Lecture 13 - Observers Introduction

K. J. Åström • The white box view of dynamical systems

• State a number of variables that summarizes the past that
1. Introduction is useful for prediction
2. The Observer Problem • The pole placement can be solved if all states are mea-
3. Combining with State Feedback sured
4. Integral action • What to do if only outputs are known?
5. Response to Reference Signals • How to determine the state from the output?
6. Summary • Combine it with state feedback

Theme: Estimating the state and using the estimate for • A new view on integral action
feedback.

Computing the State from the Output An Other Attempt

  The problem can be solved if the observability matrix has full

   y  rank. Can we avoid differentiations?
System 
 


C 
 
 dy  


 
 
 


 CA 
 
 dt   dx
dx 
 
 
  = Ax + Bu
= Ax 


.. 


x = Wo x = 

 ..  
 dt
dt 
 . 
 
 .  
 y = Cx
y = Cx   
 n− 1  
CA n− 1 
 d y 

dt n − 1
Differentiate the output Simulate a model of the system

y = Cx d x̂
Wo observability matrix = A x̂ + Bu
dt
dy  
= CAx y 
dt 
  Introduce error x̃ = x − x̂

 dy  
.. 
 

. 
 

−1  
 dt  d x̃
x = Wo  
 = A( x − x̂) = A x̃
d n− 1 y 
 .
. 
 dt
= CAn−1 x 
 . 
dtn−1  n− 1 
 


 d y 

n − 1
Works if A stable, but it makes no use of u.
dt

c K. J. Åström August, 2000 1
 Two Ways to Calculate the State An Alternative Solution
By differentiating the output dx
= Ax + Bu
dt
 T
dy d n− 1 y y = Cx
x = Wo−1 y ...
dt dtn−1
Drive the model by both u and y
This method has the drawback that it only uses the output and
d x̂
that it is differentiated many times which is highly sensitive to = A x̂ + Bu + K ( y − C x̂ )
dt
noise.
Introduce error x̃ = x − x̂
By integrating the input d x̃
= A( x − x̂ ) − K ( y − C x̂) = ( A − K C ) x̃
dt
d x̂
= A x̂ + Bu Notice that use of output y gives extra freedom.
dt

This method has the drawback that it only uses the input and Determine matrix K so that the matrix A − K C has all its eigenval-
that it only works if the matrix A is stable. ues in the left half plane, then x̃ will go to zero!

Can we combine the methods?

Example - The Car Example - The Car ...

  The matrix
dx 0 1 0 
= Ax + Bu = x+ u − k1 1
dt 0 0 1 A− KC =
y = Cx = ( 1 0)x − k2 0

Observability matrix has the characteristic polynomial


1 0 
Wo = s + k1 −1
0 1 det ( sI − A + K C) = det = s2 + k1 s + k2
k2 s
has full rank. The system is observable and the observer is
Choosing
  
d x̂ 0 1 0 k1
= A x̂ + Bu = x̂ + u+ ( y − x̂1 ) k1 = 2ζ ω o
dt 0 0 1 k2
k2 = ω 2o

− k1 1
We have A − K C = gives the characteristic polynomial s 2 + 2ζ ω o s + ω 2o .
− k2 0

c K. J. Åström August, 2000 2
 Interpretation of the Observer Comparison with State Feedback
The observer is a dynamical system with two inputs u and y State feedback design: Find matrix L so that the matrix A − B L
and two outputs x 1 and x2 . has prescribed eigenvalues.
  
d x̂ 0 1 0 k1 Observer design: Find Matrix K so that the matrix A − K C has
= x̂ + u+ ( y − x̂1 )
dt 0 0 1 k2 prescribed eigenvalues.

Input-output relations A matrix and its transpose have the same eigenvalues. We
have
1 2ζ ω o s + ω 2o
X̂ 1 ( s) = U ( s) + Y ( s) ( A − K C ) T = AT − C T K T
s2 + 2ζ ω o s + ω 2o s2 + 2ζ ω o s + ω 2o
s + 2ζ ω o ω 2o s
X̂ 2 ( s) = U ( s) + 2 Y ( s)
s2 + 2ζ ω o s + ω o
2 s + 2ζ ω o s + ω 2o State feedback and observer design are the same problem.
The same software can be used. Use state feedback program
to obtain the observer gain by feeding it by A T and C T instead
dy of A and B . Transposing L give the observer gain.
Physical interpretation. Compare with x̂2 =
dt

Observable Canonical Form Observable Canonical Form

       
 −a1 1 0 ... 0  b1   −a1 1 0 ... 0  b1 

 
 
 
 
 
 
 

   b2     b2 
 −a2
 0 1 0 
 
  −a2
 0 1 0 
 

dz  

 ..









.. 

 dz  

 ..









.. 


= 
 z + 
 u =  z +  u
dt 

. 
  . 
 
 dt 


. 

  . 

 


       

 − a 0 0 1
  

  
 − a 0 0 1  
 
     

n− 1 n− 1
     
−an 0 0 0 bn −an 0 0 0 bn
   
y = 1 0 0... 0z y = 1 0 0... 0z

      
 1 0 0 ... 0

 1 0 ... 0

 −a1 1 0 ... 0  b1   k1 

 
 
 
 
 
 
 
 
 

 −a1 ...     0  b2   k2 

 1 0 0 
 − a1 1 ... 0  −a2
 0 1 
 
 
 
 



 a − a2 −a1 1 ... 0

 

 0

 d ẑ 

 ..





 .





 .



−a2 −a1 ... =     
2
W̃o = 

 1

,
 W̃o−1 =




 
 . 
 ẑ + 
 .
. 
 u+
 .
. 
 ( y − ẑ1)

 
 
 
 dt 
 
 
 
 
 



.. 
 

.. 
 
 
 
 
 
 


 . 
 
 . 
 
 − a n− 1 0 0 1 
 bn−1 
 
 kn−1 

         
. . . ... 1 −an − a n− 1 ... 1 −an 0 0 0 bn kn


c K. J. Åström August, 2000 3
 Summary Output Feedback
If the system Consider a system which is observable and controllable
dx dx
= Ax = Ax
dt dt
y = Cx y = Cx

is observable its state can be determined by the observer Determine the state with an observer and use state feedback
from the observed state. The controller is
d x̂
= A x̂ + Bu + K ( y − Cx)
dt u = − L x̂
d x̂
• Combine measurements with mathematical model - = A x̂ + Bu + K ( y − C x̂ )
dt
indirect measurements -redundancy
• Sensor fusion in AI terminology The controller is a dynamical system whose dynamics is
represented by the observer
• Determination of filter gain K similar to state feedback
design. Relation to the Kalman filter

The Closed Loop System Block Diagram of Process and Observer

Process
dx
= Ax + Bu u x y
dt B Σ ∫ C
y = Cx
u = − L x̂ A

d x̂
= A x̂ + Bu + K ( y − C x̂)
dt Observer contains a copy of the process model
y
Introduce the state x̃ = x − x̂ instead of x̂ . Σ −1

dx
= Ax + Bu = Ax − B Lx̂ = Ax − B L( x − x̃) = ( A − B L) x + B Lx̃ K
dt
d x̃
= ( A − K C) x̃
dt u
Σ ∫
x yˆ
B C
Characteristic equation
A
det ( sI − A + B L) det ( sI − A + K C) = 0

c K. J. Åström August, 2000 4
 Block Diagram of Closed Loop system Observers and State Feedback - A Summary
Design of a controller for
xˆ u y dx
Observer −L Process = Ax + Bu, y = Cx
dt

can be split into two problems

• Design of a state feedback and an observer
A few things remains • The closed loop system has poles corresponding to the
• Integral action eigenvalues of the state feedback A c = A − B L and the
observer A o = A − K C matrices
• How to introduce reference signals
• The problems are similar. The observer is obtained by the
transformation A → A T , B → C T and L → K T . The same
computer program can be used.
• Interesting interpretation of controller

Integral Action by Explicit Disturbance Modeling Modeling Disturbances

There are many ways to introduce integral action. Here we will Classical disturbance models:
give a method based on explicit modeling of disturbances. dv
• Step = 0, v(t) = a = constant
• What is the meaning of the model dt
d2 v
dx • Ramp 2 = 0, v(t) = a + bt
= Ax + Bu dt
dt
d2 v
• Sinusoid 2 + ω 2 v = 0, v(t) = a sin(ω t + b)
• Why do we introduce integral action? dt
• Disturbances!
• How to describe disturbances


c K. J. Åström August, 2000 5
 Constant Disturbance at Process Input Example - The Car
Constant disturbance x 3 = v, physical interpretation.
dx
= Ax + B (u + v)    
dt 0 1 00
dx
dv
=0 = 0 0 1x +1u
dt dt
0 0 0 0
Feedback from estimated states y = Cx = ( 1 0 0 ) x

u = − L x̂ − v̂ System is observable (check this!). The observer

     
Observer 0 1 0 0 k1
d x̂ 
     = 0 
0 1 x̂ + 1 u + k2  ( y − x̂1 )
  
d x̂ A B x̂ B K dt
= + u+ ε 0 0 0 0 k3
dt v̂ 0 0 v̂ 0 Kv
ε = ( y − C x̂)

Example - The Car ... How to Introduce Reference Signals

The matrix To introduce the reference values we use the standard configu-
  ration with two degrees of freedom.
− k1 1 0
A − K C =  − k2 0 1 ur Model and u ff
− k3 0 0 Feedforward
xm u fb y
Generator
∑ L ∑ Process
has the characteristic polynomial −
 
s + k1 −1 0 xˆ
det ( sI − A + K C) =  k2 s −1  = s3 + k1 s2 + k2 s + k3 Observer

k3 0 s
• A nice structure
choose k1 = a + 2ζ ω o, k2 = 2aζ ω o + ω 2o , and k3 = aω 2o to give
characteristic polynomial • Decoupling of estimation of disturbances K
• Reduction of disturbances L
( s + a)( s2 + 2ζ ω o s + ω 2o ) = s3 + ( a + 2ζ ω o ) s2 + (2aζ ω o + ω 2o ) s + aω 2o
• Response to reference signals

c K. J. Åström August, 2000 6
 Example - The Car Example - The Car ...

    The complete controller

0 1 0 0
dx 
= 0 0 1 x + 1u
  u = l1 ( xm1 − x̂1 ) + l2 ( xm2 − x̂2 ) − x̂3 + u m
dt  
0 0 0 0 2ζ ω m
u m = ω 2m r − x m2 − x m1
y = Cx = ( 1 0 0 ) x ω 2m
 
dxm 0 1 0
= x m + r
Desired response in Laplace transforms dt −ω m −2ζ ω m
2
ω 2m
     
0 1 0 0 k1
ω 2m d x̂
Ym = X m1 = M y R = R, X m2 = sX m1 =  0 0 1  x̂ +  1  u +  k2  ( y − x̂1 )
s2 + 2ζ ω m s + ω 2m dt
0 0 0 0 k3
Hence
ω 2m s2
U m = P− 1 M y R = R
s2 + 2ζ ω m s + ω 2m
 2ζ ω m s + ω 2m   2ζ 
= ω 2m 1 − 2 R = ω 2m R − sYm − Ym
s + 2ζ ω m s + ω m 2 ωm

A Simulation Summary
2 y och ym • An elegant approach to control system design
1 • Interesting system structure with three components:
0 – state feedback
0 5 10 15
– observer
20 uff och ufb
– reference trajectory generator
10
• Strong similarity between state feedback and observers
0
• Sensor fusion and diagnosis y − ŷ
0 5 10 15
v and ve • A new interpretation of integral action: disturbance estima-
1
tor. Can be adapted to other signals.
0.5

0
0 5 10 15


c K. J. Åström August, 2000 7