3. SOLO
Reduced Order Observers for Linear Systems
∈∈∈+=
∈∈∈∈+=
pxmpxnpx
nxmnxnmxnx
RDRCRyuDxCy
RBRARuRxuBxAx
1
11
Plant:
We want to construct a Observer such that it’s output will asymptotically converge
to .x
xˆ
4. SOLO
Reduced Order Observers for Linear Systems
∈∈∈+=
∈∈∈∈+=
pxmpxnpx
nxmnxnmxnx
RDRCRyuDxCy
RBRARuRxuBxAx
1
11
Plant:
Assume: npCrank pxn ≤=
Find:
( )xnpn
RC −
⊥ ∈ such that:
nxn
C
C
⊥
is nonsingular.
Solution: Find the Singular Value Decomposition (SVD) of C
( )[ ] H
CpnpxCCpxn nxnpxppxp
VUC −Σ= 0
where H means Transpose of a matrix and complex conjugate of it’s elements, and:
nC
H
C
H
CCpC
H
C
H
CC IVVVVIUUUU ==== ;
( )
( ) ( )nxnnpxpp
ppC
diagIdiagI
diagpxp
1,,1,1,1,,1,1
0,,,, 2121
==
>≥≥≥=Σ σσσσσσ
( ) ( ) ( ) ( ) ( ) ( )
[ ] H
CCxppnCxnpn nxnpnxpnpnxpn
VUC −−⊥−−⊥
Σ= −−⊥ 0Then:
UC is any orthogonal matrix and ΣC is any non-zero diagonal matrix.
5. SOLO
Reduced Order Observers for Linear Systems
Define:
We have : x
C
C
p
uDy
=
−
⊥
( ) 1
: xpn
RpxCp −
⊥ ∈=
( )
−
=
−
= ⊥
−
⊥ p
uDy
CC
p
uDy
C
C
x ††
1
or : ( ) pCuDyCx
††
⊥+−=
where:
( ) nxpTT
RCCCC ∈=
−1† is the Right Pseudo-Inverse of C or pICC =†
( ) ( )pnnxTT
RCCCC −−
⊥⊥⊥⊥ ∈=
1† is the Right Pseudo-Inverse of C or pnICC −⊥⊥ =†
Then:
( ) ( )
( ) ( )
=
=
−−
−
⊥⊥⊥
⊥
⊥
⊥ pnxppn
pnpxp
I
I
CCCC
CCCC
CC
C
C
0
0
††
††
††
( ) nICCCC
C
C
CC =+=
⊥⊥
⊥
⊥
††††
6. SOLO
Reduced Order Observers for Linear Systems
We have:
+=
+=
uDxCy
uBxAx
( ) pCuDyCx
††
⊥+−=and:xCp ⊥=
( ) ( )[ ]{ }uBpCuDyCACuBxACxCp ++−=+== ⊥⊥⊥⊥
††
or:
( ) uBCuDyCACpCACp ⊥⊥⊥⊥ +−+= ††
We want to obtain an estimation of . If we add we can see that:ppˆ ( )uBxCyL −−
( )[ ]
( )
( )
0ˆ
ˆˆ
0
††
††
=−−−−=
−+−−=−−
−
⊥
⊥
uDpCCuDyCCy
uDpCuDyCCyuDxCy
pnpxpI
Apparently does not contain any information on , but let compute .py y
( ) ( )[ ]{ } uDuBpCuDyCACuDuBxACuDxCy +++−=++=+= ⊥
ˆ††
7. SOLO
Reduced Order Observers for Linear Systems
We have:
Therefore contains the information on .py
( ) ( )[ ]{ } uDuBpCuDyCACuDuBxACuDxCy +++−=++=+= ⊥
ˆ††
( ) uDuBCpCACuDyCACy +++−= ⊥
ˆ††
Let estimate by using:p
( )
( )[ ]uDuBCpCACuDyCACyL
uBCuDyCACpCACp
−−−−−+
+−+=
⊥
⊥⊥⊥⊥
ˆ
ˆˆ
††
††
or:
( )[ ] ( ) ( )[ ]uBuDyCApCACLCuDyLp
td
d
+−+−=−− ⊥⊥
†† ˆˆ
( ) pCuDyCx
††
⊥+−=
8. SOLO
Reduced Order Observers for Linear Systems
We have:
( )[ ] ( ) ( )[ ]uBuDyCApCACLCuDyLp
td
d
+−+−=−− ⊥⊥
†† ˆˆ
( ) pCuDyCx ˆˆ ††
⊥+−=
9. SOLO
Reduced Order Observers for Linear Systems
We also have:
( )[ ] ( ) [ ]uBxACLCuDyLp
td
d
+−=−− ⊥
ˆˆ
( ) pCuDyCx ˆˆ ††
⊥+−=
10. SOLO
Reduced Order Observers for Linear Systems
One other
form: ( )[ ] ( )
( ) ( ) ( ) uBCLCuDyCACLC
pCACLCuDyLp
td
d
−+−−+
−=−−
⊥⊥
⊥⊥
†
† ˆˆ
( ) pCuDyCx ˆˆ ††
⊥+−=
11. SOLO
Reduced Order Observers for Linear Systems
And
another
form:
( )[ ] ( ) ( )[ ]
( ) ( )( ) ( ) uBCLCuDyLCCACLC
uDyLpCACLCuDyLp
td
d
−+−+−+
−−−=−−
⊥⊥⊥
⊥⊥
††
† ˆˆ
( )[ ] ( )( )uDyLCCuDyLpCx −++−−= ⊥⊥
††† ˆˆ
12. SOLO
Reduced Order Observers for Linear Systems
We have:
( )
( )[ ]uDuBCpCACuDyCACyL
uBCuDyCACpCACp
−−−−−+
+−+=
⊥
⊥⊥⊥⊥
ˆ
ˆˆ
††
††
Subtract those equations:
Define the estimation error:
( )
( )[ ]uDuBCpCACuDyCACyL
uBCuDyCACpCACp
−−−−−+
+−+=
⊥
⊥⊥⊥⊥
††
††
( ) ( )ppCACLppCACpp ˆˆˆ ††
−−−=− ⊥⊥⊥
ppp ˆ:~ −=
( ) pCACLCp ~~ †
⊥⊥ −=
p~We can see that ( the estimation error) is uncontrollable and is stable iff.
( )[ ] iCACLCi ∀<− ⊥⊥ 0Real
†
λ ppp →→ ˆ&0~
13. SOLO
Reduced Order Observers for Linear Systems
Note:
Define:
( )
( )
( ) ( ) ( )
( )
( ) ( ) ( )
H
C
pnxpnCxppn
pnpxC
pnxpnCxppn
pnpxC
xnpn
pxn
nxn
pxppxp
V
U
U
C
C
Σ
Σ
=
−−−
−
−−−
−
−⊥ ⊥⊥
0
0
0
0
( )
( )
( ) ( ) ( ) ( ) ( ) ( )
−
=
−
=
−−⊥−⊥−−−
−
− pxnxppnxnpn
pxn
xnpn
pxn
pnxpnxppn
pnpx
xnpn
pxn
CLC
C
C
C
IL
I
T
C pxp
0
:
Since:
( )
[ ]
( ) ( )
=
− −−
⊥
⊥− pn
p
pn
p
pn
p
I
I
IL
I
CC
C
C
IL
I
0
000 ††
Define:
[ ] [ ]
( )
[ ]†††††
0
: ⊥⊥
−
⊥ +=
= CLCC
IL
I
CCMH
pn
p
14. SOLO
Reduced Order Observers for Linear Systems
Note (continue – 1):
Define: CLCT −= ⊥:
Then:
( )[ ] ( ) ( )[ ]
( ) ( )( ) ( ) uBCLCuDyLCCACLC
uDyLpCACLCuDyLp
td
d
−+−+−+
−−−=−−
⊥⊥⊥
⊥⊥
††
† ˆˆ
( )[ ] ( )( )uDyLCCuDyLpCx −++−−= ⊥⊥
††† ˆˆ
[ ] [ ]†††
: ⊥⊥+= CLCCMH
( )uDyLpz −−= ˆ:
( )
( )
−+=
+−+=
uDyHzMx
uBTuDyHATzMAT
td
zd
KF
ˆ
( )
( )
( )
( ) AT
C
T
HMAT
C
T
HATMAT
I
T
C
MH
I
I
MH
T
C
KF
n
pn
p
=
=
=
=
−
0
0
Those are the well known
Reduced Order Observer
Equations
15. SOLO
Reduced Order Observers for Linear Systems
Note (continue – 2):
Then:
( )
( )
−+=
+−+=
uDyHzMx
uBTuDyHATzMAT
td
zd
GF
ˆ
( )
( )
( )
( ) AT
C
T
HMAT
C
T
HATMAT
I
T
C
MH
I
I
MH
T
C
GF
n
pn
p
=
=
=
=
−
0
0
( ) CGCHATTMIAT
TMATATTFAT
DHSuSyHzMx
DGBTJuJyGzF
td
zd
n ==−=
−=−
−=++=
−=++=
:ˆ
:
=+
=+
−=
=−
−
0DHS
ITMCH
DGBTJ
CGTFAT
valueseigenstablehasF
n
nxpnxmnxq
qxpqxmqxq
nq
HSM
GJF
xz
xx
yHuSzMx
yGuJzFz
RRR
RRR
RR
∈∈∈
∈∈∈
∈∈
→
++=
++=
,,
,,
,ˆ,
ˆ
ˆ
16. SOLO
Observers
Generic Observer for a Linear Time Invariant (LTI) System
pxmpxnnxmnxn
pmn
DCBA
yux
uDxCy
uBxAx
RRRR
RRR
∈∈∈∈
∈∈∈
+=
+=
,,,
,,
Observer
nxpnxmnxq
qxpqxmqxq
nq
RSM
GJF
xz
xx
yRuSzMx
yGuJzFz
RRR
RRR
RR
∈∈∈
∈∈∈
∈∈
→
++=
++=
,,
,,
,ˆ,
ˆ
ˆ
A Necessary Condition for obtaining an Observer is that (A,C) is Observable.
The Observer will achieve
if and only if:
xx →ˆ
=+
=+
−=
=−
−
0DRS
ITGCR
DGBTJ
CGTFAT
valueseigenstablehasF
n
L.T.I. System
[ ] [ ]†††
: ⊥⊥+= CLCCMH CLCT −= ⊥:
HATGMATF == :&:
17. SOLO
Reduced Order Observers for Linear Systems
Let use a constant feedback from the
Reduced Order Observer to control the plant:
xK ˆ
The control is
xKrpCKxKru ˆ~†
−=+−= ⊥
( )[ ]
( ) ( )[ ]ppCpCuDyCKr
pCuDyCKrxKru
ˆ
ˆˆ
†††
††
−−+−−=
+−−=−=
⊥⊥
⊥
The augmented system is
( )
[ ]
[ ]
+
−=+=
−+=+−=
+
−
=
⊥
⊥⊥
⊥⊥
rD
p
x
CKKDCuDxCy
p
x
CKKrpCKxKru
u
B
p
x
CACLC
A
p
x
†
††
†
~
00
0
~
18. SOLO
Reduced Order Observers for Linear Systems
The augmented system is
( )
[ ]
[ ]
+
−=
+
−
+
−
=
⊥
⊥
⊥⊥
rD
p
x
CKKDCy
r
B
p
x
CKK
B
CACLC
A
p
x
†
†
†
000
0
~
or
The poles of the closed loop system are given by:
( )
[ ]
+
−=
+
−
−
=
⊥
⊥⊥
⊥
rD
p
x
CKKDCy
r
B
p
x
CACLC
CKBKBA
p
x
†
†
†
00~
( ) ( )
[ ] ( ) ( )[ ]
ObservertheofPoles
pn
ControllertheofPoles
n
pn
n
CACLCIsKBAIs
CACLCIs
CKBKBAIs †
†
†
detdet
0
det ⊥⊥−
⊥⊥−
⊥
−−⋅+−=
−−
−+−
Hence the Reduced Order Controller has the “Separation Property” of the Controller and
Observer.
19. SOLO
Reduced Order Observers for Linear Systems
Compensator Transfer Function
By tacking the Laplace Transform of the compensator dynamics we obtain:
( )[ ] ( ) [ ]uBxACLCuDyLp
td
d
+−=−− ⊥
ˆˆ
( ) pCuDyCx ˆˆ ††
⊥+−=
xKu ˆ−=
( ) ppCCuDyCCxC
I
ˆˆˆ †
0
†
=+−= ⊥⊥⊥⊥
( )[ ] ( ) ( ) xKBACLCuDyLxCs ˆˆ −−=−− ⊥⊥
( ) ( ) ( )[ ]( ) xKBACLCKDLCsyLs xnpn
ˆ
−⊥⊥ −−−−=
( ) ( ) ( )[ ] ( ) ( ) 1
†ˆ mxxmpnpnnx yLsKBACLCKDLCsx −−⊥⊥ −−−−=
where
Therefore
( ) ( ) ( )[ ]( ) ( ) ( ) ( )[ ] ( ) ( )pnpnnxxnpn
IKBACLCKDLCsKBACLCKDLCs −−⊥⊥−⊥⊥ =−−−−−−−−
†
( ) ( ) ( )[ ] ( ) ( ) 1
†
mxxmpnpnnx yLsKBACLCKDLCsKu −−⊥⊥ −−−−−=
20. SOLO
Reduced Order Observers for Linear Systems
How to Find K & L For a Stable Closed Loop
This can be done by solving the following
Asymptotic Minimum Variance Control Problem:
xKu ˆ−=
( ) ( ){ } ( ) ( ) ( ){ }
+=
−===++=
uDxCy
tWwtwEtwExwuBxAx T
τδτ 0,0,00
( ) ( ) ( ) ( ){ } )(0lim definitepositiveRtuRtutxQtxEJ TT
t
>+=
∞→
System with no output noise to allow us to use a Reduced Order Observer.
The solution to this problem is:
where: PBRK T1−
=
and P is the solution of the Algebraic Riccati Equation:
01
=−++ −
PBRBPQAPPA TT or:
HT
T
ARic
AQ
BRBA
RicP =
−−
−
=
−1
Minimize:
A stabilizing solution (and unique) exists iff:
1 (A,B) is stabilizable
2 AH has no jω axis eigenvalues
If Q ≥ 0 then P ≥ 0
21. SOLO
Reduced Order Observers for Linear Systems
How to Find K & L For a Stable Closed Loop (continue – 1)
( ) wCuBCuDyCACpCACp
inputknown
⊥⊥⊥⊥⊥ ++−+=
1
††
( ) pCuDyCx
††
⊥+−=and:xCp ⊥=
and
( )wuBxACxCp ++== ⊥⊥
( ) ( )[ ]{ }
( ) wCuDuBCuDyCACpCAC
uDwuBpCuDyCACuDwuBxACuDxCy
inputknown
+++−+=
++++−=+++=+=
⊥
⊥
2
††
†† ˆ
The measurements are given by (instead of )y y
Let define:
†*†*
****
:,:
:,:,:,:
⊥⊥⊥
⊥
==
====
CACCCACA
wCvwCwyypx
22. SOLO
Reduced Order Observers for Linear Systems
How to Find K & L For a Stable Closed Loop (continue – 2)
( ) wCuBCuDyCACpCACp
inputknown
⊥⊥⊥⊥⊥ ++−+=
1
††
( ) wCuDuBCuDyCACpCACy
inputknown
+++−+= ⊥
2
††
Define:
†*†*
****
:,:
:,:,:,:
⊥⊥⊥
⊥
==
====
CACCCACA
wCvwCwyypx
The Estimation Problem becomes:
( ) ∗∗∗∗
++= winputknownxAx 1
( ) ∗∗∗∗
++= vinputknownxCy 2
[ ]
=
=
⊥
⊥⊥⊥
**
**
00
00**
*
*
:
RS
SP
CWCCWC
CWCCWC
vw
v
w
E TTT
TT
TT
23. SOLO
Reduced Order Observers for Linear Systems
How to Find K & L For a Stable Closed Loop (continue – 3)
The Estimation Problem:
( ) ∗∗∗∗
++= winputknownxAx 1
( ) ∗∗∗∗
++= vinputknownxCy 2
[ ]
=
=
⊥
⊥⊥⊥
**
**
00
00**
*
*
:
RS
SP
CWCCWC
CWCCWC
vw
v
w
E TTT
TT
TT
The Solution to the Estimation Problem is:
( )( ) ( )[ ]( ) 1
0
†
0
1 −
⊥⊥
−∗∗∗
+=+= TTTT
CWCCACYCWCRCYSL
or
( )[ ] ( ) 1
0
†
0
−
⊥⊥ += TTT
CWCCCAYWCL
where
( )[ ] ( )
( )( ) ( )[ ]
−−−−
−−
=
∗−∗∗∗∗−∗∗∗
∗−∗∗∗−∗∗∗
CRSASRSP
CRCCRSA
RicY
T
TT
11
11
24. SOLO
Reduced Order Observers for Linear Systems
How to Find K & L For a Stable Closed Loop (continue – 4)
In explicit form (the Algebraic Riccati Equation) is:
But
( )[ ] ( )[ ] ( ) ( )( ) 0
1111
=−+−−+− ∗−∗∗∗∗−∗∗∗−∗∗∗∗−∗∗∗ TTT
SRSPYCRCYCRSAYYCRSA
( ) ( )
( ) ( )[ ] ( )pnnx
TT
nxnpn
TT
CACCWCCWIC
CACCWCCWCCACCRSA
−⊥
−
−⊥
⊥
−
⊥⊥⊥
∗−∗∗∗
−=
−=−
†1
00
†1
00
†1
( ) ( ) ( )
( ) ( ) †1
0
†
†1
0
†1
⊥
−
⊥
⊥
−
⊥
∗−∗∗
=
=
CACCWCCAC
CACCWCCACCRC
TTTT
TT
( ) ( )( ) ( )
( )[ ] TTT
n
TTTTT
CWCCWCCWIC
CWCCWCCWCCWCSRSP
⊥
−
⊥
−
⊥⊥⊥
∗−∗∗∗
−=
−=−
0
1
00
0
1
000
1
25. SOLO
Reduced Order Observers for Linear Systems
How to Find K & L For a Stable Closed Loop (continue – 5)
Therefore Y(n-p)x(n-p) is given by the following (n-p) Algebraic Riccati Equation:
( )[ ] ( )[ ]{ }T
TT
n
TT
n CACCWCCWICYYCACCWCCWIC
†1
00
†1
00 ⊥
−
⊥⊥
−
⊥ −+−
( ) ( ) YCACCWCCACY TTTT †1
0
†
⊥
−
⊥− ( )[ ] 00
1
00 =−+ ⊥
−
⊥
TTT
n CWCCWCCWIC
Note:
1 ( )[ ] ( ) ( ) ( )
( ) PCCWCCWI
CCWCCCWCWCCWCCWCCWICCWCCWIP
TT
n
T
I
TTTTT
n
TT
n
=−=
+−=−=
−
−−−−
1
00
1
00
1
00
1
00
21
00
2
2:
This is a Projection, since P2
= P, but oblique because P is
not symmetrical.
2 For W0 = In we get: ( ) ( ) ⊥⊥
−−
=−=−=− CCCCICCCCICCWCCWI n
TT
n
TT
n
††11
00
( ) ( ) ( ) ††††1
0
†
⊥⊥⊥
−
⊥ = CACCACCACCWCCAC TTTTTT
( )[ ] ††††1
00 ⊥⊥⊥⊥⊥⊥⊥
−
⊥ ==−
−
CACCACCCCACCWCCWIC
pnI
TT
n
( )[ ] TT
I
TTT
n CCCCCCCWCCWCCWIC ⊥⊥⊥⊥⊥⊥⊥
−
⊥ ==−
†
0
1
00
26. SOLO
Reduced Order Observers for Linear Systems
How to Find K & L For a Stable Closed Loop (continue – 6)
Hence, for W0=In, Y(n-p)x(n-p) is given by the following (n-p) Algebraic Riccati Equation:
Notice (continue – 1):
3 With this L , A*- L C* will have stable eigenvalues, but
2
( ) ( ) ( ) 0
†††††
†
=+−+ ⊥⊥⊥
−
⊥⊥⊥⊥⊥
⊥⊥
T
CCI
TTT
CCYCACCACYCACYYCAC
n
( )[ ] ( ) ( ) ††
0
1†
†
†
CACYCCCCAYCL TTCC
C
TTT
⊥
=
−
⊥⊥
⊥⊥
=+=
( ) †††
** ⊥⊥⊥⊥⊥ −=−=− CACLCCACLCACCLA
Therefore has stable eigenvalues, and the
Reduced Order Estimator is stable
( ) †
⊥⊥ − CACLC
27. SOLO
Reduced Order Observers for Linear Systems
How to Find K & L For a Stable Closed Loop (continue – 7)
Notice (continue – 2):
4 Following P.J. Blanvillain and T.L. Johnson
(IEEE Tr. AC., Vol. AC-23, No.1, June 1978) this
Problem is equivalent to the following
( ) ( )
=
+=
xCy
WNxuBxAx 0,0~0
Given
Find the Dynamic Compensator Parameters (F, G, H, M)
+=
+=
yMzHu
yGzFz
Compensator
Which minimizes the Quadratic Performance Index:
( ) ( ) ( ) ( ) ( )[ ]
+= ∫
∞
0
,,, dttuRtutxQtxEMHGFJ TT
28. SOLO
Reduced Order Observers for Linear Systems
Let append to the Reduced Order Observer the Stable Transfer Matrix
( ) ( )
=+−=
−
DC
BA
DBAIsCsQ
ˆˆ
ˆˆ
:ˆˆˆˆ 1
=−−− uBCpCACyCACy ˆ††
The input to the Stable Transfer Function will be the same
as for the Reduced Order Observer.
29. References
SOLO
Kwakernaak, H., Sivan, R., “Linear Optimal Control Systems”, Wiley Inter-science,
1972, pg.335
Reduced Order Observers for Linear Systems
Gelb A. Ed, “Applied Optimal Estimation”, The Analytic Science Corporation, 1974,
pg.320
30. August 13, 2015 30
SOLO
Technion
Israeli Institute of Technology
1964 – 1968 BSc EE
1968 – 1971 MSc EE
Israeli Air Force
1970 – 1974
RAFAEL
Israeli Armament Development Authority
1974 –2013
Stanford University
1983 – 1986 PhD AA