at the output of the plant to
reconstruct the state vector. The result is then quantized and sent over the
network to the controller.
The observer has as inputs the output and input of the plant. In the
implementation of this set up, to acquire the input of the plant, which is at
the other side of the communication link, the observer can have a version of
the model and controller, and knowledge of the update time h. In this way, the
output of the controller, that is the input of the plant, can be simultaneously
and continuously generated at both ends of the feedback path with the only
requirement that the observer makes sure that the model has been updated.
This last requirement ensures that both the controller and the observer are
synchronized. Handshaking provided by most network protocols can be used.
Again, we use the plant model,

x =

A x+

Bu, to design the state observer.
See Figure 5. The observer has the form of a standard state observer with
gain L. In summary, the system dynamic equations are for t [t
k
, t
k+1
):
Plant: x = Ax +Bu, y = Cx +Du
Model:

x =

A x +

Bu, y =

C x +

Du
Controller: u = K x
Observer:

x = (

AL
C) x +
_

B L

D L
_
u
T
y
T

T
We now proceed in a similar way as in the previous case of full feedback. There
will be an update interval h, after which the observer updates the controllers
model state x with its estimate x. Dene an error e that will be the dierence
between the controllers model state and the observers estimate: e = x x.
Also dene the modeling error matrices in the same way as before:

A = A

A,
B = B

B,

C = C

C , and

D = D

D. Dene z =
_
x
T
x
T
e
T

T
, and
o
=
_
_
A BK BK
LC

AL
C +

BK +L

DK
BK L

DK
LC L

DK L
C

AL

DK
_
_
Theorem 2 [13] The Output Feedback MB-NCS is globally exponentially sta-
ble around the solution z =
_
x
T
x
T
e
T

T
= 0 if and only if the eigenvalues
of M
o
=
_
_
I 0 0
0 I 0
0 0 0
_
_
e
o
h
_
_
I 0 0
0 I 0
0 0 0
_
_
are inside the unit circle.
The eigenvalues of the test matrix M
o
can be studied in a similar fashion as
in the state feedback case. By replacing h by t, applying the Laplace transform
and isolating the nonzero upper left block of M
o
we obtain:
P =
_
_
_
sI

A

BK
_
1
+
1
(sI A)
1
BK
_
sI

A

BK
_
1
3
_
sI

A+L
C
_
1
+
2
_
_
with
1
= (sI A)
1
(

A+

BK)(sI

A

BK)
1
2
= (sI

A+L
C)
1
_
BK
_
AL
C
_
(sI A)
1
BK
_
(sI

A

BK)
1
3
= (sI

A+L
C)
1
_
AL
C +

BK
_
AL
C
_
(sI A)
1
_
A+

BK
__
(sI

A

BK)
1
It is clear that if

A 0 ,

B 0 , and

C 0 then
1
0 ,
2
0 , and
3
0. By doing so we obtain:
P
L
= lim
0
P =
_
_
_
sI

A

BK
_
1
(sI A)
1
BK
_
sI

A

BK
_
1
0
_
sI

A+L
C
_
1
_
_
From here it can be seen that by making the error between the plant and
the model zero, the systems compensated eigenvalues would be placed at
eig
_
e
(

A+

BK)h
_
eig
_
e
(

AL
C)h
_
. Similarly to the full state feedback case,
there are perturbation matrices
1
,
2
, and
3
that can be made small by
reducing the error between the actual plant and the model dynamic equations.
The perturbation is over a matrix that has the eigenvalues of the compensated
model. Finally note that the separation principle of classic control where con-
troller and observer can be designed independently cannot be used here.
2.3 State Feedback MB-NCS with Network Induced Delays
Previously we assumed that the network delays were negligible. This is usually
true for plants with slow dynamics relative to the network bandwidth. When
this is not the case the network delay cannot be neglected. Network delays
can occur for many reasons. There are three important delay sources:
Processing time.
Media access contention.
Propagation and transmission time.
The rst one, processing time, occurs at both ends of the communication
channel. On the transmitter, the processing time is the time elapsed between
the time the transmitting process makes the request to the operating system
to transmit a message, to the time the message is ready to be sent. And on
the receiver this is the time interval that occurs between the last bit of the
message is received by the receiver, and when the time the message is delivered
by the operating system to the receiver process.
The media access contention time is the time the transmitter has to wait
until the communication channel is not busy. This is usually the case when
several transmitters have to share the same media.
The propagation and transmission time is the time the message takes to
be placed on the network media and to travel through the network to reach
the receiver. In local area networks the time the message takes to travel or
propagate through the media is small in comparison to wide area networks or
internetworks like the Internet. The time the message takes to be placed on
the network depends on the size of the message and the baud rate.
If the control network is a local area network, as is common practice in in-
dustry, the propagation and transmission time can be established beforehand
with good accuracy. Similarly for the processing time. If real time operating
systems are used the processing time can be accurately calculated. Finally,
media access contention delay can be xed with the use of a communication
protocol with scheduling. Fast data communication networks like Token Ring,
Token Bus, and ArcNet fall into this category. Industry oriented control net-
works like Foundation Fieldbus also implements a scheduler through its LAS
or Link Active Scheduler. Even the inherently non-deterministic Ethernet has
addressed the problem of not having a specied contention time with the
so-called Switched Ethernet.
In conclusion, most of these delays can at least be bounded if the network
conditions are appropriate.
In the following, we extend our results to include the case were transmis-
sion delay is present. We will assume that the update time h is larger than
the delay time . As before we will assume that the update time h and delay
are constant. We will present here the case of full state feedback systems.
So, at times kh the sensor transmits the state data to the con-
troller/actuator. This data will arrive seconds later. Therefore, at times
kh the controller/actuator receives the state vector value x(kh ). The
main idea is to use the plant model in the controller/actuator to calculate
the present value of the state. After this, the state approximate obtained can
be used to update the controllers model as in previous setups. The system is
depicted in Figure 6.
The Propagation Unit uses the plant model and the past values of the
control input u(t) to calculate an estimate of actual state x(kh) from the
Fig. 6. Proposed conguration of a state feedback networked control system in the
presence of network delays .
received data x(kh ). This estimate is then used to update the model that
with the controller will generate the control signal for the plant.
The system is described by the following equations:
Plant: x = Ax +Bu
Model:

x =

A x +

Bu
Controller: u = K x, t [t
k
, t
k+1
)
Propagation Unit:

x =

A x +

Bu, t [t
k+1
, t
k+1
]
Update law: x x, t = t
k+1
and x x, t = t
k+1
(2)
We dene the errors e = x x and e = x x. We also make the following
denitions:
A = A

A
B = B

B
,
d
=
_
_
A+BK BK BK
A+

BK

A

BK
BK
0 0

A
_
_
, z =
_
_
x
e
e
_
_
Theorem 3 [13] The State Feedback MB-NCS with networked induced delay
is globally exponentially stable around the solution z =
_
x
T
e
T
e
T

T
= 0
if and only if the eigenvalues of M
T
=
_
_
I 0 0
0 I 0
0 0 0
_
_
e
_
_
I 0 0
0 0 0
0 I I
_
_
e
T
(h)
are
inside the unit circle.
It is interesting to note that the results on Theorem 3 can be seen as a
generalization of Theorem 1. This can be shown by driving to zero.
Example 3 Here we present a numeric example with the same plant we have
been using before A =
_
0 1
0 0
_
, B =
_
0
1
_
, with randomly generated plant model
A =
_
0.3444 0.9225
0.3089 0.3560
_
,

B =
_
0.0098
1.3159
_
, and controller K =
_
1 2
.
Figure 7 shows the plots for the maximum eigenvalue magnitude as a function
of h for 3 dierent values of . The maximum value h can have to preserve
stability is reduced when is increased.
0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
0.8
0.85
0.9
0.95
1
1.05
1.1
1.15
m
a
x

e
ig
e
n
v
a
lu
e

m
a
g
n
it
u
d
e
update times h
Fig. 7. Maximum eigenvalue of test matrix M versus the update times h for = 0,
0.25, and 0.50 sec.
3 A Performance Index for Linear MB-NCS with
Constant Update Times
The performance characterization of Networked Control Systems under dier-
ent conditions is also of mayor concern together with stability. It is clear that,
since the MB-NCS is h-periodic, there is no transfer function in the normal
sense whose H2 norm can be calculated [2]. For LTI systems the H2 norm can
be computed by obtaining the 2-norm of the impulse response of the system.
We will extend this denition to specify an H2 norm, or more accurately, to
dene an H2-like performance index [2]. We will call this performance index
Extended H2 Norm. We will study the Extended H2 Norm of the MB-NCS
with output feedback studied in subsection 2.2. A disturbance signal w and a
performance or objective signal z are included in the setup.
We will start by dening the system dynamics:
Plant Dynamics: Observer Dynamics:
x = Ax +B
1
w +B
2
u

x = (

AL
C
2
) x + (

B
2
L

D
22
)u +Ly
z = C
1
x +D
12
u Model Dynamics:
y = C
2
x +D
21
w +D
22
u

x =

A x +

B
2
u
Controller:
u = K x
(3)
Dene the following:
=
_
_
A B
2
K B
2
K
LC
2

AL
C
2
+

B
2
K +L

D
22
K
B
2
K L

D
22
K
LC
2
L

D
22
K L
C
2

AL

D
22
K
_
_
M(h) =
_
_
I 0 0
0 I 0
0 0 0
_
_
e
h
, B
N
=
_
_
B
1
LD
21
LD
21
_
_
, C
N
=
_
C
1
D
12
K D
12
K
Theorem 4 [21] The Extended H2 Norm, G

xh2
, of the Output Feedback
MB-NCS described in Equation (3) is given by G
xh2
= trace
_
B
T
N
XB
N
_
where X is the solution of the discrete Lyapunov equation M(h)
T
XM(h)
X + W
o
(0, h) = 0 and W
o
(0, h) is the observability Gramian computed as
W
o
(0, h) =
_
h
0
e
T
t
C
T
N
C
N
e
t
dt.
Note that the observability Gramian can be factorized as C
aux
T
C
aux
=
_
h
0
e
T
t
C
T
N
C
N
e
t
dt. This allows one to compute the Extended Norm as the
regular H
2
norm of a discrete LTI system.
Corollary 1 [21] Dene C
aux
T
C
aux
=
_
h
0
e
T
t
C
T
N
C
N
e
t
dt and the auxiliary
discrete system G
aux
with parameters A
aux
= M(h), B
aux
= B
N
, C
aux
, and
D
aux
= 0, then the following holds.
G
xh2
= G
aux
2
Example 4 We now present an example using a double integrator as the
plant. Let the plant dynamics be given by: A =
_
0 1
0 0
_
, B
1
=
_
0.1
0.1
_
,
B
2
=
_
0 1
T
, C
1
=
_
0.1 0.1
, C
2
=
_
1 0
, D
11
= 0 , D
12
= 0.1 ,
D
21
= 0.1 , D
22
= 0 . We will use the state feedback controller K =
_
1 2
.
A state estimator with gain L =
_
20 100
T
is used to place the state observer
eigenvalues at 10. We will use a plant model with the following parame-
ters:

A =
_
0.1634 0.8957
0.1072 0.1801
_
,

B
2
=
_
0.1686
1.0563
_
,

C
2
=
_
0.8764 0.1375
D
22
= 0.1304 .
In Figure 8 we plot the extended H2 norm of the system as a function of
the updates times. Note that as the update time of the MB-NCS approaches
zero, the value of the Extended H2 norm approaches the H2 norm of the non-
networked compensated system. Also note the performance degradation as the
update time h is increased.
4 Stability of MB-NCS with Time-Varying Update
Times
In this section we relax our assumption that the update times h(k) are con-
stant. Here we study the state feedback MB-NCS shown in Figure 1. The
0 1 2 3 4 5 6
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
|
|
G
z
w
|
|
2
x
h
2
update times h (sec)
0.0394
Fig. 8. Extended H2 norm of the system as a function of the update times.
packets transmitted by the sensor contain the measured value of the plant
state and are used to update the plant model on the actuator/controller node.
These packets are transmitted at times t
k
. We dene the update times as the
times between transmissions or model updates: h(k) = t
k+1
t
k
. Previously
we made the assumption that the update times h(k) are constant. This might
not always be the case in applications. The transmission times of data pack-
ets from the plant output to the controller/actuator might not be completely
periodic due to network contention and the usual non-deterministic nature of
the transmitter task execution scheduler. Soft real time constraints provide a
way to enforce the execution of tasks in the transmitter microprocessor. This
allows the task of periodically transmitting the plant information to the con-
troller/actuator to be executed at times t
k
that can vary according to certain
probability distribution function. This translates into an update time h(k)
that can acquire a certain value according to a probability distribution func-
tion. Most work on networked control systems assumes deterministic commu-
nication rates [17, 20] or time-varying rates without considering the stochastic
behavior of these rates [9, 19]. Little work has concentrated in characterizing
stability or performance on a networked control system under time-varying,
stochastic communication.
We rst study the stability properties of the feedback MB-NCS assuming
that the update times can take any values in an interval [h
min
, h
max
]. In
this case we will assume that we dont have any statistical knowledge about
the update times. We analyze the stability properties of this system using
Lyapunov techniques. This is the strongest type of stability presented and will
provide a rst cut on the characterization of the stability properties perhaps
for comparison purposes.
Next, two types of stochastic stability are discussed, namely Almost Sure
or Probability-1 Asymptotic Stability and Mean Square or Quadratic Asymp-
totic. The rst one is the one that mostly resembles deterministic stability
[3]. Mean square stability is attractive since it is related to optimal control
problems such as the LQR.
Two dierent types of time-varying transmission times are considered for
each case of stochastic stability criterion. The rst assumes that the trans-
mission times are independent identically distributed with probability distri-
bution function that may have support for innite update times. The second
type of stochastic update time assumes that the transmission times are driven
by a nite Markov chain. Both models are common ways of representing the
behavior of network transmission and scheduler execution times.
4.1 Lyapunov Stability of MB-NCS
The stability criterion derived in this section is the strongest and most con-
servative stability criterion. It is based on the well-known Lyapunov second
method for determining the stability of a system. We will assume that the
properties of h(k) are unknown but h(k) is contained within some interval.
This criterion is not stochastic but provides a rst approach to stability for
time-varying transmission times NCS.
Denition 1 The equilibrium z = 0 of a system described by z = f (t, z)
with initial condition z(t
0
) = z
0
is Lyapunov asymptotically stable at large
(or globally) if for any > 0 there exists > 0 such that the solution of
z = f (t, z) satises z (t, z
0
, t
0
) < , t > t
0
and lim
t
z (t, z
0
, t
0
) = 0
whenever z
0
< .
Theorem 5 [14] The State Feedback MB-NCS is Lyapunov Asymptotically
Stable for h [h
min
, h
max
] if there exists a symmetric positive denite matrix
X such that Q = X MXM
T
is positive denite for all h [h
min
, h
max
]
where M =
_
I 0
0 0
_
e
h
_
I 0
0 0
_
.
Theorem 5 may be used to derive an interval [h
min
, h
max
] for h for which
stability is guaranteed. It is clear that the range for h, that is the interval
[h
min
, h
max
], will vary with the choice of X. Another observation is that the
interval obtained this way will always be contained in the set of constant
update times for which the system is stable (as derived using Theorem 1).
That is, an update time contained in the interval [h
min
, h
max
] will always be
a stable constant update time.
Several ways of obtaining the values for h
min
and h
max
can be used. One is
to rst x the value of Q, obtain the solution X of the Lyapunov equation in
Theorem 5 for a value of h known to be stable. Then, using this value of X,
the expression X MXM
T
can be evaluated for positive deniteness. This
can be repeated for all the values of h known to stabilize the system to obtain
the widest interval [h
min
, h
max
].
4.2 Almost Sure or Probability-1 Asymptotic Stability
We will use the denition of almost sure asymptotic stability [3] that pro-
vides a stability criterion based on the sample path. This stability denition
resembles more the deterministic stability denition [4], and is of practical
importance. Since the stability condition has been relaxed, we expect to see
less conservative results than those obtained using the Lyapunov stability con-
sidered in the previous section. We now dene Almost Sure or Probability-1
Asymptotic stability.
Denition 2 The equilibrium z = 0 of a system described by z = f (t, z)
with initial condition z(t
0
) = z
0
is almost sure (or with probability-1) asymp-
totically stable at large (or globally) if for any > 0 and > 0 the solution of
z = f (t, z) satises lim
P
_
sup
t
z (t, z
0
, t
0
) >
_
= 0 whenever z
0
< .
This denition is similar to the one presented for deterministic systems
in Denition 1. We will examine the conditions under which the full state
feedback continuous networked system in Figure 1 is stable.
MB-NCS with Independent Identically Distributed Transmission
Times
Here we will assume that the update times h(k) are independent identically
distributed (iid) with probability distribution function F(h). We now present
the conditions under which the full state feedback MB-NCS with iid update
times is asymptotically stable with probability-1. We will use a technique
similar to lifting [2] to obtain a discrete linear time invariant representation
of the system. It can be observed that the system can be described by
k+1
=
k
k
, with
k
L
2e
and
k
= z(t +t
k
), t [0, h
k
). (4)
Here L
2e
stands for the extended L
2
. It can be shown that the operator
k
can be represented as:
(
k
)(t) = e
t
_
I 0
0 0
_
h(k)
_
0
( h(k))()d (5)
With (t) representing the impulse function. Now we can restate the de-
nition on almost sure stability or probability-1 stability given in Denition 2
to better t the equivalent system representation (4).
Denition 3 The system represented by (4) is almost sure stable or stable
with probability-1 if for any > 0 and > 0 the solution of
k+1
=
k
k
satises: lim
P
_
sup
k
k
(t
0
, z
0
)
2,[0,t
k
]
>
_
= 0 whenever z
0
< . Where
the norm
2,[0,h(k)]
is given by
k
2,[0,h(k)]
=
_
h(k)
_
0
k
()
2
d
_
1/2
.
This denition allows us to study almost sure stability of systems such as
(4) when the probability distribution function for update times has innite
support. Based on this denition the following result can now be shown.
Theorem 6 [14] The State Feedback MB-NCS, with update times h(j) in-
dependent identically distributed random variable with probability distribu-
tion F(h) is globally almost sure (or with probability-1) asymptotically stable
around the solution z =
_
x
T
e
T

T
= 0 if N = E
_
_
e
2 ()h
1
_
1/2
_
<
and the expected value of the maximum singular value of the test matrix M,
E [M] = E [
M
], is strictly less than one, where M =
_
I 0
0 0
_
e
h
_
I 0
0 0
_
.
Note that the condition may give conservative results if applied directly
over the test matrix. To avoid this problem and make the condition tighter
we may apply a similarity transformation over the test matrix M.
The condition on the matrix N ensures that the probability distribution
function for the update times F(h) assigns smaller occurrence probabilities to
increasingly long update times, that is F(h) decays rapidly. In particular we
observe that N can always be bounded if there exists h
m
such that F(h) = 0
for h larger than h
m
. We can also bound the expression inside the expecta-
tion to obtain E
_
_
e
2 ()h
1
_
1/2
_
< E
_
e
()h
and formulate the following

corollary:
Corollary 2 The State Feedback MB-NCS, with update times h(j) that are
independent identically distributed random variable with probability distribu-
tion F(h) is globally almost sure (or with probability-1) asymptotically sta-
ble around the solution z =
_
x
T
e
T

T
= 0 if T = E
_
e
()h
< and
the expected value of the maximum singular value of the test matrix M,
E [M] = E [
M
], is strictly less than one.
Note that Corollary 1 condition T = E
_
e
()h
< is automatically
satised if the probability distribution function F (h) doesnt have innite
support. It otherwise indicates that F (h) should roll o fast enough as to
counteract the growth of Ms maximum singular value as h increases.
Example 5 We use the unstable double integrator plant. We now assume
that h(k) is a random variable with a uniform probability distribution func-
tion U (0.5, h
max
). The plot of the expected maximum singular value of a
similarity transformation of the original test matrix M is shown in Figure 9.
The similarity transformation used here was one that diagonalizes the matrix
M for h = 1.
We see that the maximum value for h
max
is around 1.3 seconds (maximum
constant update time for stability is h = 1 second.) So we see that, the double
integrator with uniformly distributed update time between 0.5 and 1.3 seconds
is stable, while the same system but with a constant update time of 1 second
is unstable. This also represents an improvement over the result that we may
have obtained by using the previously discussed Lyapunov Stability condition
in which the maximum update time obtainable would have been less than 1sec.
0.5 1 1.5 2
0.6
0.8
1
1.2
1.4
1.6
1.8
2
2.2
E
[
|
|
M
|
|
]

f
o
r

h
~
U
(
0
.
5
,
h
m
a
x
)
h
max
(sec)
Fig. 9. Average Maximum Singular Value for h U (0.5, h
max
) as a function of
h
max
, zero dynamics plant model.
1 2 3 4 5 6
0.5
0.6
0.7
0.8
0.9
1
1.1
1.2
E
[
|
|
M
|
|
]

f
o
r

h
~
U
(
0
.
5
,
h
m
a
x
)
h
max
(seconds)
Fig. 10. Average Maximum Singular Value for h U (0.5, h
max
) as a function of
h
max
, improved plant model.
The advantage of using a model-based approach resides in its ability to
reduce the amount of bandwidth required. The previous example shows sta-
bility conditions for a model that represents a zero order hold, that is, the
control value is kept constant until the next update time. We will now show
the same plots for a model that better resembles the plant; this was done
by randomly perturbing the plant matrices. The plant model matrices are:
A =
_
0.0844 0.9353
0.0476 0.0189
_
,

B =
_
0.0871
1.0834
_
. Figure 10 shows that stability is
maintained for update times that have uniform distribution with a max up-
date time of 5.5 seconds. This shows that improved knowledge over the plant
dynamics can translate into a signicant improvement in terms of stability.
MB-NCS with Markov Chain-Driven Transmission Times
In certain cases it is appropriate to represent the dynamics of the update times
as driven by a Markov chain. A good example of this is when the network
experiences trac congestion or has queues for message forwarding. We now
present a stability criterion for the model-based control system in which the
update times h(k) are driven by a nite state Markov chain. Assume that the
update times can take a value from a nite set:
h(k) {h
1
, h
2
, ..., h
N
} and h
i
= , i [1, N] (6)
Let us represent the Markov chain process by {
k
}with state space {1, 2, ..., N}
and transition probability matrix and NN matrix with elements p
i,j
and
initial state probability distribution
0
=
_
1

2
...
N
T
. The transition
probability matrix entries are dened as p
i,j
= P{
k+1
= j|
k
= i}. We can
now represent the update times more appropriately as h(k) = h
k
.
A sucient condition for the almost sure stability of the system under
Markovian jumps is given in the following theorem.
Theorem 7 [14] The State Feedback MB-NCS with update times h(k) =
h
k
= driven by a nite state Markov chain{
k
} with state space {1, 2, ..., N}
and transition probability matrix with elements p
i,j
and initial state proba-
bility distribution
0
=
_
1

2
...
N
T
is globally almost sure asymptotically
stable around the solution z =
_
x
T
e
T

T
= 0 if the matrix T has all its
eigenvalues inside of the unit circle, where:
T =
_
_
M|
h
1
0 0 0
0 M|
h
2
0 0
0 0 ... 0
0 0 0 M|
h
N
T
; M|
h
i
=
_
I 0
0 0
_
e
h
i
_
I 0
0 0
_
If is irreducible it follows that, since M is non-negative, T is also
irreducible. Then it can be shown using the Perron-Frobenius theorem as in
[6], that Ts maximum magnitude eigenvalue is real and sometimes referred
to as the Perron-Frobenius eigenvalue.
4.3 Mean Square or Quadratic Asymptotic Stability
We now dene a dierent type of stability, namely Mean Square Asymptotic
Stability:
Denition 4 The equilibrium z = 0 of a system described by z = f (t, z) with
initial condition z(t
0
) = z
0
is mean square stable asymptotically stable at large
(or globally) if the solution of z = f (t, z) satises: lim
t
E
_
z (t, z
0
, t
0
)
2
_
= 0
A system that is mean square stable will have the expectation of system
states converging to zero with time in the mean square sense. This denition
of stability is attractive since many optimal control problems use the squared
norm in their formulations. We will study the two cases studied in the previous
section under this new stability criterion.
MB-NCS with Independently Identical Distributed Transmission
Times
We present the conditions under which the state feedback networked control
system is mean square stable, we also discuss how these conditions relate to
the ones for probability-1 stability.
Theorem 8 [14] The State Feedback MB-NCS with update times h(j) in-
dependent identically distributed random variable with probability distribu-
tion F(h) is globally mean square asymptotically stable around the solution
z =
_
x
T
e
T

T
= 0 if K = E
_
_
e
()h
_
2
_
< and the maximum singu-
lar value of the expected value of M
T
M,
_
_
E
_
M
T
M
_
_
=
_
E
_
M
T
M
_
, is
strictly less than one, where M =
_
I 0
0 0
_
e
h
_
I 0
0 0
_
.
We note the similarity between the conditions given by Theorems 6 and
8. For the rst one we require the expectation of the maximum singular value
of the test matrix to be less than one. While for the second stability criterion
it is required to have the maximum singular value of the expectation of the
test matrix (multiplied by its transpose) to be less than one.
MB-NCS with Markov Chain-Driven Transmission Times
The type of stability criteria above depend on our ability to nd appropriate
P(i) matrices. Several other results in jump system stability [5, 7] can be
extended to obtain other conditions on stability of networked control systems.
Note though, that most of the results available in the literature deal with
similar but not identical type of systems.
Theorem 9 The State Feedback MB-NCS with update times h(k) = h
k
=
driven by a nite state Markov chain {
k
} with state space {1, 2, ..., N}
and transition probability matrix with elements p
i,j
is globally mean square
asymptotically stable around the solution z =
_
x
T
e
T

T
= 0 if there exists
positive denite matrices P(1), P(2), . . . , P(N) such that:
_
_
N
j=1
p
i,j
_
H(i)
T
P(j)H(i)
_
P(i)
_
_
< 0, i, j = 1..N with H(i) = e
h
i
_
I 0
0 0
_
5 Stability of Linear MB-NCS with Quantization
Here we extend our stability results to consider the case where quantization
errors occur. In particular a state space MB-NCS is considered. Two static
quantizers are studied: the Uniform Quantizer and the Logarithmic Quantizer.
These are called static since they partition the state space into invariant re-
gions that are xed in time. We also note that these quantizers represent
two of the most common data representations: xed point format for uniform
quantizers and oating point format for logarithmic quantizers. We will as-
sume that the transmission times are constant. We will also assume that the
compensated networked system without quantization is stable which in view
of Theorem 1 there exists positive denite P that satises:
_
e
(

A+

BK)
T
h
+(h)
T
_
P
_
e
(

A+

BK)h
+(h)
_
P = Q
D
(7)
where (h) = e
Ah
_
h
0
e
A
(

A+

BK)e
(

A+

BK)
d and with Q
D
symmetric and
positive symmetric. Note that e
(

A+

BK)h
+(h) was previously dened in (1).
5.1 State Feedback MB-NCS with Uniform Quantization
Dene the Uniform Quantizer as a function q : R
n
R
n
with the following
property z q(z) , z R
n
, > 0.
Theorem 10 [21] The plant state of the State Feedback MB-NCS satisfying
(7) and using the Uniform Quantizer will enter and remain in the region
x R dened by:
R =
_
e
(

A+

BK)h
+
max
(h)
_
r +
_
e
(A)h
+
max
(h)
_
where r =
_
max((e
Ah
(h))
T
P(e
Ah
(h))
T
)
2
min
(Q
D
)
and
max
(h) =
_
h
0
e
(A)(h)

_
A+

BK
_
e
(

A+

BK)
d
5.2 State Feedback MB-NCS with Logarithmic Quantization
Dene the Logarithmic Quantizer as a function q : R
n
R
n
with the follow-
ing property z q (z) z , z R
n
, > 0.
Theorem 11 [21] The State Feedback MB-NCS satisfying (7) and using the
Logarithmic Quantizer is exponentially stable if:
<
min
(Q
D
)
max
_
(e
Ah
(h))
T
P (e
Ah
(h))
_
Example 6 For our example we will use the following plant model parame-
ters: A =
_
0 1
1 3
_
, B =
_
0
1
_
. The plant will be a perturbed version of our plant
model:

A =
_
0.0689 0.9757
1.0396 3.0720
_
,

B =
_
0.0707
1.0187
_
. Both are unstable plants. The
controller is designed using the plant model: u =
_
2 5
x. This controller
places both eigenvalues of the compensated plant model at 1. Using update
time of h = 0.6 s, we will test two logarithmic quantizer functions q1 with a
mantissa word length of 12 bits and q2 with mantissa word length of 13 bits.
Their relative errors for the two dimensional space they will work on are for
q1 : 0.33 and for q2 : 0.20. Initializing the plant at
_
2
T
3
T

T
, we observe
from Figure 11 that the system working with quantizer q1( = 0.33) is unsta-
ble, while with q2( = 0.20) is stable (Figure 11). By using Theorem 11 and
a Q
D
= I we obtain a maximum relative error of 0.1241.
0 10 20 30 40
100
50
0
50
100
time (sec)
p
l
a
n
t

s
t
a
t
e

f
o
r

q
1
0 5 10 15 20
3
2
1
0
1
2
3
time (sec)
p
l
a
n
t

s
t
a
t
e

f
o
r

q
2
Fig. 11. Plant state time response for q1 and q2.
6 Stability of a Class of Non-Linear MB-NCS
We will determine sucient conditions for the stability of a state feedback
MB-NCS when the plant and controller are nonlinear. Let the plant, plant
model and controller be given by:
plant: x = f(x) +g(u); model:

x =

f(x) + g(u); controller: u =

h( x) (8)
Also dene e = x x as the error between the plant state and the plant model
state. From (8) we obtain:
x = f(x) +g
_
h( x)
_
= f(x) +m( x);
x =

f( x) + g
_
h( x)
_
=

f( x) + m( x)
(9)
We will also assume that the plant model dynamics dier from the actual
plant dynamics in an additive fashion:
f() = f() +
f
(); m() = m() +
m
() (10)
So we can rewrite (9) as:
x = f(x) +m( x)
x = f( x) +m( x) +
f
( x) +
m
( x) = f( x) +m( x) +( x)
(11)
We will now assume that f and satisfy the following local Lipschitz condi-
tions for x, y B
L
with B
L
a ball centered on the origin:
f(x) f(y) K
f
x y ; (x) (y) K
x y (12)
At this point it is to be noted that if the plant model is accurate the Lipschitz
constant K
will be small.
We will assume that the non-networked compensated plant model is ex-
ponentially stable when x(t
0
) B
S
, with x(t) B
h
for t [t
0
, t
0
+ h) with
B
S
and B
h
balls centered on the origin.
x(t) x(t
0
) e
(tt
0
)
, with , > 0. (13)
Theorem 12 [21] The non-linear MB-NCS with dynamics described by (8),
that satises the Lipschitz conditions described by (12) and with exponentially
stable compensated plant model satisfying (13) is asymptotically stable if:
_
1
_
e
h
+
K
_
e
K
f
h
e
h
_
_

K
f
+
___
> 0
Theorem 12 presents a sucient condition for stability of a class of nonlinear
systems. Note that if the model has the exact same dynamics as the plant,
that is if K
= 0, then the condition will be satised for arbitrarily large h.

Example 7 We use the inverted pendulum in Figure 12 as example.
Fig. 12. Inverted Pendulum.
The parameters for the plant are g = 10, L = 10, k = 0.1, and m = 1.01.
The model parameters are the same as the plant ones, except for the mass that
is m = 1.00. Finally the controller is given by: =
_
316 316
x. Using the
following Lipschitz and exponential stability constants: K
f
= 1.0507, K
=
0.0450, = 1.5, and = 0.6, Theorem 12 predicts stability for update times
between 0.55 and 2.55 sec. Simulations show that the system is unstable for h
greater that approximately 4.5 sec.
Conclusions
The MB-NCS control architecture presented in this chapter represents a
natural way of placing critical information about the plant on the network so
to reduce the data trac load. By making the sensor and actuator more in-
telligent the networked control system is able to predict the future behavior
of the plant, and send the precise information at critical times so to ensure
plant stability.
Model-based networked control systems are only one of the various ap-
proaches to networked control. We study these systems since the benets and
properties they have make them well suited for a variety of practical applica-
tions. Furthermore, the control structures presented appear to be amenable
to detailed analysis.
Acknowledgements
The partial support of the Army Research Oce (DAAG19-01-1-0743) and
of the the National Science Foundation (NSF CCR-02-08537 and ECS-02-
25265) is gratefully acknowledged.
References
1. P. Antsaklis and A. Michel, Linear Systems, McGraw-Hill, New York 1997.
2. T. Chen and B. Francis, Optimal Sampled-Data Control Systems, 2
nd
edition,
Springer, London 1996.
3. F. Kozin, A Survey of Stability of Stochastic Systems, Automatica Vol5 pp.
95-112, 1968.
4. M. Mariton, Jump Linear Systems in Automatic Control Marcel Dekker, New
York 1990.
5. O.L.V. Costa and M.D. Fragoso, Stability Results for Discrete-Time Linear Sys-
tems with Markovian jumping Parameters, Journal of Mathematical Analysis
and Applications, Vol 179, pp. 154-178, 1993
6. A. Dembo and O.Zeitouni, Large Deviation Techniques and Applications,
Springer-Verlag, New York, 1998.
7. Y. Fang, A New General Sucient Condition for Almost Sure Stability of Jump
Linear Systems, IEEE Transactions on Automatic Control, Vol 42, No3, pp.
378-382, March 1997.
8. S. Woods, Probability, Random Processes, and Estimation Theory for Engi-
neers, 2
nd
edition, Prentice Hall, Upper Saddle River 1994.
9. A.S. Matveev and A.V. Savkin, Optimal Control of Networked Systems via
Asynchronous Communication Channels with Irregular Delays, 40
th
IEEE Con-
ference on Decision and Control, Orlando Florida, December 2001, pp. 2323-2332.
10. L.A. Montestruque and P.J. Antsaklis, Model-Based Networked Control Sys-
tems: Stability, ISIS Technical Report ISIS-2002-001, University of Notre Dame,
www.nd.edu/isis/, January 2002.
11. L.A. Montestruque and P.J. Antsaklis, Model-Based Networked Control Sys-
tems: Necessary and Sucient Conditions for Stability, 10
th
Mediterranean Con-
ference On Control And Automation, July 2002.
12. L.A. Montestruque and P.J. Antsaklis, State And Output Feedback Control In
Model-Based Networked Control Systems, 41st IEEE Conference on Decision
and Control, December 2002.
13. L.A. Montestruque and P.J. Antsaklis, On the Model-Based Control of Net-
worked Systems, Automatica, Vol 39, pp 1837-1843.
14. L.A. Montestruque and P.J. Antsaklis, Stability of Networked Systems with
Time Varying Transmission Times, Transactions of Automatic Control, 2004,
to appear.
15. G.N. Nair and R.J. Evans, Communication Limited Stabilization of Linear
Systems, Proceedings of the Conference on Decision and Control, 2000, pp. 1005-
1010.
16. , Networked Control Sessions, Proceedings of the Conference on Decision and
Control, Proceedings of the American Control Conference, 2003.
17. H. Ishii and B. Francis, Stabilizing a Linear System by Switching Control and
Output Feedback with Dwell Time, Proceedings Of The 40
th
IEEE Conference
On Decision And Control, December 2001, pp. 191-196.
18. J.K. Yook, D.M. Tilbury, and N.R. Soparkar, Trading Computation for Band-
width: Reducing Communication in Distributed Control Systems using State Es-
timators, IEEE Transactions on Control Systems Technology, July 2002, Vol 10,
No.4, pp. 503-518.
19. G. Walsh, H. Ye, and L. Bushnell, Stability Analysis of Networked Control
Systems, Proceedings of American Control Conference, June 1999.
20. D. Hristu-Varsakelis, Feedback Control Systems as Users of a Shared Network:
Communication Sequences that Guarantee Stability, Proceedings of the 40
th
IEEE Conference on Decision and Control, December 2001, pp. 3631-3636.
21. L.A. Montestruque and P.J. Antsaklis, Technical Reports, University of Notre
Dame, www.nd.edu/isis/.

Networked Control Systems: A Model-Based Approach

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Networked Control Systems: A Model-Based Approach

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Networked Control Systems: A Model-Based Approach

Uploaded by

Copyright:

Available Formats

Networked Control Systems: A Model-Based

. Usually it is assumed that the actuator/controller will

Theorem 4 [21] The Extended H2 Norm, G

and formulate the following

= 0, then the condition will be satised for arbitrarily large h.

You might also like