Lyapunov Based Redesign

• Motivation

Consider x  f (t , x)  G (t , x)u , x  R n , u  R p
f : R  D  R n
G : R  D  R n p

But the real system is

x  f (t , x )  G (t , x )u  G (t , x ) (t , x, u )
 f (t , x)  G (t , x)(u   (t , x, u ))

 : R  D  R p  R p
 is unknown but not necessarily small. We assume
it has a known bound.
 (t , x, u )  (t , x, u )

11-1
Problem

Find a state feedback controller so that the closed loop system is stable in a
sufficiently strong sense.

Approach : u   (t , x)  v(t , x )
(i) (ii)

(i) chosen so that nominal closed loop system is

asymptotically stable.

(ii) chosen so as to cancel the effect of uncertainty.

11-2
Solution

Assume that u   (t, x) results in the uniformly asymptotically stable

nominal closed loop system,
x  f (t , x )  G (t , x ) (t , x)

Suppose also that V (t , x) is a Lyapunov function that proves

the following.

1 (|| x ||)  V (t , x)   2 (|| x ||) : l.p.d. and dec.

V V
  f (t , x)  G (t , x) (t , x)   3 (|| x ||) : l.n.d.
t x
where  i  K , i.e.  i : R   R ,  i (0)  0,  i () is strictly increasing.

11-3
Solution (Continued)

Suppose finally that ||  (t , x, (t , x)  v) ||  (t , x)  k || v ||,

where 0  k  1,  : R  D  R

The actual system with u   (t , x)  v becomes

x  f (t , x )  G (t , x ) (t , x)  G (t , x )[v   (t , x, (t , x )  v )]
V V V
Then V   ( f  G )  G (v   )
t x x
V
  3 (|| x ||)  G (v   )
x
V
Let wT  G
x
Then V   3 (|| x ||)  wT v  wT 

11-4
Solution (Continued)
Due to the matching condition, v can wipe out  .
There are two ways at least : when  is bounded in ||  ||2 or in ||  || .
i) ||  ||2
||  (t , x, (t , x)  v) ||2   (t , x)  k || v ||2 , 0  k  1
 T v   T    T v  ||  ||2 ||  ||2   T v  ||  ||2   (t , x)  k || v ||2 
Let  (t , x )   (t , x), (t , x)  R  D
and choose
 (t , x) 
v
1  k ||  ||2
 (t , x)   (t , x)
Then  T v    T  ||  ||2
1  k ||  ||2 1 k
 (t , x)  (t , x)
T v  T   ||  ||2   ||  ||2  ||  ||2 k
1 k 1 k
1 k
 (  ) ||  ||2   ||  ||2
1 k 1 k
   ||  ||2   ||  ||2  0
11-5
Solution (Continued)
ii) ||  ||
||  (t , x, (t , x)  v) ||   (t , x)  k || v || , 0  k  1

Now    i i  max |  i |  | i |  ||  ||1||  ||
T
i
i 1

 T v   T    T v  ||  ||1||  ||
  T v  ||  ||1   (t , x)  k || v || 
Let  (t , x)   (t , x), (t , x)  R  D
 (t , x)
and choose v  sgn( )
1 k
  (t , x)   k 
Then  T v   T    T   sgn( )   ||  ||1   (t , x) 
 1 k   1  k 
 k
  ||  ||1   ||  ||1  ||  ||1
1 k 1 k
 1 k 
     ||  ||1   ||  ||1
 1  k 1  k 
   ||  ||1   ||  ||1  0
11-6
Smooth Control

• Smooth Control (  2 case)

  (t , x) 
 1  k ||  || , if  (t , x) ||  ||2  
 2
v
2
   (t , x)  , if  (t , x) ||  ||  
 1 k 
2

 
V
where  T  G (t , x)
x

Obviously when  (t , x) ||  ||2    V  0

Analyze what happens when  (t , x ) ||  ||2  

11-7
We have
  2
 
V   3 (|| x ||2 )    
T
 
 1 k  
2
  3 (|| x ||2 )  ||  ||22   ||  ||2  k ||  ||2 || v ||2
(1  k )
2 2 k 2
  3 (|| x ||2 )  ||  ||2   ||  ||2  ||  ||22
(1  k ) (1  k )
2 2 2
  3 (|| x ||2 )  ||  ||2   ||  ||2    3 (|| x ||2 )  ||  ||22  ||  ||2
 
1 2 2z 
Here f ( z )   z  z is maximized at 1  z 
  2
1 1 2  
and max( z 2  z)    
  4 2 4

Thus V   3 (|| x ||2 )  , when  (t , x) ||  ||2  
4

On the other hand, when  (t , x) ||  ||2   , V is


V   3 (|| x ||2 )   3 (|| x ||2 ) 
4

Thus V   3 (|| x ||2 )  irrespective of the value of  (t , x) ||  ||2
4

11-8
Smooth Control (Continued)

Then take r large, so that 1  

 3 (r )  i.e. r ( )   31 ( )
2 4 2

1 1 
Then V ( x)    3 (|| x ||2 )   3 (|| x ||2 ) 
2 2 4
1
   3 (|| x ||2 ),  || x ||2  r ( )
2

Q  
Q   x :  3 (|| x ||)  
 2

r ( )  0 as   0

11-9
 - When  is chosen small, we can arrive at a sharper result.
n
Assume that D  R
2
such that  3 (|| x ||2 )   ( x)
 (t , x)  0  0,  (t , x)  1 ( x) where  : R n  R .
Then, when  (t , x) ||  ||2   ,
2

V   3 (|| x ||2 )  ||  ||22   ||  ||2

2 02
  ( x)  ||  ||22  1 ( x) ||  ||2

T
1 2 1   ( x)   1  1    ( x) 
   ( x)    
2 2 ||  ||2    1 202 /   ||  ||2 

P
where P is positive definite if   202 12 .
202
Thus choosing   we
, have V   1  2 ( x ).
12 2

when  (t , x) ||  ||2   .
2 1 2
Also V   ( x)   2  ( x )
1 2
We conclude V   2  ( x) which shows that the origin is uniformly
asymptotically stable.
11-10
Example
Ex: x1  x2
x 2   a[sin( x1  1 )  sin 1 ]  bx2  cu
Let aˆ and ĉ denote the nominal values of a and c.
u
Let u  cu
ˆ u 
cˆ
Choose u   ( x)  aˆ  sin( x1  1 )  sin 1   (k1 x1  k2 x2 )
where k1 & k2 are chosen so that
0 1 
 k k  b  is Hurwitz.
 1 2 

Then
x1  x2
x 2  k1 x1  (k2  b) x2   ( x, u )
ˆ  acˆ 
 ac  c  cˆ 
 ( x, u )  
ˆ   sin( x1   1 )  sin( 1 )    ˆ   k1x1  k2 x2 
 c   c 
11-11
Example (Continued)
ˆ  acˆ
ac c  cˆ c  cˆ
Hence  ( x, ( x)  v)  x1  k1 x1  k2 x2  v
cˆ cˆ cˆ
 1 || x ||2  k || v ||2
âac  acˆ c  ĉc c  ĉc
where 1   || k ||2 , k  .
cˆ cĉ ĉc
c  cˆ
We need  k 1
cˆ

Calculate the control v. Let V  xT Px be the Liapunov function

for the nominal closed loop system where P is defined by

0 k1  0 1 
1 k  b  P  P  k k  b    I  solve P
 2   1 2 
V  P11 P12  0 
T
Then w  G  2 x PB  2  x1 x2  
T
 
x  P21 P22  1 
 2( P12 x1  P22 x2 )

11-12
Example (Continued)

Choose

  (t , x) 
 1  k ||  || , if  ||  ||2  
 2
v
2
   (t , x)  , if  ||  ||  
 1 k 
2

202
where   2
1

Then the control u is

u 1
u   ( ( x)  v )
cˆ cˆ
 stabilize the origin globally!!

11-13
Backstepping

Consider a system
  f ( )  g ( ) 1
  u 2
where   R n , u  R and f () with f (0)  0 and g () are smooth in D  R n .

Design a state feedback controller to stabilize the system at (  0,   0)

u  
 g ( ) + 

f ( )
f ()

11-14
Backstepping (Continued)

Suppose (1) is stabilized by    ( ) with  (0)  0, i.e., the origin of

  f ( )  g ( ) ( ) is asymptotically stable. Furthermore, suppose that
we know a positive Lyapunov function V ( ) where
V ( )
 [ f ( )  g ( ) ( )]   W ( ),   D A
where W ( ) is positive definite.
Then, from (1),(2)
  [ f ( )  g ( ) ( )]  g ( )[   ( )]
  u

u 
 + g ( ) +  

f ()  g () ()

  ( )

11-15
Backstepping (Continued)

Let denote z     ( ), then

  [ f ( )  g ( ) ( )]  g ( ) z
z  u  

u z g ( ) 
+  + 

f ()  g () ()

 

backstepping   ( ) through the integrator

where    [ f ( )  g ( ) ]

11-16
Backstepping (Continued)
Let v  u   , then
  [ f ( )  g ( ) ( )]  g ( ) z 3
z  v 4
which is similar to the original system but  has an asymptotically
stable origin when the input is 0.
Let Va ( ,  )  V ( )  12 z 2 , then
Va  V [ f ( )  g ( ) ( )]  V g ( ) z  zv   W ( )  V g ( ) z  zv
  

Thus choosing v   V g ( )  Kz, K  0, then

Va   W ( )  Kz 2
which shows that th e origin (  0, z  0) is asymptotically stable.
Since  (0)  0, the origin (  0,   0) is asymptotically stable
(z     ( )).
Then the state feedback control law
u   [ f ( )  g ( ) ]  V g ( )  K [   ( )] ( u    v)

11-17
Lemma & Example

Lemma: Consider the system defined by (1), (2). Let  ( )

be a stabilizing state feedback control for (1)
with  (0)  0 and V ( ) be a Lyapunov function
satisfying A . Then the above state feedback control
u stabilizes the origin of (1), (2) with
V ( )  12 [   ( )]2 as a Lyapunov function.

Ex: x1  x12  x13  x2

x 2  u
x1  x12  x13  x2 x1   x1  x13
let x2   ( x1 )   x12  x1

Then V ( x1 )  12 x12  V   x12  x14   x12 , x1  R

Thus

11-18
Example (Continued)
u  x1 ( x12  x13  x2 )  xV1  [ x2   ( x1 )]
 ( 2 x1  1)( x12  x13  x2 )  x1  ( x2  x12  x1 )
Va ( x)  12 x12  12 ( x2  ( x12  x1 ))2
 12 x12  12 ( x2  x12  x1 ) 2

Let’s consider
  f ( )  g ( )
  f a ( , )  g a ( , )u
If g a ( , )  0 over the domain of interest, choose
u 1
g a ( , )
[ua  f a ( , )]
Then   u a . Thus from the previous lemma
u   a ( ,  )  1
 
g a ( , ) 

[ f ( )  g ( ) ]  V g ( )  K [   ( )]  f a ( ,  )
with Va ( ,  )  V ( )  12 [   ( )]2
as a stabilizin g state feedback control and a Lyapunov function.

11-19
Recursive Backstepping

Consider the following strict feedback system

x  f 0 ( x)  g 0 ( x) z1
z1  f1 ( x, z1 )  g1 ( x, z1 ) z2
z 2  f 2 ( x, z1 , z2 )  g 2 ( x, z1 , z2 ) z3 f i and gi depend
 only on x, z1 ,, zi
z k 1  f k 1 ( x, z1 ,, zk 1 )  g k 1 ( x, z1 ,, zk 1 ) zk
z k  f k ( x, z1 ,, zk )  g k ( x, z1 ,, zk )u

where x  R n , z1 to z k are scalars and f 0 to f k vanish at the origin.

We assume g i ( x, z1 ,, zi )  0, for 1  i  K over the domain of interest.

11-20
Recursive procedure

Recursive procedure
x  f 0 ( x)  g 0 ( x) z1
 Consider a stabilizing feedback control z1  0 ( x) with 0 (0)  0 and
V0
V0 ( x) such that x
[ f 0 ( x)  g 0 ( x)0 ( x)]   W ( x)
p.d.f

 Consider
x  f 0 ( x)  g 0 ( x) z1
z1  f1 ( x, z1 )  g1 ( x, z1 ) z2

Then using the previous result, obtain

1 ( x, z1 )  g11 [ x0 ( f 0  g 0 z1 )  Vx0 g 0  K1 ( z1  0 )  f1 ] , K1  0

V1 ( x, z1 )  V0 ( x)  12 [ z1  0 ( x)]2

11-21
Recursive procedure (Continued)

Next consider x  f 0 ( x)  g 0 ( x) z1
z1  f1 ( x, z1 )  g1 ( x, z1 ) z2
z 2  f 2 ( x, z1 , z2 )  g 2 ( x, z1 , z2 ) z3

Then we recognize that

 x    z2  f  g 0 z1   0  fa  f2
   , , f  0  , g    ,
 z1  u  z3  f1   g1  g a  g 2

Thus, similarly, obtain the state feedback control

2 ( x, z1 , z2 )  
1 1
g 2 x
( f 0  g 0 z1 ) 
1
z1
( f1  g1 z2 ) 
V1
z1
g1  K 2 ( z2  1 )  f 2 
and

V2 ( x, z1 , z2 )  V1 ( x, z1 )  12 [ z2  1 ( x, z1 )]2

11-22
Extended Linearization (Gain scheduling method)

 Motivation
 Plant – nonlinear
 Controller – linear
 Design method – classical linearization
 Assumption – no single linear controller satisfies the
performance specification
 Idea – design a set of controllers, each good at a particular
operating point, and switch (schedule) the gains of the controllers
accordingly
 Problem – now we have a nonlinear (piecewise linear) system
with time dependent jump
 Solution – no good tool but some theory is being developed
mostly simulation in the past

11-23
Structure & Examples
gains Gain Operating
Structure Scheduler point

Controller Plant y

-Examples
Linearized model at an operating point,
 Tank system
qin  qin0 , h  h 0
qin 
G p ( s) 
h s 
1
a where  
A(h 0 )
qout
d h qin0 a 2 gh 0
 A( x)dx  qi  a 2 gh
dt 0  0 0

2 A(h )h 2 A(h 0 )h 0
where A(h) is the cross section of the tank at height h,
a is the cross section of the outlet pipe.
11-24
Control Goal

Control goal : h  h 0

h0 +
Gc Gp
-

1
Use PI controller : Gc ( s )  K (1  )
Ti s
Choose K and Ti so that the closed loop system has the natural frequency
 and relative damping  . Then K and Ti should be chosen as
qin0 2  
K  2A(h )  0
0
 K
2h 
2 qin0 2  
Ti    Ti 
 2 A(h 0 )h 0 2 2
Thus for a desired h  h 0 , we schedule the gain appropriately.
Problem : How will the system behave with switches?

11-25
Nonlinear Actuator

 A different angle – nonlinear actuator

r actuator y
Gc () f () Plant G p

Assume
1
f(u) G p ( s) 
( s  1)3
1
Gc ( s )  K (1  )
Ti s

u f (u )  u 4 (nonlinear value) so u 4 is good enough

 (operating point is not at 0)
Large gain
only open or close

11-26
Step Responses

with K  0.15, Ti  1, step responses at different operating conditions are

Introduce gain scheduling through the inverse

f 1 (u ) : f ( f 1 (u ))  u
Since f () may be unknown or for simplicity, we may not want
f 1 (u ) but an estimate fˆ 1 (u ) so that
f ( fˆ 1 (u ))  u
11-27
Approximation

Now we use this approximate inverse for gain scheduling :

r + u yp
Gc (s ) fˆ 1 (u ) f ( fˆ 1 (u )) v G p (s )
-
Now u  v

Approximate fˆ by two linear function

f(u)
16

fˆ 0u3
f ˆ 1  0.433u
Then f (u )  
3 0.0538u  1.139 3  u  16
1.3 2
Domain Domain

11-28
Results

Resulting behavior for the same plant with controller

No measurement of external conditions is required.

11-29
Classification

( ) Scheduling on the operating conditions

scheduler Operating point
gain
+ Gc (s ) plant
-
(  ) Scheduling on the reference signal
gain scheduler
r (t ) + Gc (s ) plant
-

( ) Scheduling on the plant output

gain scheduler

r (t ) y
+ Gc (s ) plant
-

11-30
Issues

In all cases, nonlinear time varying systems, many become unstable

even if at each frozen time all are fine.
The system is time varying
b
G p ( s)  with frozen time
s( s  a)
Controller : Gc ( s )  K (u (t ))(1  Td (u (t )) s )

Consider two systems

x  f (t , x), x  R n (1)
f : R  R n  R n
and x  f (r , x), x  R n , r  R (2)
frozen time system
It is easy to show that (1) can be unstable even if (2) is exponentially stable.

11-31
 Example

Ex: x  A(t ) x
  1  a cos2 t 1  a sin t cos t 
A 2 
, a0
 1  a sin t cos t  1  a sin t 
det A  2  (2  a )  (2  a )  0
For a  2, Re   0. Thus the frozen system is exponentially stable. But
 e( a 1)t cos t e t sin t 
 (t ,0)   ( a 1)t 
 e sin t e t cos t 
So for 1  a, the system is unstable. However if A (t ) is small the above
problem doesn't occur.
Theorem: Consider x  A(t ) x. Suppose A(t ) is differenti able and Re i ( A(t ))  0,
t  0, i. Then 0 is uniformly asym. stable provided
sup A (t ) is sufficient ly small.
t

Proof: See Ch 5 in Nonlinear System Analysis

11-32
Formalization

 A version of scheduling on the output

 y 
x     f ( y, z )  Bu where y  R m , z  R n m , u  R m
 z 
f (0,0)  0 twice conti. differentiable in all variables.
 ueq ( y ) and zeq ( y ) such that
f ( y, z ( y ))  Bu ( y )  0
   eq    eq  
Family of equilibrium point parameterized by y

A linearization at y  y0 is :
d  y  y0  f  y  y0 
 z  z ( y )  ( y , z ( y ))
0    B (u  ueq ( y0 ))
dt  eq 0  x
0 eq
 z  z eq ( y )
0 


Design a controller Ac ( y0 ), Bc ( y0 ), Cc ( y0 )

11-33
Block Diagram

Block diagram
gain scheduler
gains

r e u u y
Gc (s ) plant

ueq H

In the scheduler
x c  Ac ( y (t )) xc  Bc ( y (t ))e
u  Cc ( y (t )) x

gain scheduling on the current output

11-34
Conditions

The theory gives condition under which y   the system is stable.

These conditions are that y   and nonoutput nonlinearity approaches zero

 scheduling variable should vary slowly.

 scheduling variable should capture the plant' s nonlinearities.

Ref : Analysis of Gain Scheduled Control for Nonlinear Plants

by J.S. Shamma, M. Athans
IEEE Tr. on Automatic Cont ol , vol.35, no.8, pp. 898 - 907, Aug.1990

11-35