24 Feedback Linearization
24 Feedback Linearization
24 Feedback Linearization
2011
- 1 -
24 Feedback Linearization
Similarly to the approach taken in sliding mode control, in this section we wish to exploit
the form of the system equations to modify the dynamics to something more convenient.
We consider a class of nonlinear systems of the form
where :
n
f D9 and :
n p
G D
9 are defined on a domain
n
Dc9 , which contains
the origin, and pose the question whether there exists a state feedback control
and a change of variables
that transforms the nonlinear system into an equivalent linear system.
If the answer to this question is positive, we can induce linear behavior in nonlinear
systems and apply the large number of tools and the well established theory of linear
control to develop stabilizing controllers.
24.1 Motivation
Example 24.1: To introduce the idea of feedback linearization, let us start with the
problem of stabilizing the origin of the pendulum equation.
( )
1 2
2 1 2
sin sin
x x
x a x bx cu o o
=
( = + +
If we choose the control
( )
1
sin sin
a v
u x
c c
o o ( = + +
We can cancel the nonlinear term ( )
1
sin sin a x o o ( +
. This cancellation results in the
linear system
( ) ( )
( )
x f x G x u
y h x
= +
=
( ) ( ) u x x v o | = +
( ) z T x =
- 2 -
1 2
2 2
x x
x bx v
=
= +
Thus, the stabilization problem for the nonlinear system has been reduced to a
stabilization problem for a controllable linear system. We can proceed to design a
stabilizing linear state feedback control
1 1 2 2
v k x k x =
to locate the eigenvalues of the closed loop system
( )
1 2
2 1 1 2 2
x x
x k x k b x
=
= +
in the open left half plane. The overall state feedback control law is given by
( ) ( )
1 1 1 2 2
1
sin sin
a
u x k x k x
c c
o o
| |
( = + +
|
\ .
Clearly, we should not expect to be able to cancel nonlinearities in every nonlinear
system. There must be a certain structural property of the system that allows us to perform
such cancellation. Therefore, the ability to use feedback to convert a nonlinear state
equation into a controllable linear state equation by canceling nonlinearities requires the
nonlinear space equation to have the structure
Where A is n n , B is n p , the pair ( ) , A B is controllable, the functions :
n p
o 9 9
and :
n p p
9 9 are defined in the domain
n
Dc9 that contains the origin, and the
matrix ( ) x is nonsingular for every x D e . If the state equation takes the form
( ) ( ) x Ax B x u x o ( = +
, then we can linearize it via the state feedback
where ( ) ( )
1
x x |
\
|
= =
=
` `
`
and the nonlinearities can be cancelled by the control
2
1
2
1
cos
u x v
a x
= +
which is well defined for
2
2 2 x t t < < . The state equation in the new coordinates can
be found by inverting the transformation to express ( )
1 2
, x x in the term of ( )
1 2
, z z ; that is,
1 1
1 2
2
sin
x z
z
x
a
=
| |
=
|
\ .
which is well defined for
2
a z a < < . The transformed state equation is given by
( ) ( ) u x x Kx o | =
- 4 -
( )
1 2
1 2 2
2 1
cos sin
z z
z
z a z u
a
=
| |
| |
= +
| |
\ . \ .
which is in the required form to use state feedback
Definition 24.1: A nonlinear system
where :
n
f D9 and :
n p
G D
9 are sufficiently smooth on a domain
n
Dc9 , is
said to be feedback linearizable (or input state linearizable) if there exists a
diffeomorphism :
n
T D9 such that ( )
z
D T D = contains the origin and a change of
variables ( ) z T x = transforms the system ( ) ( ) x f x G x u = + into the form
with ( ) , A B controllable and ( ) x nonsingular for all x D e .
Definition 24.2: f is called smooth if f C
=
=
| |
=
|
\ .
While the state equation is linear, solving a tracking control problem for y is still
complicated by the nonlinearity of the output equation.
24.1.1 Transformation Matrix ( ) T x
With the help of a Matrix ( ) T x we want to transform
into the new system
( ) ) ( ) ( x u x B Az z o + = `
Given the previous system with a coordinate transformation ( ) z T x = we derive
( )
( )
( ) ) ( ) ( ) (
) ( ) (
) ( ) (
) ( ) (
x u x B x AT
x u x B Az
u x G x f
x
x T
x
x
x T
z
o
o
+
+
+
c
c
=
c
c
=
=
=
`
`
and obtain a system of partial differential equations
=
c
c
=
c
c
) ( ) (
) ( ) ( ) ( ) (
x B x G
x
T
x x B x AT x f
x
T
o
By solving this set of partial differential equations for T(x), the transformation is found.
( ) ( ) x f x G x u = +
- 6 -
Remark: ( ) T x is not uniquely defined by this system of differential equations as shown
when we consider any linear transformation z Mz = . Then ( ) ( ) T x MT x = will also
satisfy this system but with
1
A MAM
= and B MB = .
We use this fact to choose the matrix A and B to be in the canonical controllability form.
For simplicity we consider this for a single-input system. In this case, the control
canonical form is as follows.
(
(
(
(
=
1 1 0
1
0 1 0
n
C
A
o o o .
.
.
,
(
(
(
(
=
1
0
0
.
C
B
Where the
i
o are the coefficients of the characteristic polynomial of A:
=
=
1
0
) det(
n
i
i
i
s A sI o
Note that ) , (
C C
B A is feedback equivalent to a chain of integrators. i.e. with the mapping
=
+
1
0
n
i
i
i
s u u o , the system is transformed to a chain of integrators. Thus we consider
the following system matrices:
(
(
(
(
=
0 0
1
0 1 0
0
. .
.
.
A ,
(
(
(
(
=
1
0
0
0
.
B
Then with
we obtain
2
0
( )
( ) ( ) ( ) ( )
( )
( ) ( )
n
T x
T
AT x B x x f x
T x x
x x
o
o
(
(
c
(
= =
( c
(
( )
( )
( )
1
n
T x
T x
T x
(
(
=
(
(
- 7 -
) (
) (
0
0
) (
0
x G
x
T
x
x B
c
c
=
(
(
(
(
.
Solving this system of partial differential equations we can find ( ) T x .
The following theorem, stated without proof, gives necessary and sufficient conditions for
feedback linearizability of a single-input affine-in-control nonlinear system.
We need some notation first. Let f, g be smooth vector fields on . The Lie-bracket of f
and g is defined as .
We also define
A distribution D is a mapping that assigns to each point in the state space a subspace of
the tangent space at that point. In other words, a distribution is a family of smooth vector
fields that span a subspace of at each point. A distribution is involutive if the Lie
bracket of any two vector fields in the distribution lies in the distribution. To wit, if D is
spanned by , then for each i, j, x, we have
where are smooth functions. Note that this is a natural generalization of the
concept of linear dependence of vectors. We define the following special distributions:
Theorem 24.1: A single input system at a point x if and only if
- is a linearly independent family of vectors, and
- is involutive.
24.2 Input-Output Linearization
Consider the single-input-single-output system
where f , g , and h are sufficiently smooth in a domain
n
Dc9 . The mappings
:
n
f D9 and :
n
g D9 are called vector fields on D.
( ) ( )
( )
x f x g x u
y h x
= +
=
- 8 -
Derive conditions which allow us to transform the system such that the input output map
is linear.
The derivative y is given by
where
is called the Lie Derivative of h with respect to f .
If ( ) 0
g
L h x = , then ( )
f
y L h x = , independent of u . If we continue to calculate the second
derivative of y , denoted by
( ) 2
y , we obtain
Once again, if ( ) 0
g f
L L h x = , then
( )
( )
2 2
f
y L h x = , independent of u . Repeating this
process, we see that if ( ) h x satisfies
( )
1
0
i
g f
L L h x
= , 1, 2,..., 1 i = ; ( )
1
0
g f
L L h x
=
then u does not appear in the equations of y , y ,,
( ) 1
y
and appears in the equation of
( ) p
y with a nonzero coefficient:
The forgoing equation shows clearly that the system is input-output linearizable, since the
state feedback control
reduces the input-output map to
( ) ( ) ( ) ( )
f g
h
y f x g x u L h x L h x u
x
c
( = + = +
c
( ) ( )
f
h
f x L h x
x
c
=
c
( )
( )
( ) ( ) ( ) ( )
2 2
f
f g f
L h
y f x g x u L h x L L h x u
x
c
( = + = +
c
( )
( ) ( )
1 p
f g f
y L h x L L h x u
= +
( )
( )
1
1
f
g f
u L h x v
L L h x
( = +
- 9 -
which is a chain of integrators. In this case, the integer is called the relative degree
of the system.
Example 24.3: Consider the controlled van der Pol equation
( )
1 2
2
2 1 1 2
1
x x
x x x x u c
=
= + +
with output
1
y x = . Calculating the derivatives of the output, we obtain
( )
1 2
2
2 1 1 2
1
y x x
y x x x x u c
= =
= = + +
Hence, the system has relative degree two in
2
9 . For the output
2
1 2
y x x = + ,
( )
2
2 2 1 1 2
2 1 y x x x x x u c
(
= + + +
and the system has relative degree one in
{ }
2
0 2
0 D x x = e9 =
Example 24.4: Correspondence to relative degree for Linear Systems
Consider a linear system represented by the transfer function
where m n < and 0
m
b = . A state model for the system can be taken as
where
( )
y v
=
( )
1
1 0
1
1 0
...
...
m m
m m
n n
n
b s b s b
H s
s a s a
+ + +
=
+ + +
x Ax Bu
y Cx
= +
=
- 10 -
0 1 1
0 1 0
0 0 1 0
1 0
0 0 1
m n
n n
A
a a a a
(
(
(
(
=
(
(
(
(
(
,
1
0
0
0
1
n
B
(
(
(
(
=
(
(
(
(
| |
0
1
0 0
m
n
C b b
=
This linear space model is a special case of ( ) ( ) x f x g x u = + , ( ) y h x = , where
( ) f x Ax = , g B = , and ( ) h x Cx = . To check the relative degree of the system, we
calculate the derivative of the output. The first derivative is
If 1 m n = , then
1
0
n
CB b
= = and the system has relative degree one. Otherwise,
0 CB = and we continue to calculate the second derivative
( ) 2
y . Noting that CA is a row
vector obtained by shifting the elements of C one position to the right, while
2
CA is
obtained by shifting the elements of C two positions to the right, and so on, we see that
1
0
i
CA B
= =
Thus, u appears first in the equation of
( ) n m
y
, given by
and the relative degree of the system is n m (the difference between the degrees of the
denominator and numerator polynomials of ( ) H s ).
Now let
y CAx CBu = +
( ) 1 n m n m n m
y CA x CA Bu
= +
- 11 -
where ( )
1
x | to ( )
n
x
|
are chosen such that
( ) 0
i
g x
x
| c
=
c
, for 1 i n s s
This condition ensures that when we calculate
the term u cancels out. It is now easy to verify that the change of variables ( ) z T x =
transforms the system into
( )
( ) ( )
0
,
C C
C
f
A B x u x
y C
q q
o
=
( = +
=
where
p
e9 ,
n
q
e9 , ( ) , ,
C C C
A B C is a canonical form representation of a chain of
integrators, where
( ) ( )
1
g f
x L L h x
= and ( )
( )
( )
1
f
g f
L h x
x
L L h x
o
=
We have kept o and expressed in the original coordinates. These functions are
uniquely determined in terms of f , g , and h . They are independent of the choice of | .
They can be expressed in the new coordinates by setting
( ) ( ) ( )
1
0
, T z o q o
= and ( ) ( ) ( )
1
0
, T z q
=
( )
( )
( )
( )
( )
( )
( )
1
1
n
def def
f
x
x x
z T x
h x x
L h x
|
| | q
(
(
(
(
(
(
(
(
(
= = = =
(
(
(
(
(
(
(
(
(
( ) ( ) f x g x u
x
|
q
c
( = +
c
( ) ( )
( )
1
0
,
x T z
f f x
x
|
q
=
c
=
c
- 12 -
which, of course, will depend on the choice of | . In this case, the equation can be
rewritten as
The three equations of the new system are said to be in the normal form. This form
decomposes the system into an external part and an internal part q . The external part is
linearized by the state feedback control
where ( ) ( )
1
x x |
c
c
+
c
c
=
+ +
c
c
=
_
`
Where 0
1
> k follows from local continuity and differentiability of
0
f , and 0
2
> k follows
from P. By choosing k large enough, we guarantee 0 V < .
Remark 1: This is a local result
Remark 2: If the origin 0 q = is globally asymptotically stable for ( )
0
, 0 f q q = one
might think that system can be globally stabilized. This is not the case.
Remark 3: Global stability can be guaranteed if the system is input to state stable and
linearizable.
Remark 4: One may think that the system can be globally stabilized, or at least semi-
globally stabilized, by designing the linear feedback control v K = to assign the
eigenvalues of ( ) A BK far to the left in the complex plane so that the solution of
( ) A BK = decay to zero arbitrarily fast. Then, the solution of ( )
0
, f q q = will
quickly approach the solution of ( )
0
, 0 f q q = , which is well behaved, because its origin is
globally asymptotically stable. But consider the following example.
Example 24.5: Consider the third-order system
- 14 -
The linear feedback control
assigns the eigenvalues of
at k and k . The exponential matrix
shows that as k , the solution ( ) t will decay to zero arbitrarily fast. Notice,
however, that the coefficient of the ( ) 2,1 element of the exponential matrix is a quadratic
function of k . It can be shown that the absolute value of this element reaches a maximum
value / k e at 1 k . While this term can be made to decay to zero arbitrarily fast by
choosing k large, its transient behavior exhibits a peak of the order of k . The interaction
of peaking with nonlinear growth could destabilize the system. In particular, for the initial
states ( )
0
0 q q = , ( )
1
0 1 = , and ( )
2
0 0 = , we have ( )
2
2
kt
t k te
= and
During the peaking period, the coefficient of
3
q is positive, causing ( ) t q to grow.
Eventually, the coefficient of
3
q will become negative, but that might not happen soon
enough , since the system might have a finite escape time. Indeed the solution
shows that if
2
0
1 q > , the system will have a finite escape time if k is chosen large
enough.
( )
3
2
1 2
2
1
1
2
v
q q
= +
=
=
2
1 2
2
def
v k k K = =
2
0 1
2
A BK
k k
(
=
(
( )
( )
( )
2
1
1
kt kt
A BK t
kt kt
kt e te
e
k te kt e
( +
=
(
( )
2 3
1
1
2
kt
k te q q
=
( )
( )
2
2 0
2
0
1 1 1
kt
t
t kt e
q
q
q
=
( + + +
- 15 -
Remark: It is useful to know that the zero dynamics can be characterized in the original
coordinates. Noting that
we see that if the output is identically zero, the solution of the state equation must be
confined to the set
and the input must be
Example 24.6: The system
has an open-loop equilibrium point at the origin. The derivatives of the output are
Therefore, the system has relative degree two in
3
9 . Using ( ) 1
g f
L L h x = and
( )
2
1 3 f
L h x x x = , we obtain
1 = and ( )
1 3
x x x o =
( ) ( ) ( ) ( ) ( )
0 0 y t t u t x t o
( ) ( ) ( ) { }
* 1
0
0
f f
Z x D h x L h x L h x
= e = = = =
( ) ( )
*
*
def
x Z
u u x x o
e
= =
2
3
1 1 2
3
2 3
3 1 3
2
2
1
x
x x u
x
x x
x x x u
y x
+
= +
+
=
= +
=
2 3
3 1 3
y x x
y x x x u
= =
= = +
- 16 -
To characterize the zero dynamics, restrict to
and take ( )
*
0 u u x = = . This process yields
which shows that the system is minimum phase. To transform it into the normal form, we
want to choose a function ( ) x | such that
( ) 0 0 | = , ( ) 0 g x
x
| c
=
c
and
is a diffeomorphism on some domain containing the origin. The partial differential
equation
can be solved by separating variables to obtain
which satisfies the condition ( ) 0 0 | = . The mapping ( ) T x is a global diffeomorphism, as
can be seen by the fact that for any
3
z e9 , the equation ( ) T x z = has a unique solution.
Thus, the normal form
is defined globally.
{ }
*
2 3
0 Z x x x = e9 = =
1 1
x x =
( ) ( )
2 3
T
T x x x x | ( =
2
3
2
1 3 3
2
0
1
x
x x x
| | + c c
+ =
c + c
( )
1
1 3 3
tan x x x x |
= + +
( )
( )
2
1 2
2 3 2 2
2
1 2
1
2 2 2 2
1
2
tan 1
1
tan
x
u
y
q q
| | +
= + + +
|
+
\ .
=
= + + +
=
- 17 -
Remark: (Robustness) Feedback linearization is based on exact cancellation of
nonlinearities, which is, in practice, not often practically possible. Most likely we have
only approximations o ,
| and ( )
f
A B x KT
q q
o o |
=
(
= +
( )
( ) ( )
, f
A BK B z
q q
o
=
= +
( ) ( )
( ) ( ) 2 2 2
z KT K T T o o o | | |
(
=
3
x ax bx u = +
3
x Kx bx =