Linear State Observer

(Deterministic Estimation)

- Luenberger Observer
- Reduced Order Observer
Introduction
Observer
- The poles of a controllable system can be assigned arbitrary by linear state
feedback. It means that whole state variables of the system should be known.
- However, in many practical situation we cannot measure all state variables
directly.
- Often, the state variables must be estimated from measurable outputs, and
the number of outputs is normally fewer than the number of state variables.
- In these notes, we discuss the state estimation problem from a deterministic
point of view; i.e., we do not have any uncertainty about model parameters
and system inputs.

2
Introduction
We consider a system described by
dx
= Ax + Bu , x(t0 ) = x0 ,
dt
y = cx
where x : n - dimensional state vector
u : m - dimensional state vector
y : r - dimensional state vector

It is assumed that the system is controllable and observable.

(Controllability is not essential.)

3
Open-loop observer (Real Time Simulator)

Plant
x y
x = Ax + Bu c

Computer
u x̂ y
xˆ = Axˆ + Bu c

e = x − xˆ
x − xˆ = Ax + Bu − Axˆ − Bu
= A( x − xˆ )  e = Ae, e0 = x0 − xˆ0
If A is stable, e(0) will be decayed out.
e(t ) → 0 as t → 
Note) i) If A is stable xˆ → x
ii) If ( A, B ) is not known, or ( A, B ) is incorrectly known
→ convergence will not be garuanteed
iii) Disturbance is introduced,

4
Closed-Loop Observer (Luenberger observer)
➢ Full order observer
An obvious mistake in the open loop observer is that it does not make use
of the system output, y, which includes a lot of information on the state
vector, x. so, let us consider an estimator which has the error between the
plant output and predicted output Cxˆ as an input.
 xˆ = Axˆ + Bu + F ( y − yˆ ), xˆ (t0 ) = 0
 (*)
 yˆ = Cxˆ
where F is an n  r matrix ( observer gain )
x − xˆ = Ax + Bu − Axˆ − Bu − F (Cx − Cxˆ )
= ( A − FC )( x − xˆ )
 e = AF e, e(0) = e0
where AF  A − FC
lim e → 0 if AF is stable.
t →

 F 를 잘 선택함으로서 AF 를 stable하게 할 수 있다.

5
Closed-Loop Observer (Luenberger observer)
Since the system is observable, the rank of the observability matrix is n.
Noting that
M T =  C T AT C T ( AT ) n −1 C T  ,
" ( A, C ) is observable " is equivalent to "( AT , C T ) is controllable"
( This is known as the duality between controllability and observability ) .

Note) Can F stabilize AF ?

Ans) Yes, if ( A, C )is observable.
(Recall, Ak = A − BK , Ak can be arbitrary (stabilized) if ( A, B) controllable )
( A − FC )T = AT − C T F T
 ( AT , C T ) controllable  AF T could be arbitrary(stabilized)
  C T AT C T ( AT ) n −1 C T  = N T controllable
 C 
 CA 
N =   ( A, C ) observable. qed #?
 
 n −1 
CA 
6
Closed-Loop Observer (Luenberger observer)
Since the system is observable, the rank of the observability matrix is n.
Noting that
M T =  C T AT C T ( AT ) n −1 C T  ,
" ( A, C ) is observable " is equivalent to "( AT , C T ) is controllable"
( This is known as the duality between controllability and observability ) .

Note) Can F stabilize AF ?

Ans) Yes, if ( A, C ) is observable.
(Recall, Ak = A − BK , Ak can be arbitrary (stabilized) if ( A, B) controllable )
( A − FC )T = AT − C T F T
 ( AT , C T ) controllable  AF T could be arbitrary(stabilized)
 C T AT C T ( AT ) 2 C T ( AT ) n −1 C T  = N T controllable
 C 
 CA 
N =   ( A, C ) observable. #
 
 n −1 
CA 
7
Closed-Loop Observer (Luenberger observer)

We can conclude that the eigenvalues of AT - C T F T can be assigned arbitrarily

by proper choice of F . Therefore, we have a complete control of the
convergence speed that the estimation error approach zero.

+ x 1 x y
u B s C
+

A nxn

+
F -
+
+ x̂ 1 x̂
B s C ŷ
+

A
nxn

 state observer(F = 0인 경우 open loop observer)

 시스템의 차수가 증가한다.
8
Closed-Loop Observer (Luenberger observer)
Bass − Gura algorithm for observer: single output system
F = [( N T w)T ]−1 ( − a )
 C 
 CA 
where N =  
 
 n −1 
CA 

9
Separation Theorem
v u

The linear state feedback control law was

u (t ) = − Kx(t ) + v(t )
If x is not directly measurable, a natural replacement the above equation is
u (t ) = − Kxˆ (t ) + v(t )
The plant and observer define the following 2n dimensional system
dx
plant → = Ax + B ( − Kxˆ + v )
dt
dxˆ
observer → = Axˆ + B ( − Kxˆ + v ) + FC ( x − xˆ )
dt
10
Separation Theorem
or
d  x  A − BK   x B
  =    + v
ˆ
dt  x   FC A − BK − FC   x   B 
ˆ

The poles of this system can be found from det ( sI 2 n − m0 ) = 0, However,

  A − BK   sI − A BK 
det  sI 2 n −   =
A − BK − FC    − FC sI − A + BK + FC 
det
  FC  

 add ( n + 1) , ( n + 2 ) , , 2n columns to 1, 2, , n columns, respectively

 sI − A + BK BK 
= det  
 sI − A + BK sI − A + BK + FC 
 Substract 1, 2, , n rows from ( n + 1) , ( n + 2 ) , , 2n rows, respectively
 sI − A + BK BK 
= det    square matrix
 0 sI − ( A − FC ) 
= det  sI − ( A − BK )   det  sI − ( A − FC ) 

11
Separation Theorem
We can find that n poles are due to the linear state feedback controller and
other n poles are due to the observer.
Thus one can conclude that poles assigned from controller and observer
design are completely separated.

 Separation principle or Eigenvalue separation

 Controller and observer can be designed separately without consideration
of their coupling

12
➢ 아래 이후는 2014년2학기 수업에서 제외되었음

13
Separation Theorem
Note

u ( s ) = − Kxˆ
= − K ( Is − A + BK + FC ) F y
−1

D(s)

Output feedback control law

u ( s ) = − Dy → 3차 compensator or PID controller 등도
= K ( Is − Ac ) Fy 모두 output observer 이다.
−1

where Ac = A − BK − FC
= Ak − FC
= AF − BK  K 와 F 를 결정하는게 어렵다.
dxˆ K → optimal control law
= Axˆ + B ( − Kxˆ + v ) + F ( y − Cxˆ )
dt F → Kalman Filter
sXˆ ( s ) = ( A − BK − FC ) Xˆ ( s ) + FY ( s ) + BV ( s ) 14
Disturbance Observer
Consider the system with disturbance
d1 d2
x = Ax + Bu + E d1
y = Cx + D d 2
E D
Let's define u u
x = Ax + Bu C
d 
E =  E 0   1 
 d2 
+
d  B + I y
D =  0 D   1  s
C
 d2  +

15
 Disturbance Observer
Then x = Ax + Bu + Ed
y = Cx + Dd
Let's assume that the disturbance, d, satisfies a known differential equation
Unknown exogenous input
d = A0 d + W  (t ) 
(Random with zeromean)
 only for mathematical convenience
 A0 는 많은 경우에 known(A0 = 0, W = I )
 x   A E   x  B 0
 d  =  0 A   d  +  0  u (t ) + W   (t )
   0      
 x
y = C D   
d 
xe = Ae xe + Beu + We (t )
 (*)
y = Ce xe

16
Disturbance Observer
Applying observer equation to (*) gives
F
xˆe = Ae xˆe + Beu + Fe ( y − Ce xˆe ), where Fe =   − (**)
ˆ  Fd  disturbance observation gain
Saperating (**) into xˆ and d yields
xˆ = Axˆ + Bu + Edˆ + F ( y − Cxˆ − Ddˆ )
Residual !
dˆ = A0 dˆ + Fd ( y − Cxˆ + Ddˆ )
The error equation becomes Cx + Dd

x − xˆ = Ax + By + Ed − Axˆ − Bu − Edˆ − Fcx − FDd + Fcxˆ + FDdˆ

= ( A0 − FdD)(d − dˆ ) Fd c( x − xˆ ) + W 
 x − xˆ   A − FC E − FD   x − xˆ   0 
 =    +    (t )
 d − dˆ   − Fd C ˆ
A0 − Fd D   d − d   W 
 
 ee = Aee + We (t )
If  (t ) = 0, then ee will converges to zero provided that A is stable. One should
confirm that (A e ,Ce ) pair in observable; (H.W. proof)
17
 Disturbance Observer
Note
ⅰ) In the care that A0 = 0i,e, d is constatnt, disturbance estimation,
dˆ , is a integral of residual as you see in (☆)
ⅱ) In equ (**), one should notice that there is not term for unknown input,  (t )
ⅲ) If E ( (t )) = 0 instead of  (t)=0, ee will converge to zero for state Ae
ⅳ) Suppose  (t ) is a series of finite impulsive function, i.e.

satisfyiry E ( (t )) = 0, Then d (t ) becomes stepwise function as shown below;

18
 Example (DC servo motor)
 (t )  0
 (t ) → 1
충분히 시스템의 b.w 에 비해 느리게
s
외란이 띄엄띄엄 들어오는 경우
d (t )
+
r + v ke 1  1
km
−  1s + 1 + J s 
−

B
kb

 For simplisity  1 = 0, ke = km = k
J  = km kev + d − Bv = km ke r − k m kekb + d − B
where v = r − kb  emf voltage
 input voltage
d=
 k k k +B 1  k m ke 
   − m e b      r   0
d = J  + J +  (t )
  1 
J
   d  
 0 0  0 
19
 Example (DC servo motor)
k m k e kb + B k k
Let = 10, m e = 300
J J
    −10 1     300   0 
 d  =  0 0  d  +  0  r +  1  
        
Design observer [xˆ = Axˆ + Bu + F ( y − Cxˆ )]
d  ˆ   −10 1   ˆ   300   f1 
 ˆ=   ˆ+  r +   ( − ˆ )
dt  d   0 0  d   0   f2 
F = [ f1 f 2 ] can be found from pole assignment
I.e.
  −10 1   f1  
det( Is − ( A − FC )) = det( Is −    −   ( ) )
1 0
  0 0   f2  
 s + 10 + f1 −1
= det   = s ( s + 10 + f1 ) + f 2
 0 s 
 f1 = −8, f 2 = 2
ˆ = −10ˆ + dˆ + 300r − 8 + 8ˆ
= −2ˆ + dˆ + 300r − 8
dˆ = 2ˆ − 2
20
 Disturbance observer with disturbance-free residual
Let xˆ (t ) = x(t ) + vdˆ (t ) ─①
where x is observer with d (t ) = 0, i.e.
x(t ) = Ax + Bu + F ( x) ─②
where v and F will be determined
Eq① with * and☆ yields
xˆ = x + vdˆ
= Ax + Bu + F ( y − Cx ) + V ( A0 dˆ + Fd ( y − c ( x + vdˆ ) − Ddˆ ))
= Ax + Bu + ( F + vF )( y − Cx ) + [vA − vF (Cv + D )]dˆ
x 0 d

But from ☆ with①and②

xˆ = A( x + vd ) + Bu + Edˆ + F [ y − Cx − (Cv + D )dˆ ]
= Ax + Bu + F ( y − Cx ) + [ Av + E − F (Cv + D )]dˆ
Thus we have two expersions for xˆ. In order for them both to hold for all seridual y − cx
F + vF = F or vF = F − F ─③
x d

vA0 − vFd (Cv + D ) = Av + E − F (Cv + D ) ─④

21
 Disturbance observer with disturbance-free residual
Equation ( 3) and ( 4 ) gives
vA0 − ( F − F )(Cv + D ) = Av + E − F (Cv + D )
 vA0 − ( A − FC )v = E − FD
 vA0 − Av = C ( 5)
where A = A − FC and C = E − FD

The existence of a solution to Equation ( 5 ) depends on the eigenvalue

of A0 and A. In the special case in which d is a constant, i.e. A0 = 0,
a solution to (5) exists if A is nonsingular. (One can make A = A − FC
be nonsigular and stable if ( A, C ) is observable.) In this case

v = −( A − FC ) −1 ( E − FD) ( 5)
22
Disturbance observer with disturbance-free residual
Finally, consider the equation for estimating the disturbance.
dˆ = A0 dˆ + Fd [ y − c( x + vdˆ ) − Ddˆ ]
= [ A − F (Cv + D)]dˆ + F ( y − cx)
0 d d

= Ad dˆ + Fd ( y − cx) (6)
where Ad = A0 − Fd (Cv + D)

23
 Disturbance observer with disturbance-free residual
Disturbance observer with disturbance-free residual ((1),(2),(6))

Fd d̂ 1
s d̂

Ad
y + V
− F
+ x x +
u B
1
x̂
+ s +

A
C

24
 Disturbance observer
Disturbance observer ( ) • 위 그림이 더 단순하다.
• 이 라인을 제거
D
+ d̂ 1
Fd d̂
+ s

A0
y + −
F E
−
u 1
B s
x̂

C 25
 Disturbance observer
Note that the residual, y - Cx, does not involve any (unknown) exogenous
disturbance input.

Design Procedure
Step 1 : Design and observer for the process without disturbance
(i.e. find F )
Luenberger observer
Step 2 : Find the matrix v using (5) or (5) when A0 = 0.
Step 3 : Find Fd so that dynamics of the estimator of the disturbance has
the desired pole in (6).

26
 Reduced-order observer
Suppose that there is one output for every state variable :
y = Cx

where C is a nonsingular matrix. In the case, there is no need for a

dynamic observer.; we can get x directly from y :
x = xˆ = C −1 y (19)

On intuition one might suspect that it should be possible to get by using

an observer of order n - r , where n is the order of the system and r is the
number of independent output. In many applications it is possible to
group the state variables into two sets : those that can be measured direc-
tly and those that depend indirectly on the domain.

27
 Reduced-order observer
The state vector is partitioned accordingly :
 x1 
x= 
 
 x2 
with x1 = A11 x1 + A12 x2 + B1u
(20)
x 2 = A21 x1 + A22 x2 + B2u

The output is given by

y = C1 x1 (21)

The statandard observer for (20) and (21) is

xˆ1 = A11 xˆ1 + A12 xˆ2 + B1u + K1 ( y − C1 xˆ1 ) (22)
xˆ2 = A21 xˆ1 + A22 xˆ2 + B2u + K 2 ( y − C1 xˆ1 ) (23) 28
 Reduced-order observer
But there's no reason to the trouble to implement the the observer
equation (22) for xˆ1 when we can solve for x1 directly from (21).
x1 = xˆ1 = C1−1 y (24)

In this case the observer (23) for those states that cannot be measured
directly becomes.
xˆ2 = A21C1−1 y + A22 xˆ2 + B2u (25)

which is a dynamic system of the same order so the number of state

variables that cannot be measured directly.

29
 Reduced-order observer
The dynamice behavior of the reduced-order observer is governed by
the eigenvalues of A22 which is a submatrix of the open-loop dynamic
matrix A. If the eigenvalues of A22 are suitable, then (25)

30
 Reduced-order observer
Could be a satisfactory observer. Since there is no assurance that the eigen
value of A22 are suitable, however, we need a more general system for the
reconstruction of x̂2 .

A suitably general structure for the estimation of x2 is given by

x̂2 = Ly + z (26)

Where z is state of a ( n − r ) th-order system.

z = Fz + Gy + Hu (27)

As we did for the full-order observer, we define the estimation error

 x1 − xˆ1   e1 
e = x − xˆ =  =

 
 
(28)
 x2 − xˆ2   e2 
By (24), e1 = x1 − xˆ1 = 0 (29)
So we are concerned only with e2 , the differential equation for which is
e2 = x2 − xˆ2 = A21 x1 + A22 x2 + B2u − Ly − z
y = C1 x1  y = C1 x1
= A21 x1 + A22 x2 + B2u = L C1 ( A11 x1 + A12 x2 + B1u ) 
− Fz − Gy − Hu (30)
 Reduced-order observer
But from (26)
z = xˆ2 − Ly = x2 − e2 − Ly = x2 − e2 − LC1 x1 (31)

So (30) becomes
e2 = Fe2 ( A21 − LC1 A11 − GC1 + FLC1 ) x1
+ ( A22 − LC1 A12 − F ) x2 + ( B2 − LC1 B1 − H )u (32)
In order for the error to be independent of x1 , x2 and u, the matrices
multiplying x1 , x2 and u must vanish ;
F = A22 − LC1 A12 (33)
H = B2 − LC1 B1 (34)
GC1 = A21 − LC1 A11 − GC1 + FLC1 (35)
Then, (32) becomes e2 = Fe2 and hence, for asymptotic stability,
the eigen-values of F must lie in the left half or the s-plane. Comparing (33)
with 4-(5) we see that selecting the matrix L in (33) to place the eigen-values
of F is the same type of problem as selecting the gain matrix K to place the
eiven-values of Â.
The submatrix A22 has the role of A in 4-(5) and product C1 A12 has the role of
C in 4-(5)
(*) in full order observer
 Reduced-order observer
In order to place the pole of F , it is necessary that the rank of the
corresponding controllability test matrix
W1 =  A12
'
C1' , A22
' '
A12 C1' , '
, ( A22 ) n − r −1 A12
'
C1' 
be of rank n − r. Luenberger has shown that this requirement in satisfied, if the
full-static is controllable. Having selected the matrix L to place the reduced–
order observer poles, the matrix H is determined from (34) and the matrix G is
determined from (35).
G = ( A21 − LC1 A11 ) C1 + FL
−1 (36)

Using (36), (27) given

z = F ( xˆ2 − Ly ) + Gy + Hu = Fxˆ2 + ( G − FL ) y + Hu
= Fxˆ2 + ( A21 − LC1 A11 ) C1−1 y + Hu (37)
G
This equation, together with (24) and (26) define the reduced-order observer.
A block-diagram representation of three equation is given in the figure.
Reduced-order observer
y x̂1
C1−1

G L

+ +
u z 1 z x̂2
H + s
+
F

F = A22 − LC1 A12

H = B2 − LC1 B1 Separation principle로 넘어갈 것!
G = ( A21 − LC1 A11 ) C1−1
 Reduced-order observer
The case that the static variables are not possible to group into two sets ;
those that can be measured directly and not we need more treatment.
We assume that C is of full rank of r , number of outputs. Then we can find M
so that the transformation matrix
C  r
T =   (38)
 M  n − r
Be nonsingular, denote the inverse by

T −1 =  P Q n
(39)
r n−r

Since TT −1 = T −1T = I, the following relation must hold between C , M , P and Q.

C 
  P Ir 0 
  Q = 
I n − r 
(40)
 M  0

and
C 
P Q    = PC + QM = I n (41)
 M 
 Reduced-order observer
We define the vector W by
W1 
W =   = Tx x = T −1W (42)
W2  or

Then the new state equation for W becomes

W = AW + Bu (43)
y = CW (44)

where A = TAT −1 , B = TB, C = CT −1 \ (45)

You should notice that
W1 
y = CW = CT −1W = C  P Q    =  I r 0 = W1 (46)
W2 
Rewritting (42) gives
W1 
x = T −1W =  P Q    = PW1 + QW2 (47)
W2 
or for estimation of x , can be expressed by
x̂ = PWˆ1 + QWˆ2 (48)
Reduced-order Observer

➢ By (46), ˆ , can be diretly estimated by measurement of output y, i.e.,

ˆ1 = y
And using the results of the previous analysis in section 4.2.5, we obtain the estimate
ˆ 2 = Ly + z (49)
where
z = Fˆ 2 + Gy + Hu
F = A22 − LA12
H = B2 − LB1 (50)
G = A21 − LA11

 A11 A12  C   CAP CAQ 

 = TAT =  M  A  P Q  =  MAP MAQ 
−1
A=
 A21 A22     
 B1  C   CB 
B =   = TB =   B =  
 
B 2
M
   MB 
C =  I r 0

37
Reduced-order Observer

➢ Thus, F = MAQ − LCAQ = ( M − LC ) AQ

H = MB − LCB = ( M − LC ) B (51)
G = MAQ − LCAP + F = ( M _ LC ) AP + F

( Note ) Separation principle applies to the case in which the state

feedback is based on a reduced order observer,i.e., the caracteristic
equation becomes
det ( I s − Ak ) det ( I s − F ) = 0

38
Eigenvalue Separation in State Feedback System with observer

➢ x = Ax + Bu
y = Cx
The linear state feedback control law was
u (t ) = − Kx(t ) + v(t ) (10)
If x is not directly measurable, a natural replacement for Eq.(10) is
u (t ) = − Kxˆ (t ) + v(t ) (11)
In this case, the plant and observer define the following 2n dimensional system
dx
plant → = Ax + B ( − Kxˆ + v )
dt
dxˆ
observer → = Axˆ + B ( − Kxˆ + v ) + FC ( x − xˆ )
dt
or
d  x  A − BK
  x B
= + (12)
dt  xˆ   FC
A − BK − FC   xˆ   B 
v

In poles of this system can be found from det ( sI 2 n − m0 ) = 0, However,

  A − BK   sI − A BK 
det  sI 2 n −   =
A − BK − FC    − FC sI − A + BK + FC 
det
  FC  
39
Eigenvalue Separation in State Feedback System with observer

 add ( n + 1) , ( n + 2 ) , , 2n columns to 1, 2, , n columns, respectively

 sI − A + BK BK 
= det  
 sI − A + BK sI − A + BK + FC 
 Substract 1, 2, , n rows from ( n + 1) , ( n + 2 ) , , 2n rows, respectively
 sI − A + BK BK 
= det    square matrix
 0 sI − ( A − FC )
= det  sI − ( A − BK )   det  sI − ( A − FC )  (14)

We can find that n poles are due to the linear state feedback controller and
other n poles are due to the observer. Thus one can conflude that poles assigned
from controller and observer design are completely separated.

 Separation principle or Eigenvalue separation

 Controller and observer can be designed separately without consideration
of their coupling

40
Eigenvalue Separation in State Feedback System with observer
Note

u ( s ) = − Kx
= − K ( Is − A + BK + FC ) F y
−1

D(s)

Output feedback control law → 3차 compensator or PID controller 등도

u ( s ) = − Dy 모두 output observer 이다.
= K ( Is − Ac ) Fy
−1

where Ac = A − BK − FC  K 와 F 를 결정하는게 어렵다.

= Ak − FC K → optimal control law
= AF − BK F → Kalman Filter
41
Robustness: Effect Of Modeling Error
Suppose that the observer is designed on the basis of a nominal plant.

xˆ = Axˆ + Bu + F ( y − Cxˆ ) ①

Note: original plant

x = Ax + Bu
where A = A +  A, B = B +  B, C = C +  C

a)Variation on A and C ( B = 0)
x = ( A +  A) x + Bu ②
② − ① : x − xˆ = e = Ax +  Ax + Bu − Axˆ − Bu − F ( y − Cxˆ )
= Ax +  Ax + Bu − Axˆ − Bu − F ((C +  C ) x − Cxˆ )
= ( A − FC )( x − xˆ ) + ( A − F  C ) x
→ e = ( A − FC )e + ( A − F  C ) x ③

Now for control, u = − Kxˆ = − K ( x − e)

42
Robustness: Effect Of Modeling Error
②  x = ( A +  A) x − BK ( x − e)
= ( A +  A − BK ) x + BKe
= ( A − BK ) x + BKe ④

 x  AK BK   x 
③,④    =   
e   AC ˆ
A   e 
where AK = A − BK = A − BK
ˆ = A − FC
A
 AC =  A − F  C

u x y
+ x = Ax + Bu C
-

K
x̂ xˆ = Axˆ + Bu + F ( y − Cxˆ )

43
Robustness: Effect Of Modeling Error
H1 ( s )

I x H1 ( s )
B + +  AC
+ s
AK  AC H 2 ( s)

K I
+ H2 H1
s +
+
Aˆ  AC
H 2 ( s)
H (s)
e→0이면 Robust +
e≠0이면 Robust가 아니다.  AC

44
Robust Stability (Absolute Stability)
0.1  K ( s )  0.2라고 가정하자.

-1/K

+ G
- -10 -5

1/K(s)

forbidden K  
zone


az If 1  absolute definition
G  
 

G  = max  G ( jw) 
wR

45
Robust Stability (Absolute Stability)
Example of stability

K: 기울기

1
K

+K

-K

46
Robust Stability (Absolute Stability)
 A = A − A  :  A의 최댓값, 보통 알 수 있다고 가정

For Robust stability

1
If  AC is bounded by  AC   , then H ( s)  


 Robust control이 이것을 만족하는 H ( s )를 찾거나 만족하는지 확인하는 과정이다.

e.q.
A=0
C = C +  C = (1 +  )C , − 0.1    0.1    = 0.1
Then  AC = − FC

47
Robust Stability (Absolute Stability)
D(s)

H (s) F H (s) C
- -

F  C 

B ( sI − AK ) −1 C
−
K ˆ ) −1
( sI − A F

D( s) = C ( sI − AK )−1 BK ( sI − Aˆ ) −1 F 
1
 

Pole placement
Open loop transfer fn Observer for Robustness
48
Robust Stability (Absolute Stability)
b)Variation on B ( A = A, C = C , B = B +  B )
Observer state feedback
 x   xˆ 
 xˆ  =  e 
   
 xˆ   AK FC   xˆ 
→ =   
  
e − BK AF 
e
where AK = A − BK , AF = A − FC

( sI − AK ) −1 x K
F

e
C ( sI − AF ) −1 B

49
Robust Stability (Absolute Stability)
If B = (1 +  ) B or  B =  B

F ( sI − AK ) −1 K
D(s)
+
-
 
C ( sI − AF ) −1 B

1
D( s) = K ( sI − AK ) FC ( sI − AF ) B 
−1 −1

 

※ 만일 Robustness가 만족이 안되는 경우

Compensator 등을 설계하여 추가하거나 K.F. 등을 변경시킨다.

50