State Observer

 A state observer estimates the state variables based on the measurements of the output and
control variables.
 The concept of observability is plays an important role in control systems.
 State observers can be designed if and only if the observability condition is satisfied.
 We shall use the notation 
x designate the observed state vector.
 The observed state vector 
x is used in the state feedback to generate the desired control vector.
 Consider the system defined by
x  Ax  Bu (1)
y  Cx (2)
 Assume that the state x is to be approximated by the state 
x the dynamic model
x  Ax
  Bu + K e ( y - Cx)
 (3)
which represents the state observer.
 Here, the state observer has y and u as inputs and 
x as output.
 The last term on the right-hand side of this model Eqn. (3), is a correction term that involves
the difference between the measured output y and the estimated output Cx .
 Matrix Ke serves as a weighting matrix.
 The correction term monitors the state 
x.
 State Observer cond. ……

 Figure 1 shows the block diagram of the system and the full-order state observer.

Full-order state observer

Fig. 1: Block diagram of system and full-order state observer, when input u
and output y are scalars.
 In the presence of discrepancies between the A and B matrices used in the model and those
of the actual system, the addition of the correction term will help to reduce the effects due
to the difference between the dynamic model and the actual system.
 The state observer for which the dynamics are characterized by A and B matrices and by the
additional correction term, which involves the difference between the measured output and
the estimated output.
 In the current discussions we assume that the A and B matrices used in the model and those
of the actual system are the same.
 Full-order State Observer
 Assume that the system is defined by Equs. (1) and (2) and the observer model is
defined by Equation (3).
 To obtain the observer error equation, let us subtract Eqn. (3) from Eqn. (1).
x - x  Ax  Ax
  K e (Cx  Cx
)
 ( A  K e C)(x  x) (4)
 Define the difference between x and 
x as the error vector e, or
e = x  x
 Then Eqn. (4) becomes
e = (A  K eC)e (5)
 The dynamic behavior of the error vector is determined by the eigenvalues of matrix
A – KeC.
 If matrix A – KeC is a stable matrix, the error vector will converge to zero for any initial
error vector e(0).
 That is 
x ( t ) will converge to x(t) regardless of the values of x(0) and x(0) .
 If the system is completely observable, then it can be proved that it is possible to
choose matrix Ke, such that A – KeC has arbitrarily desired eigenvalues.
 The observer gain matrix Ke, can be determined to yield the desired matrix A – KeC.
 Dual Problem
 The design of the full-order observer becomes that of determining an appropriate Ke,
such that A – KeC has desired eigenvalues.
 Consider the system defined by
x  Ax  B u
y  Cx
 In designing the full-order state observer, we may solve the dual problem, that is,
solve the pole-placement problem for the dual system
Z  A * z  C * v
n  B*z
 Assuming the control signal v to be
v  Kz
 If the dual system is completely state controllable, then the state feedback gain
matrix K can be determined such that matrix A* – C*K will yield a set of the desired
eigenvalues.
 If μ1, μ2, .., , μn the desired eigenvalues of the state observer matrix.
 Then by taking the same μi’s as the desired eigenvalues of the state-feedback gain
matrix of the dual system we obtain
| sI  ( A * C * K) |  ( s  1 )( s   2 )  ( s   n )
 Dual Problem cond. …….

 Noting that the eigenvalues of A* - C*K and those of A - K*C are the same, we have
| sI  ( A * C * K) |  | sI  ( A  K * C) |
 Comparing the characteristic polynomial | sI – (A – K*C) | and the characteristic
polynomial | sI – (A – KeC) | for the observer system in Eqn. (5), we find that Ke,
and K* are related by
Ke  K *
 Thus, using the matrix K determined by the pole-placement approach in the dual
system, the observer gain matrix Ke, for the original system can be determined by
using the relationship Ke = K*.
 Necessary and sufficient condition for state observation
 The necessary and sufficient condition for the determination of the observer gain
matrix Ke, for the desired eigenvalues of A – KeC is that the dual of the original
system
Z  A * z  C * v
be completely state controllable.

 The complete state observability condition for this dual system is that the rank of
[ C *  A * C *  ( A*) 2 C*    ( A*) n -1 C * ]
be n.
 This is the condition for complete observability of the original system defined by
Eqns. (1) and (2).
 This means that a necessary and sufficient condition for the observation of the
state of the system defined by Eqns. (1) and (2) is that the system be completely
observable.
 Design of Full-order State Observers
 Consider the system defined by
x  Ax  Bu (6)

y  Cx (7)

where x = state vector (n-vector)

u = control signal (scalar)
y = output signal (scalar)
A = n X n constant matrix
B = n X 1 constant matrix
C = 1 X n constant matrix
 We assume that the system is completely observable.
 We assume further that the system configuration is the same as that shown in Fig. 1.
 In designing the full-order state observer, it is convenient if we transform the system
equations given by Eqns. (6) and (7) into the observable canonical form.
 Design of Full-order State Observers Cond. ……
 Define a transformation matrix Q by
Q  ( WN * ) 1
where N is the observability matrix
N = [ C *  A * C *  ( A*) 2 C *    ( A*) n -1 C * ]
and W is defined by the following equation
 an 1 an  2  a1 1 
a an 3  1 0 
 n2
W=     
 
 a1 1  0 0
 1 0  0 0 
where al, a2, . . . , an-1 are coefficients in the characteristic equation of the original
state equation given by Eqn. (6):
| sI  A | = s n  a1s n 1    an 1 s  an  0
 Define a new state vector (n-vector) ξ by
X = Q (10)
 Design of Full-order State Observers Cond. ……
 Then Eqns. (6) and (7) become
  Q 1 AQ  Q 1Bu (11)
y = CQ (12)
0 0  0  an 
1 0  0  an 1 
 (13)
where Q AQ   0
1
1  0  an  2 
 
     
 0 0  1  a1 

 bn  an b0 
b  a b 
Q 1B   n 1 n 1 0 
   (14)
 
b  a b
 n 1 n 1 0 

CQ   0 0  0 1 (15)

 Equations (11) and (12) are in the observable canonical form.

 Design of Full-order State Observers Cond. ……

 If matrix A is already in the observable canonical form then Q = I.

 We choose the state observer dynamics to be given by
x = Ax + Bu  K e ( y  Cx)

 + Bu  K eCx
= (A  K eC)x (16)
 Now define
x = Q 
(17)
 By substituting Eqn. (17) into Eqn. (16), we have
 = Q 1 ( A  K C)Q  Q 1Bu  Q 1K CQ (18)
e e

 Subtracting Eqn. (18) from Eqn. (11), we obtain


   = Q1 ( A  K eC) Q (   (19)
 Define
    
 Then Equation (19) becomes
  Q1 ( A  K eC) Q  (20)
 Design of Full-order State Observers Cond. ……
 We require the error dynamics to be asymptotically stable and ε(t) to reach zero
with sufficient speed.
 The procedure for determining matrix Ke is first to select the desired observer
poles (the eigenvalues of A - KeC) and then to determine matrix Ke so that it will
give the desired observer poles.
 Noting that Q-1 = WN*, we have
 an 1 an 2  a1 1   C   k1 
a
 n2 an 3  1 0  CA   k2 
Q Ke   
1
       
   
 a1 1  0 0  CA n  2   kn 1 
 1  0 0   CA n 1   kn 
 k1 
k 
 2 
where K e    
 
 k n 1 
 kn 
 0 0  0  an 
1 0  0  an 1 
Design of Full-order State Observers Cond. …… 
Q 1 AQ   0 1  0  an  2 
 
     
 Since Q-1Ke is an n-vector, let us write  0 0  1  a1 
 n 
 
 n 1 
Q 1K e    
  (21)
 2 
 1 

 Then, referring to Eqn. (15), we have

 n  0 0  0 n 
  0 0  0  n1 
 n 1  
Q K eCQ =     0 0 0  1   
1
    
   

 2  0 0  0 2 
 1  0 0  0 1 
0 0  0 an   n 
1 0  0 an1   n1 

and Q ( A  K eC)Q  Q AQ  Q K eCQ = 0
1 1 1
1   an 2   n 2 
 
     
0 0  1 a1  1 
 Design of Full-order State Observers Cond. ……
 The characteristic equation
| sI  Q1 ( A  K eC)Q |  0

s 0 0  0 an   n
1 s 0  0 an 1   n 1
 0 1 s  0 an  2   n  2  0
     
0 0 0  1 s  a1  1

 s n  (a1  1 ) s n 1  ( a2   2 ) s n 2    ( an   n )  0 (22)
 It can be seen that each of δn, δn-1, δn-2, . . . , δ1 is associated with only one of the
coefficients of the characteristic equation.
 Suppose that the desired characteristic equation for the error dynamics is
( s  1 )( s   2 )  ( s   n )  s n  1 s n 1   2 s n 1     n 1 s   n  0 (23)
 Note that the desired eigenvalues µi's determine how fast the observed state
convergesto the actual state of the plant.
 Design of Full-order State Observers cond. ……
 Comparing the coefficients of terms of like powers of s in Equations (22) and (23),
we obtain
a1  1  1
a2   2   2
  
an   n   n
 From the above equations, we get
1  1  a1
 2   2  a2
  
 n   n  an
 Then, from Equation (21), we have
  n    n  an 
    a 
 n 1   n 1 n 1 
Q Ke =     
1
 
   
  2    2  a2 
 1   1  a1 
 Design of Full-order State Observers cond. ……
  n  an    n  an 
  a    a 
 n 1 n 1   n1 n 1 
 Hence, K e = Q     ( WN*)1   
    (24)

 2 2   a  
 2 2  a
 1  a1   1  a1 
 Equation (24) specifies the necessary state observer gain matrix Ke.
 The equation for the full-order state observer is
x = (A - K e C)x
 + Bu  K e y (25)
 Thus far we have assumed that the matrices A and B in the observer are exactly the
same as those of the actual plant.
 In practice, this may not be true.
 Then the error dynamics may not be given by Eqn. (20).
 This means that the error may not approach zero.
 Hence, we should try to build an accurate mathematical model for the observer to
make the error acceptably small.