EE402 - Discrete Time Systems Spring 2018

Lecture 17
Lecturer: Asst. Prof. M. Mert Ankarali

Discrete-time Luenberger Observer

In general the state, x[k], of a system is not accessible and observers, estimators, filters) have to be used to
extract this information. The output, y[k], represents the measurements which is a function of x[k] and u[k].

x[k + 1] = Gx[k] + Hu[k]

y[k] = Cx[k] + Du[k]

A Luenberger observers is built using a “simulated” model of the system and the errors caused by the
mismatched initial conditions x0 6= x̂0 (or other types of perturbations) are reduced by introducing output
error feedback.
Let’s assume that the state vector of the simulated system isx̂[k], then the state space equation of this
synthetic system takes the form

x̂[k + 1] = Gx̂[k] + Hu[k]

ŷ[k] = C x̂[k] + Du[k]

Note that since u[k] is the input that is supplied by the controller, we assume that it is known apriori.
If x[0] = x̂[0] and when there is no model mismatch or uncertainty in the system then we expect that
x[k] = x̂[k] and y[k] = ŷ[k] for all k. When x[0] 6= x̂[0], then we should observe a difference between the
measured and predicted output y[k] 6= ŷ[k]. The core idea in Luenberger observer is feeding the error in the
output prediction y[k] − ŷ[k] to the simulated system via a linear feedback gain.

x̂[k + 1] = Gx̂[k] + Hu[k] + L (y[k] − ŷ[k])

ŷ[k] = C x̂[k] + Du[k]

In order to understand how a Luenberger observer works and to choose a proper observer gain L, we define
an error signal e[k] = x[k] − x̂[k]. The dynamics w.r.t e[k] can be derived as

e[k + 1] = x[k + 1] − x̂[k + 1]

= (Gx[k] + Hu[k]) − (Gx̂[k] + Hu[k] + L (y[k] − ŷ[k]))
e[k + 1] = (G − LC) e[k]

where e[0] = x[0] − x̂[0] denotes the error in the initial condition.
If the matrix (G − LC) is stable then the errors in initial condition will diminish eventually. Moreover, in
order to have a good observer/estimator performance the observer convergence should be sufficiently fast.

17-1
17-2 Lecture 17

Observer Gain & Pole Placement

Similar to the state-feedback gain design, the fundamental principle of “pole-placement” Observer design is
that we first define a desired closed-loop eigenvalue set and compute the associated desired characteristic
polynomial.

E ∗ = {λ∗1 , · · · , λ∗n }
p∗ (z) = (z − λ∗1 ) · · · (z − λ∗n )
= z n + a∗1 z n−1 + · · · + a∗n−1 z + a∗n

The necessary and sufficient condition on arbitrary observer pole-placement is that the system should be
fully Observable. Then, we tune L such that

det (zI − (G − LC)) = p∗ (z)

Direct Design of Observer Gain

If n is small, the most efficient method could be the direct design.

Example: Consider the following DT system
   
1 0 1
x[k + 1] = x[k] + u[k]
0 2 1
 
y[k] = 1 −1 u[k]

Design an observer such that estimater poles are located at λ1,2 = 0 (Dead-beat Observer)
Solution: Desired characteristic equation can be computed as

p∗ (z) = z 2
 
l2
Let L = , then the characteristic equation of (G − LC) can be computed as
l1
 
z − 1 + l2 −l2
det (zI − (G − LC)) = det
l1 z − 2 − l1
= z 2 + z(l2 − l1 − 3) + (l1 − 2l2 + 2)

If we match the equations

l2 − l1 = 3
−l1 + 2l2 = 2
l2 = −1
l1 = −4
 
−1
Thus L =
−4
Lecture 17 17-3

Design of Observer Gain Using Observable Canonical Form

Let’s assume that the state-space representation is in observable canonical form

   
0 0 ··· 0 −an (bn − b0 an )
 1 0 · · · 0 −an−1   (bn−1 − b0 an−1 ) 
   
G =  ... ... .. .. ..
 , H=
..
   
 . . .   . 

 0 0 ··· 0 −a2   (b2 − b0 a2 ) 
0 0 ··· 1 −a1 (b1 − b0 a1 )
 
C = 0 0 ··· 0 1 , D = b0
 
ln
Let L =  ... , then (G − LC) takes the form
 

l1
 
0 0 ··· 0 −an  

 1 0 ··· 0 −an−1 
 ln
(G − LC) =  .. .. .. .. .. ..   0 ···

− 0 0 1
  
 . . . . .  . 
 0 0 ··· 0 −a2  l1
0 0 ··· 1 −a1
 
0 0 ··· 0 −(an + ln )
 1 0
 ··· 0 −(an−1 + ln−1 ) 

=  ... ... .. .. ..
 
 . . . 

 0 0 ··· 0 −(a2 + ln ) 
0 0 ··· 1 −(a1 + ln )

The characterstic equation of G − LC is simply given as

p(z) = z n + (a1 − l1 )z n−1 + · · · + (an−1 − ln−1 )z + (an − ln )

Let’s assume that desired p∗ (z) is equal to

p(z) = z n + a∗1 z n−1 + · · · + a∗n−1 z + a∗n

Then, the observer gain L is computed as

a∗n − an
 

L∗ = 
 .. 
. 
∗
a1 − a1

If the system is not in Observable canonical form, we can find a transformation that outputs the Observable
canonical form representation The Observability matrix of a state-space representation is given as
 
C
 CG 
O=
 
.. 
 . 
CG−1
17-4 Lecture 17

Let’s define a transformation matrix Q as follows:

−1
Q = (W O) , x[k] = Qx̄[k]
x̄[k + 1] = Q GQ x̄[k] + Q−1 Hu[k]
 −1 

y[k] = CQx̄[k] + Du[k]

where
 
an−1 an−2 ··· a1 1

 an−2 an−3 ··· 1 0 

.. .. . ..
W = ..
 
 . . . 

 a1 1 
1 0 ··· 0

Then it is given that

 
0 0 ··· 0 −an

 1 0 ··· 0 −an−1 

Ḡ = Q −1
GQ = 
 .. .. .. .. .. 
 . . . . . 

 0 0 ··· 0 −a2 
0 0 ··· 1 −a1
 
C̄ = CQ = 0 0 ··· 0 1

Let’s compute C̄Q−1

C̄Q−1 = C̄W O
 
an−1 an−2 ··· a1 1
 an−2 an−3 ··· 1 0 
  .. .. . ..

 T
= 0 0 ··· 0 1  .. O

 . . . 
 a1 1 
1 0 ··· 0
−1
  
= 1 0 ··· 0 0 C CG ··· CG
=C

A similar approach (but longer) can be used to show that Q−1 GQ = Ḡ.
We know how to design a observer gain L̄ for the Observable canonical form. Given L̄, Observer gain w.r.t.
original state-space representation is computed as

L = QL̄

Example 2: Consider the following DT system

   
1 0 1
x[k + 1] = x[k] + u[k]
0 2 1
 
y[k] = 1 −1 u[k]

Design an observer using the Observable canonical form such that estimater poles are located at λ1,2 = 0
(Dead-beat Observer)
Lecture 17 17-5

Solution: Characteristic equation of G can be derived as

 
z−1 0
det = z 2 − 3z + 2
0 z−2

Observability matrix can be computed as

 
1 −1
O=
1 −2

The matrix W can be computed as

 
−3 1
W =
1 0

Transformation matrix Q can be computed as

−1
Q = W OT
  −1  −1  
−3 1 1 −1 −2 1 −1 −1
= = =
1 0 1 −2 1 −1 −1 −2

Given that desired characteristic polynomial is p∗ (z) = z 2 , L̄ of observable canonical from can be computed
as
 
−a2
L̄ =
−a1
 
−2
=
3

Finally Observer Gain L can be computed as

 
−1
L = QL̄ =
−4

Not surprisingly the result is same with the one found in first example.

Closed-Loop Observer & State-Feedback

In the state-feedback control policy the input is ideally defined by the following law

u[k] = −Kx[k]

However, as mentioned in Observer lecture, in general we don’t have direct access to the all states of the
system. In this case, we learnt how to design an Observer/Estimator of the states. In this respect, it is
natural to assume that in a closed-loop system, the control policy that define the input should depend on
the estimated states

u[k] = −K x̂[k]

However the important question how this coupling affect the closed-loop behavior, and even deeper question
can be even use such a policy. The advantage of LTI systems is that state-feedback gain, and observer
gain can be seperatelly designed and we guarntee a stable closed-loop performance. In this section, we will
17-6 Lecture 17

analyze the coupled system Equations of motion for the closed-loop observer & state-feedback based control
system is given below

x[k + 1] = Gx[k] + Hu[k]

y[k] = Cx[k] + Du[k]
x̂[k + 1] = Gx̂[k] + Hu[k] + L (y[k] − ŷ[k])
ŷ[k] = C x̂[k] + Du[k]
u[k] = −K x̂[k]

If we eliminate u[k] and ŷ[k] we obtain following dynamical representation

x[k + 1] = Gx[k] − HK x̂[k]

x̂[k + 1] = Gx̂[k] − HK x̂[k] + LC (x[k] − x̂[k])
y[k] = Cx[k] − DK x̂[k]

Now let’s replace x̂[k] with e[k] = x[k] − x̂[k]

x[k + 1] = (G − HK)x[k] + HKe[k]

e[k + 1] = (G − LC)e[k]
y[k] = (C − DK)x[k] + DKe[k]
 
x[k]
Now let’s defina a state for the whole system, z[k] = then the state-space representation is given
e[k]
by
 
(G − HK) HK
z[k + 1] = z[k]
0n×n (G − LC)
 
y[k] = (C − DK) DK z[k]

The system matrix is in block diagonal form and the eigenvalues of this new system matrix is find by taking
the union of eigenvalues of (G − HK) and eigenvalues of (G − LC). Thus a seperate pole-placement can be
performed for the state-feedback controller and the observer.