Sei sulla pagina 1di 47

Controlling

Across
Networks
Controlling Across Networks

Chaouki T. Abdallah
Professor & Chair
ECE Department, The University of New Mexico

Talk Given at Shanghai Jiao Tong University


上海交通大学
阿布杜拉 乔基 

Controlling Across Networks


The Really Big Picture

Controlling
Across
Networks

Controlling Across Networks


The Big Picture

Controlling
Across
Networks
We live and operate in a networked world.
Networks provide a powerful metaphor for describing
system behavior in biology, computer science, physics,
social science, and engineering.
Complex networks are being studied for the purpose of
gaining insight into how properties such as community
structure and small-world effects emerge.
In modern cars and airplanes, as well as in networked
homes and office buildings, modern control systems
are increasingly incorporating communication networks
in feedback loops.

Controlling Across Networks


Networks and Control

Controlling
Across
Networks

Control engineers are thus forced to expand their


application domain by incorporating the communication
infrastructure into their designs, and by considering the
impact of link capacity, latency, and packet loss on the
performance of feedback control.
In this talk, ...

Controlling Across Networks


The Need to Study Networks

Controlling
Across
Networks

“The increasing complexity and interconnectedness of


energy, telecommunications, transportation, and financial
infrastructures pose new challenges for secure, reliable
management and operation.1”
“Complex interactive networks are omnipresent and critical
to economic and social well-being.”

M. Amin, National Infrastructures as Complex Interactive


Networks, Chapter 14 in Automation, Control, and
Complexity: An Integrated Approach, T. Samad and J.
Weyrauch, Eds. (John Wiley and Sons, NY, 2000).

Controlling Across Networks


Obvious Questions

Controlling
Across
Networks
How many of us would choose an old and more
expensive computer over a modern and cheaper one,
IF the first can be networked while the second can not?
It is the obviously the network!
What connects us makes us stronger (the whole is
more than the sum of it parts), but also more vulnerable
(viruses, marketing, etc.)
BUT, suppose you understand how networks come to
be, their structure, and some of their hidden properties:
Can you use that knowledge to design better processes
over the network?

Controlling Across Networks


Networks as Communications Media

Controlling
Across
Networks
Connectedness: which expresses the existence of a
path between the information transmitter and the
information receiver.
Navigability: quantified by the difficulty of finding a
connecting path. Typically, this difficulty depends on
whether the path is predetermined, or whether it is
discovered in an ad hoc fashion.
Efficiency: as represented by the latency (delays) of
each utilized path. This latency, usually a function of
the number of hops and the individual link latencies,
must be sufficient to guarantee desired end-to-end
communication latencies.

Controlling Across Networks


Connectedness

Controlling
Across
Networks

Connectedness is identical to the mathematical notion of


percolation. Percolation is illustrated by a wildfire, initiated
at a source vertex, spreading across an edge connected to
a burning vertex with a fixed probability By analyzing the
number of vertices reached by the process, it is possible to
determine whether there exists a path connecting a given
pair of nodes.

Controlling Across Networks


Power Laws in Networks

Controlling
Across
Networks
One common feature of many real world networks is a
power-law degree distribution, in which the probability of a
randomly chosen vertex having k neighbors scales as

p(k) ∝ k−α (1)

The ubiquity of the power-law degree distribution has led


network theorists to focus on graph models that exhibit this
feature, but whose topological structure is otherwise
random. Obviously, a network with many redundant paths
between all pairs of vertices becomes more robust to node
and edge failures of all kinds.

Controlling Across Networks


Random Newtorks

Controlling
Across
Networks

Figure: Random Networks

Controlling Across Networks


Power Laws

Controlling
Across
Networks

Figure: Random Networks

Controlling Across Networks


Navigability

Controlling
Across
Networks

Given that a network is connected, several paths may


typically connect a transmitter with a receiver. In network
theory, a network’s navigability is determined both by how
easily such a path can be found, and how many hops
(delays!) such a path ultimately requires. Solutions can be
grouped into two categories: Central authorities, in which
the communication path between two vertices is determined
by an external source, later mirrored by the network’s
routers, and Decentralized techniques, in which routing
decisions are made independently by network routers,
possibly in an ad-hoc fashion.

Controlling Across Networks


Efficiency

Controlling
Across
Networks

As far as control design is concerned, a communication


channel is merely a medium for obtaining or sending
information (measurement signals, or control commands).
From this perspective, what seems to be important is: how
much information can be carried, and how fast can it be
transferred.

Controlling Across Networks


Our Focus

Controlling
Across
Networks

In the remainder of the talk, I will focus on the issues of


source encoding in order to send output and control signals
across a packet network.

I will not discuss the issues of connectedness or navigability

Controlling Across Networks


Closed-Loop Type I

Controlling
Across
Networks
LTI

ENCODER

Rate: Rp packets/time_unit NETWORK

DECODER

CONTROL

Figure: Closed-loop network control system: Type 1

Controlling Across Networks


Closed-Loop Type II

Controlling
Across
Networks
LTI

ENCODER ENCODER

Rate: Rp1 packets/time_unit

NETWORK NETWORK
2 1

Rate: Rp2 packets/time_unit

DECODER DECODER

CONTROL

Figure: Closed-loop network control system: Type 2

Controlling Across Networks


Fundamental Results

Controlling
Across
Networks Let the system be

x(k + 1) = Ax(k) + Bu(k); y(k) = Cx(k)

Theorem (Data Rate Theorem)


The minimum required rate for stabilization is given by
n
X
R> log2 (|λi (A)|); (2)
i=1

where λi are the eigenvalues of an open-loop discrete linear


system.

Controlling Across Networks


Fundamental Results

Controlling
Across
Networks

Theorem (Tatikonda)
For system with (A, B) a stabilizable pair, a necessary
condition on the channel capacity for almost surely
asymptotic stabilizability is that
X
C> max { 0, log2 |λ(A)| }
λ(A)

Controlling Across Networks


Specific Problem

Controlling
Across
Networks

Consider the discrete LTI system given by

x(k + 1) = Ax(k) + Bu(k) (3)

where A is n × n and we assume that it is diagonal


A = diag(λ1 , . . . , λn ) and |λj | ≥ 1, ∀j ∈ {1, . . . , n}, and
λi 6= λj if j 6= i, x(k) is n × 1, B is n × m and u(k) is m × 1.

Controlling Across Networks


Assumptions

Controlling
Across
Networks

Assumptions
The packet based network considers a packet size of
Dmax bits used for data.
Noiseless network.
The controller does not saturate.
There are not packet losses in the network.
Synchronization between encoder and decoder: the
decoder knows exactly both the sign and the position of
each significant bit when it is encoded.

Controlling Across Networks


Notation

Controlling
Across
Networks

log() represents log2 ().


The norm symbol (k.k) denotes the Euclidean norm.
d.e is the ceil function.
The variable µ to denote the controllability index.

Controlling Across Networks


Background: Previous Results

Controlling
Across
Networks

Theorem
(L. Shi, R. Murray) Assuming B, C are invertible and the
system dimension is n. Then a sufficient condition for the
closed loop exponential stability is that the network
parameters and the system parameters satisfy the
inequality below
−R1 R2 

|A| 2 n + |B| B−1 A 2− n < 1

Controlling Across Networks


Disadvantage of the Result

Controlling
Across
Networks

Disadvantages
The assumption of an invertible B is very conservative.
Moreover, the idea of state augmentation for the
time-delay consideration is not longer valid since the
augmented B is, in general, not invertible.
We focus our work in removing this constraint.

Controlling Across Networks


Results: NCS Type I

Controlling
Across
Networks For a Closed-Loop Type I, we have the following result.
Theorem
Assuming an equal allocation of bits per state component, a
network rate, Rp of packets/bits, and (A, B) is a controllable
pair with controllability index µ, a sufficient condition for
system (3) to be asymptotically stabilizable is
 
R
Rp > ,
DMax
R
where R = n dlog (kAµ k) + 1e and every state can allocate n
bits/sample.

Controlling Across Networks


Results: NCS Type I

Controlling
Across
Networks

Corollary
Assuming an equal allocation of bits per state component
and (A, B) is a controllable pair, where B is n × 1 and the
control law, u(k), is 1 × 1, a sufficient condition for system
(3) to be asymptotically stabilizable is
 
R
Rp ≥ ,
DMax
R
where R = n dlog (kAn k) + 1e and every state allocates n
bits/sample.

Controlling Across Networks


Results: NCS Type I: Time Delay

Controlling
Across
Networks

Let us consider the network control System type 1 and the


discrete LTI system given by the following equation:

x(k + 1) = Ax(k) + Bu(k − p) (4)

where A = diag(λ1 , . . . , λn ) and |λj | ≥ 1, ∀j ∈ {1, . . . , n}, and


λi 6= λj if j 6= i, x(k) is n × 1, B is n × 1 and u(k) is 1 × 1 and
p ∈ N is the time delay. We assume here that the delay is a
constant equal to p time-steps even though that the network
probably imparts a time-varying and random delay.

Controlling Across Networks


Results: NCS Type I: Time Delay

Controlling
Across Theorem
Networks
Assuming
l an e.a.b per state component, a network rate of
m
R
Rp = DMax packets/time-step, and (A, B) is a controllable
pair. A sufficient condition for system (4) to be
asymptotically stabilizable is

R > (n + p) log(kAn+p k) + 1
 

   
A B 0 ... 0 0
 0 0 1 . . . 0 0
   
where A =  0 0 0 . . . 0 and B = 0 and every state
   
 1  .. 
  .
..
0 0 . ... 0 1
R
can allocate n+p bits/sample..
Controlling Across Networks
Results: NCS Type II

Controlling
Across
Networks

Theorem
Assuming an equal allocation of bits per mode and (A, B) is
a controllable pair, where B is n × 1 and the control law, u(k),
is 1 × 1, a sufficient condition for system (3) to be
asymptotically stabilizable is
R1
kAn k 2− n +1 + kζk ζ −1 A 2−R2 +1 < 1

where ζ = B| AB| . . . | An−1 B .


 

Controlling Across Networks


Simulations: Example 1

Controlling
Across
Networks

First, we tested the results of Theorem 1 for the system:


   
1 0 0 1 0
x(k + 1) = 0 3 0 x(k) + 1 1 u(k) (5)
0 0 4 0 1
 
−1.33
With initial condition x(0) =  3.768 .
8.44
The rate obtained is R/n = 6 bit/time-step and the
simulation for such a rate is shown in the next figure.

Controlling Across Networks


Simulations: Example 1

Controlling
Across System Evolution (Using R/n = 6 bit/time−step)
60
Networks
x1(k)
x2(k)
50 x3(k)

40
States

30

20

10

0 2 4 6 8 10 12
Time Step

R
Figure: (Type 1): Multi-Input Case using n = 6 bits/time-step.

Controlling Across Networks


Simulations: Example 2

Controlling
Across
We considered a single-input system given by:
Networks    
20 0 10 1
z(k + 1) =  0 10 0  z(k) + 1 u(k) (6)
0 10 30 1
Using a state-space transformation, we diagonalized the
system to obtain:
   
20 0 0 −1.000
x(k + 1) =  0 10 0  x(k) +  2.121  u(k) (7)
0 0 30 1.225
 
1.33
We assume the initial condition to be x(0) = 3.768.
8.44
We get R/n = 16 bit/time-step.
Controlling Across Networks
Simulations: Example 2

Controlling
Across System Evolution (R/n = 16 bit/time−step)
14000
Networks
x1(k)
12000 x2(k)
x3(k)
10000

8000

6000
States

4000

2000

−2000

−4000

−6000
0 5 10 15
Time Step

R
Figure: (Type 1): Single Input Case using n = 16. bit/time-step.

Controlling Across Networks


Is the result conservative?

Controlling
4 System Evolution (Using R/n = 14 bits/time−step)
Across x 10
12
Networks
x1(k)
10
x2(k)
x3(k)
8

4
States

−2

−4

−6
0 5 10 15 20 25
Time Step

R
Figure: (Type 1): Single Input Case using n = 14. bit/time-step.

Controlling Across Networks


Simulations: Example 3 (Time Delay)

Controlling
Across
Networks

Let us finally consider a system with time-delay p = 2


evolving according to the following dynamics:
   
2 0 1
x(k + 1) = x(k) + u(k − 2) (8)
0 1.5 1
 
1.33
with the initial condition state vector x(0) = .
30.768
For this system, our result gives a rate bounded below by
R/n = 6 bit/time-step.

Controlling Across Networks


Simulations: Example 3

Controlling
Across System Evolution (Using R/(n+p) = 6 bit/time−step)
400
Networks
x1(k)
350
x2(k)

300

250

200
States

150

100

50

−50
0 2 4 6 8 10 12 14 16 18
Time−Step

Figure: Closed-loop NCS with Time-Delay: Type 1

Controlling Across Networks


Playing Communications against Control

Controlling
Across
Networks

By modifying the communication protocol, Tanner shows


that velocity synchronization in a connected group of
autonomous mobile agents, may still be achieved when the
agent controllers use delayed information, regardless of the
size of this delay, if control and communication are properly
interleaved.
Tanner used the composition properties of graphs are used
to show that under certain assumptions on the
communication topology, delays may have no effect on
stability.

Controlling Across Networks


Designing with Delays

Controlling
Across
Consider the following second-order linear, time-invariant
Networks plant:
Y(s) 1
H(s) = = 2 (9)
U(s) s + w2n
where wn is the natural frequency of the system. Let a
static, output-feedback delay compensator be given by:

U(s) = C(s)U(s); C(s) = ke−sτ (10)

where k and τ are both design parameters. The closed-loop


transfer function is given by:

Y(s) ke−sτ
= 2 (11)
R(s) s + w2n − ke−sτ

Controlling Across Networks


Designing with Delays

Controlling
Across
Networks

Let us consider the Nyquist plot of the open-loop system


H(s)C(s) given by:

Ke−jwτ
H(jw)C(jw) = (12)
w2n − w2

We divide the Nyquist graph of (12) into three regions: the


first, when the frequency is less than the natural frequency,
or w < wn ; the second, when they are equal, i.e. w = wn ;
and the third when the frequency is greater than the natural
frequency, so that w > wn .

Controlling Across Networks


Designing with Delays

Controlling
Across
Networks Note that the magnitude kH(jw)C(jw)kw=wn is infinite, so our
analysis focuses on the cases were w 6= wn . For those
regions, the magnitude of the open-loop gain is:
K
|G(jw)C(jw)| = ; 0 ≤ w < wn
w2n
− w2
(13)
K
|G(jw)C(jw)| = 2 ; w > wn
w − w2n

and its phase is given by

θ(w) = −π − wτ ; for 0 ≤ w < wn


(14)
θ(w) = −wτ ; for w > wn

Controlling Across Networks


Designing with Delays

Controlling
Across
Networks

The intersections of the polar plot with the negative real axis
take place at the frequencies wc where
2nπ
wc = , 0 ≤ wc < wn
τ (15)
(2n + 1)π
wc = , wc > wn
τ
In order to guarantee asymptotic stability of closed-loop
system, the magnitude |G(jw)C(jw)| evaluated at wc must be
less than 1 so that the -1 point is not encircled.

Controlling Across Networks


Designing with Delays

Controlling
Across
Networks
Therefore:
k
< 1, for 0 ≤ 2nπ/τ < wn
w2n
− (2nπ)2 /τ 2
(16)
k
< 1, for (2n + 1)π/τ > wn
((2n + 1)π)2 /τ 2 − w2n

Combining the last two conditions we find the lower and


upper bounds of the stability region for positive gain k > 0

2nπ (2n + 1)π


p <τ < p
2
wn − k w2n + k
(17)
1 + 4n
0<k≤ w2
1 + 4n + 8n2 n

Controlling Across Networks


Designing with Delays

Controlling
Across
Networks

Following a similar analysis, we find the complete stability


region for negative gains k < 0 as follows

2nπ (2n − 1)π


p >τ > p
2
wn − k w2n + k
(18)
1 + 4n
0>k≥ w2
1 + 4n + 8n2 n
Combining the stability regions for positive and negative
gains (and positive-negtaive delays) , we obtain the graph
shown below for w2n = 1.

Controlling Across Networks


Controlling
Across
Networks

Figure: Stability Regions with Positive and Negative Delay

Controlling Across Networks


Designing with Delays

Controlling
Across
Networks

As expected, the stabilizing gain regions decrease as the


delay increases along the vertical axis. Using this simple
graphical tool, we may choose the delay and gain values to
guarantee that the closed-loop system is stable. In order to
illustrate the process, assume that the open-loop plant is
described by the transfer function:
1
H(s) = (19)
s2 +1

Controlling Across Networks


Designing with Delays

Controlling
Across
Networks

Due to the connecting network, the output signal is subject


to a delay that is randomly distributed between 0 and 3
seconds according to a uniform distribution. Let the initial
conditions be y(0) = y(0) = 0.1. We the show via a Matlab
simulation, how the closed-loop system is stable with a
choice of a small gain k = 0.1 despite the fact that the delay
is randomly varying. Note that this simply illustrates that the
delayed feedback controller is somewhat robust to changes
in the delay as may be encountered across a shared
communication network.

Controlling Across Networks


Designing with Delays

Controlling 0.15
Across
Networks
0.1

0.05

−0.05

−0.1

−0.15

−0.2
0 50 100 150 200 250 300 350 400

Figure: Simulation of Second Order System with Uniform Variable


Delay, Positive Feedback
Controlling Across Networks
CDC 2008-Cancun

Controlling
Across
Networks

Figure: See you there, December 9-11, 2008.

Controlling Across Networks

Potrebbero piacerti anche