Lecture 39 PDF

Lecture 39
Integrator Back-stepping;
Linear Quadratic (LQ) Observer
Dr. Radhakant Padhi
Asst. Professor
Dept. of Aerospace Engineering
Indian Institute of Science - Bangalore
Philosophy of Nonlinear Control Design

Using Lyapunov Theory
Dr. Radhakant Padhi
Asst. Professor
ADVANCED CONTROL SYSTEM DESIGN

Dr. Radhakant Padhi, AE Dept., IISc-Bangalore
Philosophy of Feedback Control

Design Using Lyapunov Theory
Motivation :
X = f ( X ,U )
Goal: Design
U = ( X ) such that
X = f ( X , ( X ) ) is asymptotically stable
Design Idea:
Choose a pdf V1 ( X )
Make V1 ( X ) V2 ( X ) , where V2 ( X ) > 0 (pdf)
Feedback Control Design Using

Lyapunov Theory: An Example
Problem:
Design a stabilizing controller for the following system

x = ax 2 x 3 + u
Solution:
Let
V1 ( X ) =
1
2
x2
V1 = x x = x ax 2 x 3 + u
= ax 3 x 4 + xu
Let us choose
V2 ( X ) = x 2
V1 ( X ) V2 ( X )
ax 3 x 4 + xu x 2
Philosophy of feedback control

xu x 2 + x 4 ax 3
i.e
Analysis:
u = x + x 3 ax 2
x = ax 2 x 3 x + x 3 ax 2
x = x
Advantage: The closed loop system is globally asymptotically stable.
Problem: The benefitial nonlinearity got cancelled.

(which is not desirable)

Let us choose:
V2 ( X ) = x 2 + x 4
Then
V1 V2 ( X )
leads to:
ax 3 x 4 + xu x 2 x 4
ax 3 + xu x 2
ax 2 + u = x
or
u = x ax 2


Closed Loop system: x = ax 2 x 3 x ax 2
i.e.
x = x3 x
The destabilizing nonlinearity got cancelled,

but the benefitical nonlinearity is retained !
Another Problem: If V2 ( X ) = x 2 ,
x = x, only if a is accurate
If the actual parameter value is a , then the feedback
loop operates with

x = x + ( a a ) x2
Can be potentially
destabilizing term
if ( a a ) is high
i.e. The global stability reduces to local stability.

This excits robustness issues!
However, if V2 ( X ) is made "Sufficiently powerful",
then the destabilizing effect can be minimized.
Hence, Lyapunov based designs can be "very robust"
Control Design Using

Integrator Back-stepping
Dr. Radhakant Padhi
Asst. Professor

Problem : Design a state feedback asymptotically stabilizing controller
for the following system

X = f ( X ) + g ( X )
= u
where
X Rn , R , u R
Note:
X
n +1
R
: State of the system, u : Control input (single input)

10
Assumptions:
f , g : D R n are smooth
f (0) = 0
Considering state as a "control input" of subsystem (1)
we assume that a state feedback control law
of the from = ( X ) , ( 0 ) = 0. Moreover,
a Lyapunov function V1 : D R + such that
V1
f ( X ) + g ( X ) ( X ) Va ( X ) , X D
X
where, Va ( X ) : D R + is a pdf function.
V1 ( X ) =

11
An important observation:
When X = 0, = ( 0 ) = 0 & X = f ( 0 ) = 0
(i.e. everything is nice)
However, when 0, X = f ( X ) and hence X 0
in general. That is the core problem!
We need some algebraic manipulation as follows.

12
Step - 1:
X = f ( X ) + g ( X ) + g ( X ) ( X ) g ( X ) ( X )
= f ( X ) + g ( X ) ( X ) + g ( X ) ( X )
= f ( X ) + g ( X ) ( X ) + g ( X ) z
By this construction, when z 0, X = f ( X ) + g ( X ) ( X )

which is asymptotically stable (i.e. X 0)!
13
z =
This is backstepping, since ( X ) is
stepped back by differentiation
= u
v
So, we have
X = f ( X ) + g ( X ) ( X ) + g ( X ) z
z=v
[ This system is equivalent to the original system ]

Note: =
X =
f ( X ) + g ( X )
X
X
T

14
Back-stepping:
Conceptual Block Diagram
Back-stepping
Ref: H. J. Marquez, Nonlinear Control

Systems: Analysis and Design,
Wiley, 2003.

15
Step-2:
Let V ( X , z ) = V1 ( X ) +
1 2
z
2
V1
V =
f ( X ) + g ( X ) ( X ) + g ( X ) z + z v
X
T
Then
Va ( X )
V1 T
Va ( X ) +
g ( X ) + v z
X
V1
g ( X ) k z, k > 0
X
Then V Va ( X ) k z 2 < 0
(ndf)
T
Let v =
ndf
nddf

16
Control Solution:
V1
v = u =
g(X )k z
X
T
V1
u =
g ( X ) k ( X )
X
T

where, =
f ( X ) + g ( X ) , k > 0
X
T
Note: In the design, there is a need to design ( X ) first

17
Integrator Back-stepping:
An Example
Problem:
x1 = ax12 x13 + x2
x2 = u
Solution :
X = x1 , f ( x1 ) = ax12 x13 , = x2 , g ( x1 ) = 1
To find ( x1 ) :
1 2
V1 ( x1 ) = x1
2
V1 = x1 x1 = x1 ( ax12 x13 + x2 ) Va ( x1 )
x12 + x14
18
An Example
ax13 x14 + x1 x2 x12 x14
x1 ( ax12 + x2 ) x12
Let
ax12 + x2 = x1
x2 = ( ax12 x1 ) ( x1 )
z
Modified system : x1 = ax12 x13 + ( x1 ) + x2 ( x1 )
z=v
x2 ( x1 )
u
1 2
z
2
V1
V1
2
3
+ v z
V = V1 + z v =
ax1 x1 + ( x1 ) +
x
x
1
1
Let V ( x1 , z ) = V1 ( x1 ) +

19
An Example
V1
v =
Let
k z , k > 0
x1
u = x1 k x2 ( x1 )
u=
ax12 x13 + x2 ) x1 k x2 ( ax12 x1 )
(
x1
= ( 2ax1 1) ( ax12 x13 + x2 ) x1 k x2 + ax12 + x1
u = (1 + 2ax1 ) ( ax12 x13 + x2 ) x1 k ( x1 + x2 + ax12 )

where k > 0
20
An Example
Note: The composite Lyapunov function is:
1 2
V = V1 + z
2
2
1 2 1
= x1 + x2 ( x1 )
2
2
1 2 1
2 2
= x1 + ( x2 + x1 + ax1 )
2
2
21
More General Case
System Dynamics:
X = f ( X ) + g ( X ) 1
1 = 2
2 = u
Idea : Successive iteration.
Note: The procedure for nth order system is entirely analogous

22
More General Case
Step-1: Consider the subsystem
X = f ( X ) + g ( X ) 1
1 = 2
Assumption:
1 = ( X ) is a stabiliting feedback law for

X = f ( X ) + g ( X ) 1
and V1 ( X ) is the coresponding Lyapunov function.
23
More General Case
By the result obtained before, we have
( X )
V1
2 =
f ( X ) + g ( X ) 1
g ( X ) k 1 ( X )
X
X
X
T
1 ( X , 1 )
(k > 0)
2
1
We also have V2 = V1 + 1 ( X )
2

24
More General Case
Step - 2:
X1
X f ( X ) + g ( X ) 1
=
+
0
1
f1 ( X1 )
2 = u
where, X 1
0
1 2

g1 ( X1 )
X

1
Using the same idea,

T
1
V2
u =
f1 ( X 1 ) + g1 ( X 1 ) 2
g1 ( X 1 ) k1 2 1 ( X 1 ) , k1 > 0
X 1
X 1
2
2
2
1
1
1
and V = V2 + 2 1 ( X 1 ) = V1 + 1 ( X ) + 2 1 ( X 1 )
2
2
2
25
Integrator Back-stepping for

Strict Feedback Systems
System Dynamics
X = f ( X ) + g ( X ) 1
1 = f1 ( X , 1 ) + g1 ( X , 1 ) 2
2 = f 2 ( X , 1, 2 ) + g 2 ( X , 1 , 2 ) 3
k = f k ( X , 1 , , k ) + g k ( X , 1 , , k ) u
Strong Assumption:
g1 ( X , 1 ) , g 2 ( X , 1 , 2 ) , , g k ( X , 1 , , k ) 0
over the domain of interest t
26
Integrator Back-stepping for

Strict Feedback Systems
Special Case:
X = f ( X ) + g ( X )
= fa ( X , ) + ga ( X , ) u
Solution:
Define = v and carryout the design for v as before.
Finally f a ( X , ) + g a ( X , ) u = v
i.e.
u=
1
v f a ( X , )
ga ( X , )
Note: By assumption, g a ( X , ) 0 t
27
Linear Quadratic (LQ) Observer
Dr. Radhakant Padhi

Asst. Professor
Why Observers?
z
State feedback control designs need the state

information for control computation
In practice all the state variables are not available

for feedback. Possible reasons are:
Non-availability of sensors
Expensive sensors
Quality of some sensors may not acceptable due to noise (its
an issue in output feedback control design as well)
A state observer estimates the state variables

based on the measurement of some of the output
variables as well as the plant information.
29
Observer
z
An observer is a dynamic system whose

output is an estimate of the state vector X
Full-order Observer
Reduced-order Observer
Observability condition must be satisfied for

designing an observer (this is true for filter
design as well)
30
Observer Design for Linear Systems

Plant : X = AX + BU
Y = CX (sensor output vector)
Let the observed state be X and the
Plant
Observer dynamics be
+ BU
+KY
X = AX
e
Error : X
( X X )
State observer
31

Error Dynamics:
X = X X
+ BU
+KY
= ( AX + BU ) AX
e
Add and Substract AX and substitute Y = CX

+ AX
AX
+ BU BU
K CX
X = AX AX
e
= ( A A ) X + A ( X X ) + ( B B ) U K eC X
+ A A K C X + B B U
= AX
e
Goals: 1. Make the error dynamics independent of X
( X may be large, even though X may be small)
2. Eliminate the effect of U from eror dynamics
32

This can be done by enforcing
Necessary and sufficient condition

for the existence of Ke :
The system should be observable.
A A K eC = 0
and B B = 0
This results in
A = A K C
e
B = B
Observer dynamics: X = AX + BU + K e Y CX
33
Observer Design: Full Order

z
Order of the observer is same as that of the system

(i.e. all states are estimated, irrespective of whether
they are measured or not).
Goal: Obtain gain Ke such that the error dynamics

are asymptotically stable with sufficient speed of
response.
This means that = A Ke C is Hurwitz (i.e. it has all
eigenvalues strictly in the left half plane.
Note: T =AT CTKeT and the eigen values of both

and T are same!
34
Comparison of Control and

Observer Design Philosophies
Control Design
z
CL Dynamics
Observer Design
z
X = ( A BK ) X
z
Objective
CL Error Dynamics
= ( A K C) X
X = AX
e
z
X ( t ) 0, as t
Objective
X ( t ) 0, as t
Notice that
T
( A K e C ) = ( A K e C )
= AT C T K eT
35
Algebraic Riccati Equation (ARE)

Based Observer Design
Dual System
System
Z = AT Z + C T V
n = BT Z
X = AX + BU
Y = CX
M = B AB
A B
N = C T AT C T
n 1
T n1
CT
M = C T AT C T
N = B A B
CT
A n 1 B
T n1
LQR Design
U = K X
36
ARE Based Observer Design

CL system (control design)
Error Dynamics
X = ( A K eC ) X
X = ( A BK ) X
X 0 as t
( A K eC )
= AT C T K eT
Analogous
1
K = R B P,
T
P>0
where,
= R CP
where,
1
PA + A P PBR B P + Q = 0
T
T
e
Acts like a
controller
gain
PAT + AP PCT R1CP + Q = 0

Observer Dynamics
X = AX + BU + K e (Y CX )
37
Continuous-time Kalman Filter Design for

Linear Time Invariant (LTI) Systems
Dr. Radhakant Padhi
Asst. Professor
Problem Statement
System Dynamics: X = AX + BU + GW W (t ) : Process noise vector
V (t ) : Sensor noise vector
Measured Output: Y = CX + V
Assumptions:
(i )
X (0) X 0 , P0 , W (t ) ( 0, Q ) and V (t ) ( 0, R )
are "mutually orthogonal"
[ X (0) :
initial condition for X ]
(ii ) W (t ) and V (t ) are uncorrelated white noise

(iii ) E W (t ) W T (t + ) = Q ( ), Q 0 (psdf)
E V (t ) V T (t + ) = R ( ), R > 0 (pdf)
39
Problem Statement
Objective:
To obtain an estimate of the state vector X ( t ) using the
state dynamics as well as a "sequence of measurements"
as accurate as possible.
i.e., to make sure that the error X (t )
X (t ) X (t ) becomes
very small ideally X (t ) 0 as t .

40
Observer/Estimator/Filter Dynamics
X = AX + BU + K e Y Y
where (i ) X = E ( X ) : Estimate of the state X

(ii ) Y = E (Y ) : Estimate of the output Y
= E ( CX + V )
= E ( CX ) + E (V )
= CE ( X )
( E (V ) = 0 )
= CX
(iii ) K e : Estimator/Filter/Kalman Gain
Problem : How to design K e ?
41
Solution: Summary
(i) Initialize X (0)
(ii) Solve for Riccati matrix P from the Filter ARE:
AP + PAT PC T R 1CP + G Q GT = 0
(iii) Compute Kalman Gain:
K e = PC T R 1
(iv) Propagate the Filter dynamics:
X = AX + BU + K e Y CX
where Y is the measurement vector

42

43

Lecture 39 PDF

Uploaded by

Document Information

Original Description:

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

Lecture 39 PDF

Uploaded by

Copyright:

Available Formats

Lecture 39

Philosophy of Nonlinear Control Design

ADVANCED CONTROL SYSTEM DESIGN

Philosophy of Feedback Control

Feedback Control Design Using

Design a stabilizing controller for the following system

Philosophy of feedback control

Advantage: The closed loop system is globally asymptotically stable.

Problem: The benefitial nonlinearity got cancelled.

Philosophy of feedback control

ADVANCED CONTROL SYSTEM DESIGN

Philosophy of feedback control

The destabilizing nonlinearity got cancelled,

Philosophy of feedback control

i.e. The global stability reduces to local stability.

Control Design Using

ADVANCED CONTROL SYSTEM DESIGN

for the following system

ADVANCED CONTROL SYSTEM DESIGN

ADVANCED CONTROL SYSTEM DESIGN

By this construction, when z 0, X = f ( X ) + g ( X ) ( X )

stepped back by differentiation

[ This system is equivalent to the original system ]

ADVANCED CONTROL SYSTEM DESIGN

Ref: H. J. Marquez, Nonlinear Control

ADVANCED CONTROL SYSTEM DESIGN

ADVANCED CONTROL SYSTEM DESIGN

Note: In the design, there is a need to design ( X ) first

Modified system : x1 = ax12 x13 + ( x1 ) + x2 ( x1 )

ADVANCED CONTROL SYSTEM DESIGN

= ( 2ax1 1) ( ax12 x13 + x2 ) x1 k x2 + ax12 + x1

u = (1 + 2ax1 ) ( ax12 x13 + x2 ) x1 k ( x1 + x2 + ax12 )

Note: The procedure for nth order system is entirely analogous

1 = ( X ) is a stabiliting feedback law for

ADVANCED CONTROL SYSTEM DESIGN

Using the same idea,

Integrator Back-stepping for

Integrator Back-stepping for

Linear Quadratic (LQ) Observer

Dr. Radhakant Padhi

State feedback control designs need the state

In practice all the state variables are not available

A state observer estimates the state variables

An observer is a dynamic system whose

Observability condition must be satisfied for

Observer Design for Linear Systems

Observer Design for Linear Systems

Add and Substract AX and substitute Y = CX

Observer Design for Linear Systems

Necessary and sufficient condition

Observer Design: Full Order

Order of the observer is same as that of the system

Goal: Obtain gain Ke such that the error dynamics

Note: T =AT CTKeT and the eigen values of both

Comparison of Control and

Algebraic Riccati Equation (ARE)

ARE Based Observer Design

PAT + AP PCT R1CP + Q = 0

Continuous-time Kalman Filter Design for

initial condition for X ]

(ii ) W (t ) and V (t ) are uncorrelated white noise

very small ideally X (t ) 0 as t .