Chapter 11: Nonlinear Observers

1 Nonlinear Observability
Consider the system

nl
_
x = f(x) +g(x)u f : R
n
R
n
, g : R
n
R
y = h(x) h : R
n
R
(1)
Notation:
x
u
(t, x
0
): solution of (1) at time t originated by u with initial state x
0
.
y(x
u
(t, x
0
): output y when the state x is x
u
(t, x
0
).
Clearly
y(x
u
(t, x
0
)) h(x
u
(t, x
0
))
Denition 1 : A pair of states (x
1
0
, x
2
0
) is said to be distinguishable if there exists
an input function u such that
y(x
u
(t, x
1
0
)) y(x
u
(t, x
2
0
))
Denition 2 :
nl
is said to be (locally) observable at x
0
R
n
if there exists
a neighborhood U
0
of x
0
such that every state x =x
0
is distinguishable
from x
0
. It is said to be locally observable if it is locally observable at each
x
0
R
n
.
This means that
nl
is locally observable in a neighborhood U
0
R
n
if
there exists an input u R such that
y(x
u
(t, x
1
0
)) y(x
u
(t, x
2
0
)) t [0, t] x
1
0
= x
2
0
1
Remarks There is no requirement in Denition 2 that distinguishability
must hold for all functions.
In the following theorem we consider a system of the form:

nl
_
x = f(x) f : R
n
R
n
y = h(x) h : R
n
R
(2)
Theorem 1 The state space realization (2) is locally observable in a neighbor-
hood U
0
D containing the origin, if
rank
_
_
_
_

_
h
.
.
.
L
n1
f
h
_

_
_
_
_
= n x U
0
(3)
Example 1 (Linear time-invariant)

l
_
x = Ax
y = Cx.
Then h(x) = Cx and f(x) = Ax, and we have
h(x) = C
L
f
h = (
h
x
x) = (CAx) = CA
.
.
.
L
n1
f
h = CA
n1
and therefore
l
is observable if and only if S = {C, CA, CA
2
, , CA
n1
} is
linearly independent or, equivalently, if rank(O) = n.
2
Example 2 Consider the following state space realization:

nl
_
_
_
x
1
= x
2
(1 u)
x
2
= x
1
y = x
1
which is of the form
_
x = f(x) +g(x)u
y = h(x)
with
f(x) =
_
x
2
x
1
_
, g(x) =
_
x
2
0
_
, h(x) = x
1
.
If u = 0, we have
rank({h, L
f
h}) = rank({[1 0], [0 1]}) = 2
and thus
_
x = f(x)
y = h(x)
is observable according to Denition 2. Assume now that u = 1. We obtain
the following dynamical equations:
_
_
_
x
1
= 0
x
2
= x
1
y = x
1
.
A glimpse at the new linear time-invariant state space realization shows that
observability has been lost.
3
1.1 Nonlinear Observers
There are several ways to approach the nonlinear state reconstruction prob-
lem, depending on the characteristics of the plant. We now discuss two rather
dierent approaches, each applicable to a particular class of systems.
2 Observers with Linear Error Dynamics
Idea:
(i) Find an invertible coordinate transformation that linearizes the state
space realization.
(ii) Design an observer for the resulting linear system.
(iii) Recover the original state using the inverse coordinate transformation
dened in (i).
More explicitly, suppose that given a system of the form
_
x = f(x) +g(x, u) x R
n
, u R
y = h(x) y R
(4)
there exist a dieomorphism T() satisfying
z = T(x), T(0) = 0, z R
n
(5)
and such that, after the coordinate transformation, the new state space real-
ization has the form
_
z = A
0
z + (y, u)
y = C
0
z y R
(6)
4
where
A
0
=
_

_
0 0 0
1 0 0
0 1 0
.
.
.
.
.
.
0 0 1 0
_

_
, C
0
= [0 0 0 1], =
_

1
(y, u)

1
(y, u)
.
.
.
.
.
.

n
(y, u)
_

_
. (7)
then, under these conditions, an observer can be constructed according to
the following theorem.
Theorem 2 (Marino-Tomei) If there exist a coordinate transformation mapping
(4) into (6), then, dening

z = A
0
z + (y, u) K(y (C
0
z
n
)), z R
n
(8)
x = T
1
( z) (9)
such that the eigenvalues of (A
0
+ KC
0
) are in the left half of the complex
plane, then x x as t .
Proof: Let z = z z, and x = x x. We have

z = z

z
= [A
0
z +(y, u)] [A
0
z +(y, u) K(y z
n
)]
= (A
0
+KC
0
) z.
If the eigenvalues of (A
0
+KC
0
) have negative real part, then we have that
z 0 as t . Using (9), we obtain
x = x x
= T
1
(z) T
1
( z) 0 as t
5
Example 3 Consider the following dynamical system
_
_
_
x
1
= x
2
+ 2x
2
1
x
2
= x
1
x
2
+ x
3
1
u
y = x
1
(10)
and dene the coordinate transformation
_
z
1
= x
2

1
2
x
2
1
z
2
= x
1
.
In the new coordinates, the system (10) takes the form
_
_
_
z
1
= 2y
3
+y
3
u
z
2
= z
1
+
5
2
y
2
y = z
2
which is of the form (6) with
A
0
=
_
0 0
1 0
_
, C
0
= [0 1], and =
_
2y
3
+ y
3
u
5
2
y
2
_
.
The observer is

z = A
0
z +(y, u) K(y z
2
)
_

z
1

z
2
_
=
_
0 0
1 0
_ _
z
1
z
2
_
+
_
2y
3
+y
3
u
5y
2
/2
_
+
_
K
1
K
2
_
(y z
2
).
The error dynamics is
_
z
1
z
2
_
=
_
0 K
1
1 K
2
_ _
z
1
z
2
_
Thus, z 0 for any K
1
, K
2
> 0.
6
3 Lipschitz Systems
Consider a system of the form:
_
x = Ax +f(x, u), A R
nn
y = Cx C R
1n
(11)
where f is Lipschitz in x on an open set D R
n
, i.e.,
f(x
1
, u

) f(x
2
, u

) x
1
x
2
x D. (12)
Now consider the following observer structure

x = A x + f( x, u) +L(y C x) (13)
where L R
n1
.
Theorem 3 Given the system (11) and the corresponding observer (13), if the
Lyapunov equation
P(A LC) + (A LC)
T
P = Q (14)
where P = P
T
> 0, and Q = Q
T
> 0, is satised with
<

min
(Q)
2
max
(P)
(15)
then the observer error x = x x is asymptotically stable.
7
Proof:
.
x = x
.
x
= [Ax +f(x, u)] [A x +f( x, u) +L(y C x)]
= (A LC) x +f(x, u) f( x, u).
To see that x has an asymptotically stable equilibrium point at the origin,
consider the Lyapunov function candidate:
V ( x) x
T
P x.

V ( x) =
.
x
T
P x + x
T
P x
.
x
=
.
x
T
Q x + 2 x
T
P[f(x, u) f( x, u)]
but
x
T
Q x
min
(Q) x
2
and
f(x
1
, u

) f(x
2
, u

) x
1
x
2
x D
so

V ( x)
min
(Q) x
2
+ 2 x
T
P[f(x, u) f( x, u)]

V ( x)
min
(Q) x
2
+ 2 x P f(x, u) f( x, u)

V ( x)
min
(Q) x
2
+ 2 x P x

V ( x)
min
(Q) x
2
+ 2P x
2

V ( x) {
min
(Q) 2P} x
2
Therefore,

V is negative denite, provided that

min
(Q) > 2P
or, equivalently
<

min
(Q)
2
max
(P)
8
Example 4 Consider the following system:
_
x
1
x
2
_
=
_
0 1
1 2
_ _
x
1
x
2
_
+
_
0
x
2
2
_
y = [1 0]
_
x
1
x
2
_
we have that
L =
_
0
2
_
A LC =
_
0 1
1 2
_
Solving the Lyapunov equation
P(A LC) + (A LC)
T
P = Q
with Q = I, we obtain
P =
_
1.5 0.5
0.5 0.5
_
which is positive denite. The eigenvalues of P are
min
(P) = 0.2929, and

max
(P) = 1.7071. We now consider the function f.
Aside
x

1
=
_

1

2
_
, x

2
=
_

1

2
_
we have that
f(x

1
) f(x

2
)
2
=
_
(
2
2

2
2
)
2
= |
2
2

2
2
|
= |(
2
+
2
)(
2

2
)|
2|
2
| |
2

2
| = 2k
2

2kx

1
x

2
for all x satisfying |
2
| < k. Thus, = 2k and f is Lipschitz x = [
1

2
]
T
:
|
2
| < k, and we have
= 2k <
1
2
max
(P)
or
k <
1
6.8284
9