A Luenberger observer estimates a state using a plant model.
Let us consider a system
x˙y=Ax+Bu+v=Cx+w,
where
x,
y, and
u are a state, output, and input;
A,
B, and
C stand for a system matrix, input matrix, and observation matrix;
v and
w are a process noise and observation noise.
The state-space representation model enables the prediction of a state
x^˙y^=Anx^+Bnu=Cnx^,
where a subscript
n denotes a nominal model.
An observer compensate a state comparing a system output and predicted state.
x^˙y^=Anx^+Bnu+L(y−y^)=Cnx^,
where
L is a observer gain.
The dynamics of a plant and observer provides the dynamics of an estimation error:
x˙−x^˙=(Ax+Bu+v)−(Anx^+Bnu+L(y−y^))=An(x−x^)+Aex+Beu−L(Cx+w−Cnx^)+v=An(x−x^)+Aex+Beu−LCn(x−x^)−LCex+v−Lw=(An−LCn)(x−x^)+(Ae−LCe)x+Beu+v−Lw,
where a subscript
e denotes a modeling error as
Xe=X−Xn.
Then, we can found
e˙e=(An−LCn)e+(Ae−LCe)x+Beu+v−Lw,≡x−x^.
Let us consider an observer-based feedback controller
u=uref−Fx^.
This controller affects a state and estimation error as
x˙e˙=Ax+B(uref−Fx^)+v=(A−BF)x+BFe+Buref+v=(An−LCn)e+(Ae−LCe)x+Be(uref−Fx^)+v−Lw=(An−LCn+BeF)e+(Ae−LCe−BeF)x+Beuref+v−Lw.
It provides a following state-space representation
xE˙=AExE+BEuref+[II]v+[0−L]w,
where
xEAEBE≡[xe]≡[A−BFAe−LCe−BeFBFAn−LCn+BeF]≡[BBe].
If a modeling error is sufficiently small enough,
AE=[A−BF0BFAn−LCn]
holds, and the characteristics of control and observation are separated.