Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
5, N
o
. 2
,
A
p
r
il
201
5, p
p
.
25
9
~
27
0
I
S
SN
: 208
8-8
7
0
8
2
59
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Recursive Least-Squares Estimati
on for the Joint Input-State
Estimation of Linear Discrete
Time Systems with Unknown
Input
Ta
lel. Bessao
u
di
1
, Fa
yç
al
. B
e
n Hmi
d
a
2
Unité d
e
r
ech
erche en
Commande, Surveillance
et Sûreté
d
e
fon
c
tionnement d
e
s Sy
stèm
es (C3S)
Tunis University
-
ESSTT, 5
av
.
Taha Hussein BP 5
6
-1008, Tun
i
s,
Tunisia
1
bessaouditalel@
y
a
hoo
.fr,
2
fa
ycal
.benhm
id
a@e
sstt.rnu.tn
Article Info
A
B
STRAC
T
Article histo
r
y:
Received Sep 9, 2014
Rev
i
sed
Jan
3, 2
015
Accepte
d
Ja
n 15, 2015
This
paper
pr
es
ents
a r
ecur
s
ive le
as
t-s
quar
e
s
approa
ch t
o
es
tim
ate
simultaneously
the state and
the unknow
n in
put of lin
ear
time var
y
ing
discrete time s
y
stems with unk
nown
input. Th
e method is based on the
assumption that no prior knowledge abou
t th
e
d
y
namical
evolu
tion of th
e
input is availab
l
e. Th
e join
t in
put
and state
estimation are o
b
tain
ed b
y
recursive least-s
quares formulation b
y
apply
i
ng
the inv
e
rsion
lemmas. The
proposed filter i
s
equivalen
t
to recursive thr
ee
step filt
er. To il
lustrat
e
th
e
performance of
the proposed
fi
lter an
example is
given.
Keyword:
In
fo
rm
ation fo
rm
ulas
In
ver
s
i
o
n l
e
m
m
a
Least-squa
res
State estim
a
tion
U
nkn
own
input esti
m
a
t
i
o
n
Copyright ©
201
5 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Talel Bessaou
di,
Un
ité
d
e
rech
erch
e en Co
mman
d
e
, Surv
eillan
ce et Sûreté de fo
n
c
tionn
emen
t d
e
s Systèmes (C
3
S
)
Tu
ni
s U
n
i
v
e
r
si
t
y
-ESSTT,
5
a
v
.
Taha
H
u
ssei
n
B
P
5
6
-
1
00
8,
Tu
ni
s, T
u
ni
si
a
Em
a
il: b
e
ssao
u
d
italel@yah
o
o
.
fr
1.
INTRODUCTION
During the
last deca
des, t
h
e
pr
oblem
of unknown input
filtering ha
s recei
ved growing at
tention
due
to
its ap
p
licatio
n
s
i
n
env
i
ro
nmen
tal state es
ti
m
a
tio
n
[1
], [2]. Th
e un
kno
wn
inp
u
t
filtering
prob
lem
h
a
s treated
in
th
e literature b
y
d
i
fferen
t
ap
pro
ach
es. Th
e
first ap
pro
a
ch
assu
m
e
s th
at th
e m
o
d
e
l fo
r d
y
n
a
m
i
cal ev
o
l
u
t
i
on
of t
h
e
u
n
k
n
o
w
n
i
n
put
i
s
avai
l
a
bl
e.
Whe
n
t
h
e
pr
ope
rt
i
e
s of t
h
e u
n
k
n
o
w
n i
n
put
are
k
n
o
w
n,
t
h
e augm
ent
e
d
st
at
e
Kalm
an
filter (
A
SKF) is a solu
tio
n
.
To
redu
ce co
m
p
u
t
atio
n
co
sts of th
e ASKF, Fried
l
an
d
[2
] propo
sed
th
e
two
stag
e Kal
m
an
filter wh
ere th
e estim
at
i
o
n
of th
e state an
d unk
nown in
pu
t are
d
e
co
up
led. Th
e seco
nd
approach treats the case
wh
en
no
t h
a
v
e
a
pr
io
r
kn
ow
ledge ab
ou
t th
e
d
y
n
a
m
i
cal ev
o
l
u
tio
n
f
o
r
th
e
unk
now
n
in
pu
t.
Kitan
i
d
i
s [1
] was th
e fi
rst to so
l
v
e t
h
e prob
le
m
using the linea
r
unbi
ased
m
i
nim
u
m
-
va
ri
ance.
Dar
oua
c
h
et al, [3
] ex
ten
d
Kitan
i
d
i
s’s filter u
s
in
g
a param
a
terizin
g
t
ech
n
i
q
u
e
to
o
b
tain
an
o
p
tim
a
l
filter (OEF).
Hsieh
[4
]
h
a
s
d
e
v
e
lop
e
d an
eq
u
i
v
a
len
t
to Kitan
i
d
i
s’s
filter
n
o
t
ed
b
y
rob
u
s
t
-
two
stag
e
Kalm
an
filter (RTSKF). Later,
Hsieh [5
] d
e
v
e
lo
p
e
d
an
o
p
t
i
m
al
m
i
n
i
m
u
m
v
a
rian
ce
filter (OM
V
F) to
solv
e th
e p
e
rfo
r
man
ce of
d
e
grad
ation
p
r
ob
lem
en
co
un
tered
in
(OEF). Gillij
n
s
& De
Moo
r
[6
]
has treated
th
e
p
r
ob
lem
to
esti
m
a
te th
e state
in
th
e
p
r
esen
ce
of
u
n
k
nown inp
u
t
wh
ich
affects only th
e syste
m
s m
o
d
e
l. Th
ey
dev
e
lop
e
d a
recu
rsi
v
e
filter wh
ich
is
o
p
tim
al in
th
e sen
s
e
o
f
m
i
n
i
m
u
m
-
v
a
rian
ce. Th
is
filter h
a
s b
e
en
ex
tend
ed
b
y
th
e sam
e
au
thors [7
] for jo
i
n
t
in
pu
t and
state esti
m
a
t
i
o
n
to
lin
ear d
i
screte-ti
m
e s
y
st
e
m
s with
d
i
rect
feed
throu
gh
whe
r
e the state and the
u
nkn
own
inp
u
t esti
m
a
tio
n
are in
terconn
ect
ed
. Th
is filter
is called
recursiv
e three step filter (RTSF)
an
d
i
s
li
mited
to
d
i
rect feed
throug
h
m
a
trix
h
a
s fu
l
l
ran
k
. Ch
e
ng
et
al
, [8]
pr
op
ose
d
a recu
rsi
v
e op
tim
a
l
filt
er with
g
l
ob
al op
ti
m
a
l
ity in
th
e sen
s
e o
f
u
n
b
i
ased
min
i
m
u
m
-
v
a
rian
ce ov
er all un
b
i
ased
estim
a
t
o
r
s,
bu
t th
is filter is
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 2, A
p
ri
l
20
15
:
25
9 – 2
7
0
26
0
li
mited to estimate the state. Recently
, the
case of an arbi
trar
y
rank has been
solved
by
Hsieh (2009)
in
the
designe
d
opti
m
al filter [9],
known as ER
TSF
(E
xten
d RTSF).
Ot
her m
e
thods
are proposed
by Gi
llijns
a
n
d
B
a
rt
de M
o
or i
n
[1
0]
,
[1
1]
an
d
[1
2]
w
h
i
c
h
u
s
e l
east
-s
qua
re
s (LS
)
t
e
c
hni
q
u
e a
n
d
t
h
e i
n
f
o
rm
ati
on f
o
rm
ulas.
In t
h
i
s
pa
per
,
we pre
s
ent
an
un
bi
ase
d
m
i
ni
m
u
m
-
vari
ance est
i
m
a
ti
on of t
h
e st
at
e and t
h
e un
kn
o
w
n
in
pu
t. Th
ese esti
m
a
tes
are obt
ai
ned
by
sol
v
i
n
g t
h
e i
n
f
o
r
m
at
i
on form
ul
as usi
n
g t
h
e w
e
i
ght
ed l
east
-
s
qua
res
m
e
t
hod.
T
h
e a
dva
nt
age
o
f
t
h
i
s
m
e
t
hod i
s
t
o
pr
ovi
de a
di
rect
est
i
m
a
t
e
of t
h
e
st
at
e an
d
u
n
k
n
o
w
n
i
n
p
u
t
i
n
a
si
ngl
e
bl
oc
k
wi
t
h
a si
m
p
l
e
cal
cul
a
t
i
on.
The
pape
r i
s
o
r
ga
ni
zed as
f
o
l
l
o
w. Sect
i
o
n
2,
prese
n
t
s
t
h
e
pr
obl
em
un
de
r co
nsi
d
e
r
at
i
o
n
and s
o
m
e
p
r
elim
in
aries. In
section
3
,
we
set up
t
h
e d
e
sign
of
th
e filter eq
u
a
ti
o
n
b
y
recursiv
ely so
l
v
ing
t
h
e
weig
h
t
ed
least-squ
a
res
prob
lem
.
An
ill
u
s
trativ
e ex
amp
l
e is p
r
esen
te
d
in
section
4
.
Fin
a
lly, in
sect
i
on
5 we c
oncl
ude
o
u
r
obt
ai
ne
d res
u
l
t
s
.
2.
PROBLEM AND PRELE
MINARIES
2.
1 Pr
obl
em
F
o
rmul
a
ti
o
n
C
onsi
d
er
t
h
e l
i
near
st
oc
hast
i
c
di
scret
e
-t
im
e sy
ste
m
with
unk
nown inp
u
t
i
n
th
e
fo
llowing
form
:
1
kk
k
k
k
k
x
Ax
G
d
w
+
=+
+
(1
)
kk
k
k
k
k
yC
x
H
d
v
=+
+
(2
)
whe
r
e
n
k
x
ÎÂ
is the state vector,
m
k
d
ÎÂ
is th
e u
nkn
own
in
pu
t v
ector
an
d
p
k
y
ÎÂ
is the m
eas
urem
ent
vector. The
process noise
n
k
w
ÎÂ
and the m
e
asurem
ent nois
e
p
k
v
ÎÂ
are assum
e
d to be m
u
tuall
y
u
n
c
or
r
e
lated
zer
o
s-
m
ean
wh
ite r
a
nd
o
m
sig
n
a
ls with
non sin
g
u
l
ar
cov
a
r
i
an
ce m
a
tr
ice
s
0
T
kk
k
Qw
w
éù
=³
êú
ëû
and
0
T
kk
k
RE
v
v
éù
=>
êú
ëû
respectively. The m
a
trices
,,
kk
k
A
GC
and
k
H
are known and
have
appropri
ate
dim
e
nsion.
We
assum
e
that
()
,
kk
A
C
is o
b
se
r
v
a
b
l
e
,
pm
and t
h
e initial state is unc
orre
lated with the
white
noises pr
ocess
e
s
k
w
and
.
k
v
The
initial state
0
x
is a Gaussia
n
ra
nd
om
var
i
able with
00
ˆ
Ex
x
,
00
00
0
ˆˆ
T
Ex
x
x
x
P
where
.
E
d
e
no
tes th
e exp
ectatio
n
operato
r.
Also, we assu
m
e
th
at
11
kk
k
rank
C
G
r
a
nk
G
,
th
e d
i
rect feed
throug
h
m
a
tri
x
k
H
has
an arbitrary
rank
.
Th
e ob
j
ecti
v
e o
f
th
is p
a
p
e
r is to
d
e
sig
n
an
op
ti
m
a
l recu
rsive filter wich
esti
m
a
tes b
o
t
h
the syste
m
state
k
x
and
th
e unk
now
n in
pu
t
k
d
b
a
sed
on th
e in
itial estimate
0
ˆ
x
a
n
d the s
e
que
nce
of m
e
asurem
ent
01
,
,
...,
k
y
yy
. N
o
pri
o
r k
n
o
wl
e
d
ge ab
out
t
h
e d
y
n
am
i
cal
evol
u
t
i
on o
f
k
d
is assumes to be available. Now
we deri
ve a Recursive
Least Square
(RLS)
proce
d
ure that propa
g
a
t
es a one
st
ep
ahead predicte
d state estim
a
te. For sim
p
licity of
deri
vations
,
we use a
stoc
ha
stic approac
h
.We ass
u
m
e
that an estim
ate
/1
ˆ
kk
x
is available
with covaria
n
ce
matrix
/1
/1
/1
ˆˆ
T
kk
k
k
k
k
kk
PE
x
x
x
x
an
d
we seek
fo
r a weig
h
t
ed
least sq
u
a
re
(WLS) th
at allo
ws to
esti
m
a
te
/
ˆ
kk
x
base
d on
/1
ˆ
kk
x
and the
ne
wly available
measurem
ent
k
y
.
The e
r
ror estimation
/1
kk
x
-
%
i
s
gi
ve
n
by
:
/1
/1
ˆ
:.
kk
k
k
k
xx
x
--
=-
%
(3
)
Usi
n
g (1
), (
2
)
and
(
3
)
,
we obt
ai
n
t
h
e f
o
l
l
o
wi
ng
eq
uat
i
o
n:
/1
/
1
1
ˆ
00
0.
0
kk
n
k
kk
kk
k
k
k
kk
n
k
k
xI
x
x
yC
H
d
v
AG
I
x
w
(4
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Recu
rsive
Lea
s
t-Sq
ua
res Estima
tion
f
o
r t
h
e
Jo
in
t
Inpu
t-S
t
ate Estima
tion
of Lin
e
ar
Discrete …
(T. Bessao
ud
i)
26
1
So,
t
h
e c
o
rres
p
on
di
n
g
WL
S
p
r
o
b
l
e
m
i
s
gi
ve
n
by
1
2
/1
,,
1
ˆ
00
mi
n
0
0
kk
k
kk
k
kk
k
k
xd
x
kk
n
k
xx
I
yC
H
d
AG
I
x
k
W
(5
)
whe
r
e
k
W
de
not
es t
h
e
wei
ght
i
n
g
m
a
t
r
i
x
.
From
(5
) t
h
e i
n
t
e
r
p
ret
a
t
i
o
n o
f
an M
V
U (
U
nbi
ase
d
M
i
ni
m
u
m
-
Vari
ance
) est
i
m
at
or i
s
obt
ai
ne
d
by
c
h
o
o
si
ng
11
1
/1
,,
kk
k
k
di
a
g
P
R
Q
k
W
. Th
e
p
r
o
p
o
s
ed so
l
u
tio
n of th
e LS
p
r
ob
lem
(5
) is
g
i
v
e
n
in the fo
llo
wi
n
g
fo
rm:
//
1
ˆ
ˆ
kk
k
k
k
k
k
dy
C
x
M
(6
)
//
1
/
1
/
ˆ
ˆˆ
ˆ
()
k
k
kk
k
k
k
k
k
k
kk
xx
K
y
C
x
H
d
(7
)
1/
/
/
ˆ
ˆˆ
kk
k
k
k
k
k
k
x
Ax
G
d
(8
)
Whe
r
e t
h
e
gain m
a
trices
mp
k
M
and
np
k
K
still h
a
v
e
to b
e
d
e
term
in
ed
later.
2
.
2
Preliminaries
Th
e
fo
llowing
le
mmas are essen
tial for later
d
e
v
e
l
o
p
m
en
ts.
Lemma
A.1
(The ma
trix
in
ve
rsi
on l
e
m
m
a
[1
1]
):
Let
nn
A
,
nm
B
,
mn
C
and
mm
D
b
e
real
m
a
trice
s
. If
A
,
1
DC
A
B
and
D
are
no
n-
si
ng
ul
ar,
t
h
e
n
1
AB
D
C
is no
n-
singu
lar, and
11
11
1
1
1
.
A
B
D
C
A
AB
D
C
AB
C
A
The
follo
win
g
fo
rm
ula pr
o
v
ides a m
a
nne
r
to in
vert a
22
b
l
o
c
k m
a
trix
b
a
sed
o
n
th
e m
a
trix
inv
e
rsion
le
mma,
1
1
1
1
11
1
0
.
0
AB
D
C
AB
I
B
D
CD
CA
I
DC
A
B
Ind
e
ed
, th
e d
i
ag
on
al en
tries of th
e
first m
a
tr
ix
on
th
e righ
t
h
a
n
d
sid
e
o
f
t
h
e eq
u
a
lity sign
can
b
e
co
m
p
u
t
ed
u
s
ing
t
h
e m
a
tri
x
inv
e
rsion
lemma.
Lemm
a A
.
2:
Let
nn
A
,
nm
B
and
mm
C
be
re
al
m
a
trices. If
A
,
C
ar
e non
-
s
i
n
gular
th
en
,
11
11
1
.
TT
A
BC
B
A
B
A
B
C
B
B
C
3.
FILTER DE
SIGN
Th
e calcu
lation
of th
e
o
p
tim
a
l
m
a
trices
k
M
and
k
K
is ad
d
r
essed
in
th
e su
b
s
ection
3.1
wh
ich
cal
l th
e
measurem
ent update,
yields an estim
a
t
e of
k
x
an
d u
nkno
wn
i
n
pu
t
k
d
. T
h
e t
i
m
e updat
e
o
f
t
h
e st
at
e
est
i
m
a
ti
on i
s
p
r
esent
e
d i
n
s
u
b
s
ect
i
on
3.
2.
3.
1.
Mea
s
ure
ment Up
da
te
The m
easurement update is derive
d from
(5) by
ex
tracting
th
e ro
ws th
at
dep
e
nd
on
ly on
k
x
and
k
d
.
This yield,
2
/1
,
ˆ
0
min
kk
kk
k
kk
kk
xd
xx
I
CH
yd
1,
k
W
(9
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
IJEC
E V
o
l
.
5, No
. 2, A
p
ri
l
20
15
:
25
9 – 2
7
0
26
2
Whe
r
e
11
/1
,
kk
k
dia
g
P
R
1,
k
W
d
e
no
tes th
e
weigh
ting m
a
trix
.
No
w we
de
ri
ve an e
xpl
i
c
i
t
up
dat
e
f
o
rm
ul
a by
sol
v
i
n
g t
h
e p
r
obl
em
st
at
e and u
n
k
n
o
w
n i
n
p
u
t
esti
m
a
t
i
o
n
s
. Fi
rstly, no
te th
at
(9) is eq
ui
val
e
nt
t
o
t
h
e l
east
-
s
qua
res
p
r
o
b
l
e
m
mi
n
2,k
k
2
kk
k
W
X
Y-
A
X
(1
0)
W
h
er
e
k
A
0
kk
I
CH
,
/1
ˆ
kk
k
x
y
k
Y
,
k
k
x
d
k
X
and
11
/1
(,
)
kk
k
di
ag
R
P
2,
k
W
(1
1)
Using t
h
e
Gauss-Markov t
h
eorem
[13], the
solution is
written as:
ˆ
-1
TT
kk
2
,
k
k
k
2
,
k
k
X=
A
W
A
A
W
Y
(1
2)
Usin
g (1
1)
the
cova
riance
m
a
trix
-1
T
k2
,
k
k
AW
A
follows as
-1
T
k2
,
k
k
AW
A
1
11
1
/1
11
TT
kk
k
k
k
k
k
k
TT
kk
k
k
k
k
PC
R
C
C
R
H
HR
C
H
R
H
(1
3)
In
t
h
e next section we will
determin
ate an
unbiased estimate of t
h
e st
ate and
unknown input
by seeking a
solution t
o
the
equation
(13).
Lemma 3.1:
T
h
e
exp
r
e
s
s
i
o
n
of
th
e e
r
r
o
r
cov
a
r
i
an
c
e
m
a
tr
i
x
/
d
kk
P
is gi
ven
by
:
1
1
/
,
dT
kk
k
k
k
PH
R
H
(1
4)
and the e
r
ror c
ova
riance
m
a
tr
ix of t
h
e
state i
s
gi
ven in t
h
e followi
ng form
:
//
1
/
dT
T
kk
kk
k
k
k
k
k
k
k
PP
K
R
H
P
H
K
(1
5)
whe
r
e
/1
T
kk
k
k
kk
RC
P
C
R
(1
6)
1
/1
T
kk
k
k
k
K
PC
R
(1
7)
Proof:
Note t
h
at,
/
kk
P
and
/
d
kk
P
can
be ide
n
tifi
e
d as error c
o
variance m
a
trices of
/
ˆ
kk
x
and
/
ˆ
kk
d
, that
is,
//
/
T
k
k
kk
kk
PE
x
x
,
//
/
dT
k
k
kk
kk
PE
d
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN:
208
8-8
7
0
8
Recursive Least-Squares
Estimation
f
o
r
t
h
e Joint Input-St
ate
Estim
atio
n o
f
Linea
r
Discre
te
…
(T.
Bessao
ud
i)
26
3
Whe
r
e t
h
e inve
rse
of
/
kk
P
and
/
d
kk
P
are give
n, respecti
v
ely, by
11
1
1
1
1
//
1
1
()
,
TT
kk
kk
k
k
k
k
k
k
k
k
k
T
kk
k
PP
C
R
C
C
R
H
H
R
H
HR
C
(1
8)
1
11
/
1
11
1
/1
.
dT
T
kk
k
k
k
k
k
k
TT
kk
k
k
k
k
k
k
PH
R
H
H
R
C
PC
R
C
C
R
H
(1
9)
The
n
,
by
a
pply
i
ng lem
m
a
A.1
the equati
on (13) is
rewritten as follows:
1
11
1
/1
/
11
/
1
11
1
11
1
/1
0
0
TT
kk
k
k
K
k
k
k
k
k
TT
d
kk
k
k
k
k
k
TT
kk
k
k
k
k
TT
kk
k
k
k
k
k
K
PC
R
C
C
R
H
P
HR
C
H
R
H
P
IC
R
H
H
R
H
HR
C
P
C
R
C
I
(2
0)
Ap
ply
i
ng the
m
a
trix inversi
on lem
m
a
A.1
to the info
rm
ation fo
rm
ulas (18
)
an
d (
1
9
)
, the err
o
r co
v
a
riance
matrix
/
d
kk
P
and
/
kk
P
ar
e give
n i
n
t
h
e f
o
llowi
ng
f
o
rm
s:
1
1
11
1
1
/1
/
1
TT
T
kk
k
k
k
k
k
k
k
k
k
d
kk
T
kk
k
H
R
H
H
RC
P
C
RC
P
CR
H
(2
1)
11
11
1
1
1
11
1
1
1
/1
1
11
TT
T
kk
k
k
k
k
kk
k
TT
T
T
k
k
kk
k
k
k
k
kk
k
k
k
k
TT
kk
k
k
k
k
HR
H
H
R
H
HR
C
PC
R
C
C
R
H
H
R
H
H
R
C
CR
C
H
R
H
(2
2)
1
/1
TT
kk
k
k
k
k
k
HR
C
P
C
H
(2
3)
1
1
T
kk
k
HR
H
(2
4)
//
1
/
dT
T
kk
kk
k
k
k
k
k
k
k
PP
K
R
H
P
H
K
(2
5)
1
11
/1
/
11
1
1
()
T
kk
k
k
k
kk
TT
kk
k
k
k
k
k
k
k
PC
R
C
P
CR
H
H
R
H
H
R
C
(2
6)
1
1
11
1
1
T
kk
k
k
k
TT
T
T
kk
k
k
k
k
k
k
k
k
k
k
k
k
PP
C
R
H
HR
H
H
R
C
P
C
R
H
HR
C
P
(2
7)
11
/
Td
T
kk
k
k
k
k
k
k
k
k
k
PP
C
R
H
P
H
R
C
P
(2
8)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN:
2
088
-87
08
IJEC
E V
o
l. 5, No
. 2, A
p
ril 20
15
:
25
9 – 2
7
0
26
4
/1
/
Td
T
T
kk
k
k
k
k
k
k
k
k
k
PK
R
K
K
H
P
H
K
(2
9)
w
h
er
e
1
11
/1
T
kk
k
k
k
k
PP
C
R
C
(3
0)
11
/1
TT
kk
k
k
k
k
k
PC
R
P
C
R
(3
1)
/1
T
kk
k
k
k
CP
R
K
(3
2)
The gain
m
a
tri
x
k
K
that m
i
nim
i
z
e
the e
r
ror covariance is
gi
ve
n
by
1
/1
.
T
kk
k
k
k
K
PC
R
(3
3)
Setting the Derivate of
(25) with respect
t
o
k
K
, we
get
/
/
Td
T
T
kk
kk
k
k
k
k
k
k
P
RK
H
P
H
K
K
(3
4)
Let re
place the
/
d
kk
P
by
e
q
uation
(
1
4)
,
we
notice t
h
at
/
0
kk
k
P
K
,
therefore the gain
k
K
m
i
nim
i
ze the trace
of
the m
a
trix covariance
/
kk
P
.
Lemma 3.2:
An unb
iased esti
m
a
te o
f
th
e
un
kno
wn
in
pu
t
k
d
can be obtai
ne
d
in
the f
o
llowi
ng
f
o
rm
:
//
1
ˆ
ˆ
kk
k
k
k
k
k
dy
C
x
M
(3
5)
whe
r
e
1
/
.
dT
kk
k
k
k
PH
R
M
(3
6)
We c
o
nsider the m
i
nim
u
m
-
variance
unbiase
d
state estim
ation
/
ˆ
kk
x
give
n in
th
e f
o
llowi
ng
f
o
r
m
:
//
1
1
/
ˆ
ˆˆ
ˆ
()
kk
kk
k
k
k
k
k
k
k
xx
K
y
C
x
H
d
(3
7)
Proof:
The e
q
uation
(
1
2
)
ca
n
be
writ
ten
as follows
1
11
1
/
/1
11
/
11
/1
/1
1
ˆ
ˆ
ˆ
.
0
TT
kk
kk
k
k
K
k
k
k
TT
kk
kk
k
k
k
T
kk
k
k
kk
T
k
kk
x
PC
R
C
C
R
H
d
HR
C
H
R
H
PC
R
x
y
HR
(3
8)
Substituting
(20)
in (38), we obtain
1
/
/
/1
/1
ˆˆ
k
k
kk
kk
k
k
xP
P
x
(3
9)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN:
208
8-8
7
0
8
Recursive Least-Squares
Estimation
f
o
r
t
h
e Joint Input-St
ate
Estim
atio
n o
f
Linea
r
Discre
te
…
(T.
Bessao
ud
i)
26
5
11
1
1
1
//
()
,
TT
T
T
k
k
kk
k
k
kk
k
k
k
k
kk
k
PC
R
P
C
R
H
H
R
H
H
R
y
and
estim
ation o
f
un
k
n
o
w
n i
n
p
u
t is
give
n i
n
th
e
follo
win
g
fo
rm
:
11
1
1
1
//
/
/
1
11
/1/1
ˆ
()
()
.
dT
d
T
T
kk
k
k
k
k
k
k
k
k
k
k
kk
k
k
k
T
kk
k
k
k
k
d
P
HR
y
P
HR
C
P
C
R
C
Px
C
R
y
(4
0)
Let us apply the le
mma
A.1
a
nd
A.
3
to t
h
e e
quatio
n
(3
9) a
n
d (
4
0
)
, t
h
e estim
ate of the u
n
k
n
o
w
n i
n
p
u
t is
give
n
by
:
11
//
/
11
1
1
1
/1
/1
/1
ˆ
ˆ
()
(
)
.
dT
dT
kk
kk
k
k
k
k
k
k
k
k
TT
kk
k
k
k
k
kk
k
k
k
d
P
HR
y
P
HR
C
PC
R
C
P
x
C
R
y
(4
1)
w
h
er
e
11
1
1
1
//
1
/
1
/
1
11
/1
ˆ
()
ˆ
.
dT
T
kk
k
k
k
k
k
k
k
k
kk
k
k
TT
kk
k
k
k
k
k
k
P
H
RC
P
C
RC
P
x
HR
H
H
R
C
x
(4
2)
1
11
1
1
1
//
1
1
//
1
.
dT
T
T
kk
k
k
k
k
k
k
k
k
k
k
k
k
dT
T
kk
k
k
k
k
k
k
k
PH
R
R
C
P
C
R
C
C
R
y
PH
R
C
P
C
y
(4
3)
11
//
1
ˆ
ˆ
TT
kk
k
k
k
k
k
k
k
k
k
dH
R
H
H
R
y
C
x
(4
4)
Rem
a
rk 3.
1:
to evaluate t
h
e
perform
a
nce
of the
filter in case where
H
k
ha
s an arbitrary
rank we
use the
heu
r
istic exte
n
s
ion
p
r
ese
n
ted
in [
9
]
by
re
placing e
q
uatio
n (
1
4)
an
d
(3
6
)
by
:
11
//
,†
dT
d
T
kk
k
k
k
k
k
k
k
k
PH
R
P
H
R
H
M
(4
5)
whe
r
e
†
the M
o
o
r-Pe
n
rose
is
ge
neralized
in
ver
s
e
1
†
TT
M
MM
M
The state estim
ation
/
ˆ
kk
x
is gi
ven
in the
followi
ng
form
:
1
/
/
/1
/1
11
1
1
1
//
ˆˆ
()
kk
kk
k
k
k
k
TT
T
T
kk
k
k
kk
k
k
k
k
k
k
k
k
k
xP
P
x
PC
R
P
C
R
H
H
R
H
H
R
y
(4
6)
Using t
h
e inve
rsion lem
m
a
A.
1
and
A.
3
we can
show
that:
11
//
1
/
1
/
1
/
1
/
1
11
1
1
/1
/1
ˆˆ
ˆ
ˆ
()
.
T
k
k
kk
kk
kk
k
k
k
k
k
k
k
TT
T
kk
k
k
k
k
k
k
k
k
k
k
k
PP
x
x
P
C
R
C
x
PC
R
H
H
R
H
H
R
C
x
(4
7)
11
1
1
1
/
11
1
1
/1
/1
1
()
()
.
TT
T
kk
k
k
k
k
k
k
k
k
k
k
TT
T
k
k
kk
k
k
k
k
k
k
kk
k
T
kk
k
PC
R
R
H
H
R
H
H
R
y
PC
R
y
PC
R
H
H
R
H
HR
y
(4
8)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN:
2
088
-87
08
IJEC
E V
o
l. 5, No
. 2, A
p
ril 20
15
:
25
9 – 2
7
0
26
6
1
//
1
/
1
/
1
11
1
1
/1
/
1
11
1
1
/1
/
1
1
ˆˆ
ˆ
ˆ
()
()
.
T
kk
kk
k
k
k
k
k
k
k
TT
T
k
k
kk
k
k
k
k
kk
k
k
k
TT
T
kk
k
k
k
k
k
k
k
k
k
k
k
T
kk
k
xx
P
C
R
C
x
PC
R
H
H
R
H
H
R
C
x
PC
R
y
PC
R
H
H
R
H
HR
y
(4
9)
11
//
1
/
1
/
1
ˆˆ
TT
kk
k
k
k
k
k
k
k
k
k
k
k
k
x
IP
C
R
C
x
P
C
R
y
(5
0)
11
1
1
/1
/
1
ˆ
()
TT
T
kk
k
k
k
k
k
k
k
k
k
k
k
k
PC
R
H
H
R
H
H
R
y
C
x
3.
2. T
i
me
Up
d
a
te
Firstly
, we e
x
tr
act fr
om
(4)
t
h
e equation that
depe
nds
on
1
k
x
.
1
kk
k
k
k
k
x
Ax
G
d
w
(5
1)
Second substituting
k
x
and
k
d
for their L
S
estim
ates
/
ˆ
kk
x
and
/
ˆ
kk
d
obtained
d
u
ri
ng
the m
easurem
ent
up
date (4
1)
an
d (5
0)
.
T
h
e
n
, we obtain
//
1
/
/
ˆ
ˆ
()
k
k
k
k
kk
k
k
kk
kk
k
A
xG
d
x
A
x
G
d
w
+
+=
-
+
+
%
%
(5
2)
The c
o
r
r
es
po
n
d
in
g L
S
pr
oble
m
is give
n
by
1/
ˆ
ˆ
mi
n
kk
k
k
k
k
xA
x
G
d
3,
k
W
(5
3)
Whe
r
e
3,
k
W
de
notes
the weighting m
a
trix
which we
c
h
oose
1
//
//
T
k
k
k
k
kk
k
k
kk
k
k
k
k
EA
x
G
d
w
A
x
G
d
w
3,
k
W
(5
4)
From
eq
uation
(5
3)
1/
/
/
ˆ
ˆˆ
kk
k
k
k
k
k
k
x
Ax
G
d
(5
5)
The e
r
ror estim
ation
1/
kk
x
is give
n by
1/
1
1
/
ˆ
kk
k
k
k
xx
x
(5
6)
//
kk
k
k
k
k
k
A
xG
d
w
(5
7)
In
co
nse
que
nc
e, the c
o
varia
n
ce m
a
trix o
f
1/
ˆ
kk
x
is give
n
by
:
1/
1
/
1/
T
kk
kk
kk
PE
x
x
(5
8)
//
//
T
xd
k
kk
k
k
kk
k
dx
d
T
kk
k
k
k
A
PP
A
GQ
PP
G
(5
9)
I
t
fo
llo
ws fro
m (3
5)
th
at
/
kk
d
is
given
by
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN:
208
8-8
7
0
8
Recursive Least-Squares
Estimation
f
o
r
t
h
e Joint Input-St
ate
Estimation of
Linear Discre
te
…
(T.
Bessao
ud
i)
26
7
//
1
kk
k
k
k
k
k
k
k
k
dI
M
H
d
M
C
x
v
(6
0)
//
/
/
T
xd
x
d
k
k
kk
kk
kk
PP
E
x
d
.
Using
(56) a
n
d (60), it
follows that
//
x
dd
kk
k
k
k
k
PK
H
P
(6
1)
4.
ILLUSTRAT
I
VE E
X
AMPLE
To show the
propose
d
re
sults, th
e n
u
m
e
rical
exam
ple given
by
Da
ro
uac
h
,
Zasadzin
ki a
n
d
B
outay
e
b
(2003) is consi
d
ere
d
, whe
r
e the pa
ram
e
ters
of system
s
(1) and (
2
) are giv
e
n as follo
ws, t
h
e pa
ram
e
ters
of the
sy
stem
(1) a
n
d
(2
) a
r
e
give
n
b
y
:
0
000
5
0
0
084
0
0
5
1
7
0
806
9
k
..
A
..
10
01
k
,C
,
0
0
129
0
1
250
4
0
k
.
G
.
0
0
036
0
034
2
0
0
342
0
324
9
k
..
,Q
..
00
1
0
00
1
6
k
.
R.
.
W
i
t
h
out loss
of
gene
rality, the initial stat
e and its es
tim
ate are both assum
e
d to be zero, and the initial
cova
riance
is
g
i
ven by
()
0
10
,
2
0
0
Pd
i
a
g
=
. The
u
n
k
n
o
w
n
i
n
p
u
t
are
give
n
by
[]
[
]
[
]
[
]
[]
[]
55
2
0
5
7
0
44
3
0
4
6
5
ss
s
k
ss
s
uk
uk
uk
d
uk
uk
uk
é
ù
--
+
-
ê
ú
=
ê
ú
--
+
-
ê
ú
ë
û
whe
r
e
[
]
s
uk
is
the unit-step function. In
th
is e
x
am
ple,
we assum
e
that the
sim
u
la
tion tim
e is 100 tim
e step.
1
10
00
k
H
,
2
10
10
k
H
3
10
01
k
and
H
Case
1
:
1
kk
H
H
Figure
1. Act
u
al and estim
a
t
e
d
value
of the
state
0
10
20
30
40
50
60
70
80
90
10
0
-0
.
5
0
0.
5
1
Ti
m
e
1s
t
e
l
em
en
t
o
f
s
t
at
e
ac
t
u
a
l
e
s
ti
m
a
te
d
0
10
20
30
40
50
60
70
80
90
10
0
-4
0
-3
0
-2
0
-1
0
0
10
Ti
m
e
2n
d e
l
em
e
n
t
o
f
s
t
at
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN:
2
088
-87
08
IJEC
E V
o
l. 5, No
. 2, A
p
ril 20
15
:
25
9 – 2
7
0
26
8
Figure
2.
Actual and estim
a
t
e
d
value
of the
input
Case 2
:
2
kk
H
H
Figure
3. Act
u
al and estim
a
t
e
d
value
of the
state
Figure
4. Act
u
al and estim
a
t
e
d
value
of the
input
0
10
20
30
40
50
60
70
80
90
10
0
-2
0
2
4
6
Ti
m
e
1s
t
e
l
em
en
t
o
f
i
n
pu
t
ac
t
u
a
l
e
s
ti
m
a
te
d
0
10
20
30
40
50
60
70
80
90
10
0
0
1
2
3
4
Ti
m
e
2n
d e
l
em
ent
of
i
n
p
u
t
0
10
20
30
40
50
60
70
80
90
100
-0
.
5
0
0.
5
1
Ti
m
e
1 s
t
e
l
em
en
t
o
f
s
t
at
e
ac
t
ual
e
s
ti
m
a
te
d
0
10
20
30
40
50
60
70
80
90
100
-4
0
-3
0
-2
0
-1
0
0
10
2n
d e
l
em
e
n
t
o
f
s
t
at
e
0
10
20
30
40
50
60
70
80
90
100
-2
0
2
4
6
Ti
m
e
1s
t
e
l
em
en
t
o
f
i
n
pu
t
ac
t
ual
es
t
i
m
a
t
e
d
0
10
20
30
40
50
60
70
80
90
100
-2
0
2
4
6
Ti
m
e
2n
d e
l
em
e
n
t
o
f
i
n
pu
t
Evaluation Warning : The document was created with Spire.PDF for Python.