TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4705 ~ 4
7
1
6
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.549
0
4705
Re
cei
v
ed
De
cem
ber 2
9
, 2013; Re
vi
sed
March 8, 201
4; Acce
pted
March 21, 20
14
Robust Centralized Fusion Kalman Filters with
Uncertain Noise Variances
Wen
-
juan Qi,
Peng Zhang
, Zi-li Deng*
Dep
a
rtment of Automatio
n
, Heilo
ng
jia
ng U
n
i
v
ersit
y
Harbi
n
, Chi
an, 150
08
0
*Corres
p
o
n
g
d
i
ng auth
o
r, e-mail: dzl@
hlj
u
.ed
u
.cn
A
b
st
r
a
ct
T
h
is pa
per stu
d
ies th
e pr
ob
le
m
of the
desi
g
nin
g
the r
o
b
u
st loca
l a
nd c
ent
rali
z
e
d
fusi
on
Kal
m
a
n
filters for mult
isens
or system w
i
th uncert
a
in n
o
is
e var
i
ances. Usi
ng
the mi
ni
max r
obust esti
mati
on
princi
pl
e, the central
i
z
e
d fusi
o
n
robust ti
me-v
aryin
g
Kal
m
a
n
filters are pres
ented
b
a
se
d o
n
the w
o
rst-cas
e
conserv
a
tive s
ystem w
i
th th
e cons
ervative
upp
er
bo
un
d
of nois
e
vari
ances. A Ly
ap
unov
appr
oac
h i
s
prop
osed
for the ro
bustn
ess
ana
lysis a
nd t
heir r
obust
acc
u
racy re
latio
n
s
are
prove
d
. It is prov
ed th
at th
e
robust
accurac
y
of rob
u
st ce
n
t
rali
z
e
d
fuser
is
hig
her t
han
th
ose of r
o
b
u
st l
o
cal K
a
l
m
an fi
lters. Speci
a
l
l
y, the
corresp
ond
in
g steady-state ro
bust loca
l an
d central
i
z
e
d fusi
on Kal
m
an filt
ers are als
o
pr
opos
ed a
nd th
e
conver
genc
e i
n
a re
ali
z
at
io
n
betw
een ti
me
-varying
an
d
steady-state K
a
l
m
a
n
filters i
s
prove
d
by t
h
e
dynamic
err
o
r system
analysis
(D
ESA) method
and dynam
ic
varianc
e error
system
analysis
(DVES
A)
meth
od. A Mon
t
e-Carlo si
mul
a
tion exa
m
ple s
how
s the robus
tness and
accu
racy relati
ons.
Ke
y
w
ords
:
mu
ltise
n
sor i
n
formatio
n
fusio
n
, central
i
z
e
d f
u
sio
n
, uncert
a
i
n
no
ise var
i
an
ce, mi
ni
max r
o
bus
t
Kal
m
a
n
filter
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
The aim
of th
e multisen
sor informatio
n f
u
sio
n
is ho
w
to com
b
ine th
e local e
s
tim
a
tors o
r
local me
asurements to o
b
t
ain the fused
estimators
, who
s
e a
c
curacy is hig
her
than that of each
local
estimat
o
r [1]. For th
e ce
ntrali
zed
fusion
optim
al Kalman filter, all the lo
cal mea
s
u
r
em
ent
data are
ca
rri
ed to the fusi
on ce
ntre to o
b
tain
a glob
al
ly optimal fused state e
s
timation [2].
The da
rwba
ck of th
e Kalm
an filter i
s
th
a
t
it only suita
b
le to h
andl
e
the
state e
s
timation
probl
me
s for system
s wit
h
exact mod
e
l para
m
et
ers and n
o
ise varian
ce
s. Howeve
r, in many
appli
c
ation
problem
s, there
exist un
ce
rta
i
nties of the
model p
a
ram
e
ters an
d/or
noise varia
n
ces.
Und
e
r
the
s
e
uncertaintie
s
the
pe
rform
a
nce of
t
he Ka
lman filter
will
deg
rade
[3], and a
n
inexa
c
t
model m
a
y cause the filte
r
to diverge.
This h
a
s
mot
i
vated the de
signi
ng of th
e rob
u
st Kal
m
an
filters, whi
c
h
guarante to h
a
ve a minima
l upper b
oun
d of the actu
al filtering error varia
n
ces
for
all admissibl
e
unce
r
taintie
s
.
In order to d
e
sig
n
the rob
u
st Kalman filter
s for the systems with t
he model pa
rameters
uncertaintie
s
,
two importa
nt appro
a
che
s
are t
he Ri
ccati equatio
n approa
ch [4-6] and the linear
matrix ineq
ua
lity (LMI) app
roa
c
h [7
-9]. T
he di
sadvant
age of the
s
e
two ap
pro
a
ch
es i
s
that onl
y
model
pa
ram
e
ters a
r
e
un
certain
whil
e t
he n
o
ise
va
ri
ances a
r
e
assume
d to
be
exactly kno
w
n
.
The
rob
u
st K
a
lman filte
r
in
g proble
m
s for
system
s
with u
n
certai
n noi
se
varia
n
ce
s
are
sel
dom
con
s
id
ere
d
[1
0, 11], and th
e rob
u
st info
rmation fu
si
on
Kalman filter are al
so
sel
d
om re
se
arch
ed
[12, 13].
In this pap
er,
using th
e mi
nimax rob
u
st
estimation p
r
inci
ple, the l
o
cal a
nd
cen
t
ralize
d
fusion
rob
u
st
time-varying
and
steady-state Ka
lma
n
filters a
r
e p
r
ese
n
ted b
a
se
d on the
wo
rst-
ca
se con
s
ervative system with the
con
s
ervativ
e
uppe
r bo
und of noi
se varian
ce
s. The
conve
r
ge
nce
in a reali
z
a
t
ion betwe
en
the ti
me-va
r
ying an
d st
eady-state K
a
lman filters is
rigo
rou
s
ly proved by the dynam
ic e
r
ror sy
stem a
nalysi
s
(DES
A) method [
14] and dyn
a
mic
variance error sy
stem analysi
s
(DVESA) me
thod [15]. Furthermor
e, a Ly
apunov equation
approa
ch i
s
pre
s
ente
d
fo
r the
rob
u
stn
e
s
s an
alysi
s
,
whi
c
h i
s
different from th
e
Ri
ccati eq
ua
tio
n
approa
ch
an
d the
LMI ap
proa
ch.
The
con
c
e
p
t of th
e ro
bu
st a
c
cura
cy is give
n an
d the
ro
bust
accuracy
rela
tions a
r
e
pro
v
ed, it is pro
v
ed that
the
robu
st a
c
curacy of
the
ce
ntralized fu
se
r i
s
highe
r than th
at of the local
robu
st Kalman filter.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4705 – 4
716
4706
The remain
d
e
r of thi
s
pa
per i
s
o
r
ga
ni
zed
as foll
o
w
s. Se
ction
2 gives th
e
probl
em
formulatio
n.
The
rob
u
st
centrali
zed
fu
sion time
-varyi
ng Kalm
an fil
t
ers are
pre
s
ented i
n
Se
ct
ion
3. The rob
u
st local and
ce
ntralized fusi
on stea
dy
-sta
te Kalman filters
are p
r
e
s
e
n
ted in Secti
on
4. The
ro
bu
st accu
ra
cy a
nalysi
s
i
s
giv
en in
Sectio
n 5. T
he
si
mulation
exa
m
ple i
s
give
n in
Section 6. Th
e con
c
lu
sio
n
is pro
p
o
s
ed in
Section 7.
2. Problem Formulation
Con
s
id
er the
muiltisen
so
r linear di
sce
r
et time-varyi
ng syste
m
with un
ce
rtai
n noise
var
a
inc
e
.
1
x
tt
x
t
t
w
t
(1)
,1
,
,
ii
yt
H
t
x
t
t
t
i
L
(2)
Whe
r
e
t
repres
ents
the
disc
rete time,
n
x
tR
is th
e
state,
i
m
i
yt
R
is the
measurement
of the
th
i
su
bsy
s
t
e
m,
r
wt
R
is the in
put noise,
t
is the com
m
on
disturban
ce
noise,
i
m
i
tR
is the measurement
noise of the
th
i
subsy
s
tem,
t
,
t
an
d
H
t
are kno
w
n
time-varying matrices with
approp
riate
di
mensi
o
n
s
.
L
is the numb
e
r of
sen
s
ors.
Assump
tion 1.
wt
,
t
and
i
t
are
u
n
co
rrelated
white noi
se
s
wi
th ze
ro m
ean
s an
d
unkno
wn un
certain a
c
tual varian
ce
s
Qt
,
Rt
an
d
i
Rt
at time
t
, res
p
ec
tively,
Qt
,
Rt
and
i
Rt
are kno
w
n con
s
e
r
vative uppe
r bou
n
d
s of
Qt
,
Rt
and
i
Rt
, s
a
tis
f
ying:
,,
ii
Q
t
Q
t
Rt
Rt
R
t
R
t
,
1,
,
iL
,
t
(3)
Assump
tion
2.
The initial state
0
x
is ind
epen
dent of
wt
,
t
and
i
vt
and h
a
s
mean value
and un
kno
w
n
uncertain a
c
t
ual varian
ce
0|
0
P
whi
c
h s
a
t
i
sf
ie
s:
0|
0
0
|
0
PP
(4)
Whe
r
e
0|
0
P
is a kn
own
con
s
e
r
vative upper b
ound of
0|
0
P
.
Assump
tion
3.
The
system (1
) an
d
(2) i
s
u
n
ifo
r
mly co
mplet
e
ly observab
l
e and
compl
e
tely co
ntrollabl
e.
Defining:
,1
,
,
ii
vt
t
t
i
L
(5)
Whe
r
e
i
vt
are
white noi
se
s with zero mean
s and
the con
s
e
r
va
tive and act
ual
varian
ce
s are
given as:
ii
v
Rt
R
t
R
t
,
ii
v
Rt
R
t
R
t
,
1,
,
iL
(6)
ij
v
Rt
R
t
,
ij
v
Rt
R
t
,
ij
(7)
From (3), we have:
ii
vv
Rt
Rt
,
1,
,
iL
,
t
(8)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Centralize
d
Fusi
on
Kalm
an Filte
r
s
with Un
ce
rtain Noi
s
e Va
riances (We
n
-j
uan Qi)
4707
3. Robus
t Ce
ntralized Fu
sion Time-v
a
r
y
i
ng Kalman Filters
Introdu
ce the
centralized fu
sion me
asure
m
ent equatio
n:
cc
c
yt
H
t
x
t
v
t
(9)
With the defin
ition:
T
TT
1
,,
cL
y
t
yt
yt
,
T
TT
,,
c
H
tH
t
H
t
,
T
TT
1
,,
cL
v
t
vt
vt
(10
)
And
c
vt
has the
conservative a
nd actu
al vari
ance matri
c
e
c
R
and
c
R
as
:
i
i
i
v
c
v
v
RR
R
R
R
RR
RR
R
,
i
i
i
v
c
v
v
RR
R
R
R
RR
RR
R
(11)
Therefore fro
m
(3) a
nd (8
),
accrodi
ng to t
he Lemma 1
and Lem
ma 2
in Appendix, we obtai
n:
cc
Rt
Rt
(12)
Based
on th
e wo
rst-case
con
s
e
r
vative
system
(1
)
and
(9) with
Assu
mption
s 1-3
and
con
s
e
r
vative uppe
r boun
d
s
Qt
and
c
Rt
, the gl
obally o
p
tima
l ce
ntrali
zed
f
u
se
d time
-va
r
ying
robu
st Kalma
n
filters are gi
ven as:
ˆˆ
|1
|
1
cc
c
c
c
x
tt
tx
t
t
K
t
y
t
(13)
=1
cn
c
c
tI
K
t
H
t
t
(14)
1
TT
=|
1
|
1
cc
c
c
c
c
c
Kt
P
t
t
H
t
H
t
P
t
t
H
t
R
t
(15)
TT
1|
|
cc
P
tt
t
P
t
t
t
t
Q
t
t
(16)
The fused co
nse
r
vative filtering e
r
ror va
rian
ce
|
c
Pt
t
is given as:
||
1
cn
c
c
c
Pt
t
I
K
t
H
t
P
t
t
(17)
It can be re
written as the L
y
apunov eq
u
a
tion:
T
T
TT
|1
|
1
11
1
cc
c
c
n
c
c
nc
c
c
c
c
Pt
t
t
P
t
t
t
t
H
t
t
Q
t
t
I
K
tH
t
K
tR
t
K
t
(18)
With the i
n
itial values
ˆ
0|
0
,
c
x
and
0|
0
0
|
0
c
PP
, where
n
I
is the
nn
identity
matrix.
The act
ual predictio
n and f
iltering e
rro
rs are obtai
ned
as:
ˆ
1|
1
1
|
1
|
cc
c
x
tt
x
t
x
t
t
t
x
t
t
t
w
t
(19)
ˆ
||
|
1
cc
n
c
c
c
c
c
x
t
t
x
t
xt
t
I
K
t
H
t
xt
t
K
t
v
t
(20)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4705 – 4
716
4708
Substituting (19) into (20)
yields:
|1
|
1
1
cc
c
n
c
c
c
c
x
tt
t
x
t
t
I
K
t
H
t
t
w
t
K
t
v
t
(21)
The a
c
tual
fuse
d filterin
g
error vari
an
ce
Τ
|
Ε
||
cc
c
Pt
t
x
t
t
x
t
t
, acc
o
rding to
(21),
we
have:
T
T
TT
|1
|
1
11
1
cc
c
c
n
c
c
nc
c
c
c
c
Pt
t
t
Pt
t
t
t
H
t
t
Q
t
t
I
K
t
H
t
K
tR
tK
t
(22)
With the initial value
0|
0
0
|
0
c
PP
.
Theorem 1.
For m
u
ltise
n
s
or un
ce
rtain
system
(1
) and (9)
with Assu
mption
s
1-3,
the
actual
central
i
zed
fusi
on
time-varyin
g
Kalm
an filters
with the con
s
ervative uppe
r boun
d
Qt
,
c
Rt
and
0|
0
c
P
are robu
st in the sen
s
e
that for all a
d
missible
act
ual varia
n
ces
,
c
Qt
R
t
and
0|
0
c
P
satisfying (3),
(4) an
d (12
)
, for arbit
r
ary time
t
, we have:
||
cc
P
tt
P
t
t
(23)
And
|
c
P
tt
is the m
i
nimal upp
er bound of
|
c
P
tt
for all admissi
ble un
certai
n
t
ies of noise
varian
ce
s. We call the a
c
tual fuse
d Kalman filter
s a
s
the robu
st ce
ntralized fusi
on Kalman filters.
Proof.
Defin
i
ng
||
|
cc
c
P
tt
P
t
t
P
tt
, subtract
ing (22
)
fro
m
(18) yiel
ds the
Lyapun
ov eq
uation.
T
|1
|
1
cc
c
c
c
P
tt
t
P
t
t
t
U
t
(24)
T
T
T
11
1
1
cn
c
c
n
c
c
cc
c
c
Ut
t
t
t
Q
t
Q
t
t
I
K
t
H
t
K
t
Rt
Rt
K
t
(25
)
Applying (3
), (12
)
and (25
)
yields that
0
c
Ut
, a
nd from (4) we have:
0
|
0
0
|0
0
|
0
0
|0
0
|
0
0
cc
c
PP
P
P
P
(26)
Hen
c
e f
r
om
(24
)
, we
hav
e
1|1
0
c
P
. Applying t
he math
emat
ical in
du
ction
method yiel
ds
|0
c
Pt
t
, for all time
t
, i.e.
the ineq
uality (23) h
o
lds. Ta
king
,
cc
Q
t
Q
t
Rt
Rt
an
d
0|
0
0
|
0
PP
, then compa
r
ing (1
8) with
(22), we h
a
ve
||
cc
P
tt
P
t
t
. For arbitra
r
y other
uppe
r bou
nd
*
|
c
P
tt
, we have
*
||
|
cc
c
P
tt
P
t
t
P
t
t
which yields that
|
c
P
tt
is the minima
l
uppe
r bou
nd
of
|
c
P
tt
. The proof
is com
p
leted
.
Corollar
y
1.
For un
ce
rtain
multisen
sor system
(1
) a
nd (2) with
A
s
sumption
s
1
-
3 and
con
s
e
r
vative uppe
r bo
und
s
Qt
and
i
v
Rt
, similar to the rob
u
st
centralized f
u
sio
n
time-va
r
ying
Kalman filters, the robu
st local time
-va
r
ying Kalman filters a
r
e give
n by:
ˆˆ
|1
|
1
ii
i
i
i
x
tt
t
x
t
t
K
t
y
t
,
1,
,
iL
(27)
=1
in
i
tI
K
t
H
t
t
,
T1
=|
1
i
ii
K
tP
t
t
Ht
R
t
(28)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Centralize
d
Fusi
on
Kalm
an Filte
r
s
with Un
ce
rtain Noi
s
e Va
riances (We
n
-j
uan Qi)
4709
T
|1
ii
iv
Rt
H
t
P
t
t
H
t
R
t
(29)
TT
1|
|
ii
P
tt
t
P
t
t
t
t
Q
t
t
(30)
||
1
in
i
i
Ptt
I
K
t
Ht
Ptt
(31)
The co
nserv
a
tive local filtering e
r
ror v
a
rian
ce
|
i
P
tt
ca
n b
e
rewritten a
s
the Lyapu
n
o
v
equatio
n [2].
TT
T
T
|1
|
1
1
1
1
i
ii
i
i
n
i
ni
i
v
i
Ptt
t
P
t
t
t
t
t
t
Q
t
t
I
K
tH
t
K
tR
tK
t
(32)
With the initial values
0|
0
0
|
0
i
PP
. And
the actual filtering e
r
ror va
rian
ce
s are given by the
Lyapun
ov eq
uation
s
.
TT
T
T
|1
|
1
1
1
1
i
ii
i
i
n
i
ni
i
v
i
Pt
t
t
Pt
t
t
t
t
t
Q
t
t
I
K
tH
t
K
tR
tK
t
(33)
Similarly, the local time
-varying Ka
lman filters a
r
e al
so
robu
st, i.e.,
||
ii
P
tt
P
t
t
,
1,
,
iL
(34)
4. Robus
t Lo
cal and Ce
ntralized Fusio
n Stead
y
-
sta
t
e Kalman Fi
lters
Theorem 2.
For multisensor
un
ce
rtain time
-inv
ariant
sy
ste
m
(1
) a
n
d
(9
)
with
Assu
mption
1 and 3,
wh
ere
t
,
t
,
H
tH
,
,,
ii
Qt
Q
R
t
R
R
t
R
,and
Qt
Q
,
Rt
R
,
ii
Rt
R
are all th
e
con
s
tant m
a
trice
s
, t
hen th
e
actual
centra
lized fu
sio
n
steady-state
Kalman filters are given by:
ˆˆ
|1
|
1
ss
cc
c
c
c
x
tt
x
t
t
K
y
t
(35)
=
cn
c
c
IK
H
,
1
TT
=
cc
c
c
c
c
c
KH
H
H
R
(36)
TT
cc
Q
,
=
cn
c
c
c
PI
K
H
(37)
The predi
ctio
n error vari
an
ce
c
satisfie
s th
e steady-stat
e
Riccati eq
u
a
tion:
1
TT
T
T
=
c
c
c
c
cc
c
c
cc
HH
H
R
H
Q
(38)
Whe
r
e the
superscri
pt s
denote
s
“ste
ady-state
”,
th
e fuse
d co
nservative filteri
ng erro
r
var
i
anc
e
c
P
is given as:
T
TT
T
cc
c
c
n
c
c
n
c
c
c
c
c
PP
H
Q
I
K
H
K
R
K
(39)
The fused act
ual filtering e
r
ror vari
an
ce
c
P
is given a
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4705 – 4
716
4710
T
TT
T
cc
c
c
n
c
c
n
c
c
c
c
c
PP
H
Q
I
K
H
K
R
K
(40)
The a
c
tual
centrali
zed fu
sion ste
ady-st
a
te Ka
lman fi
lters
(35
)
a
r
e
rob
u
st in th
e se
nse
that for all admissi
ble un
ce
rtainties of no
ise varia
n
ces
Q
and
i
v
R
satisfying
(3) an
d (8
), we have:
cc
PP
(41)
And
c
P
is the min
i
mal uppe
r bo
und of
c
P
.
Proof.
As
t
, taking
the limit
operation
s
fo
r (13)-(
18
),
(2
2)
a
nd (2
3), we obtain
(3
5
)
-
(
4
1)
. T
a
k
i
ng
,
cc
QQ
R
R
, from (39
)
a
nd (40
)
, we
have
cc
P
P
. I
f
c
P
is arbitrary othe
r uppe
r
boun
d of
c
P
for
all admi
s
sible
Q
and
c
R
sat
i
sf
y
i
ng
,
cc
QQ
R
R
, then
we
h
a
ve
cc
c
PP
P
, which
yields that
c
P
is minimal up
pe
r boun
d of
c
P
. The pro
o
f is co
mpleted.
Similarly, the actual lo
cal
steady-s
tate Kalman filters are given by:
ˆˆ
|1
|
1
ss
ii
i
i
i
x
tt
x
t
t
K
y
t
,
1,
,
iL
(42)
=
in
i
IK
H
,
1
TT
=
i
ii
i
v
KH
H
H
R
,
in
i
i
PI
K
H
(43)
The predi
ctio
n error vari
an
ce
i
satisfie
s th
e steady-stat
e
Riccati eq
u
a
tion.
1
TT
T
T
=
i
ii
i
i
v
i
HH
H
R
H
Q
(44)
The co
nserv
a
tive and actual local filteri
ng e
r
ror varian
ce
s satisfy the steady-state
Lyapun
ov eq
uation
s
.
T
TT
T
i
ii
i
i
n
i
n
i
i
v
i
P
P
Q
I
KH
KR
K
(45)
T
TT
T
i
ii
i
i
n
i
n
i
i
v
i
P
P
Q
I
KH
KR
K
(46)
The act
ual lo
cal ste
ady-state Kalman filters (42
)
are
robu
st, i.e.,
ii
PP
,
1,
,
iL
(47)
And
i
P
is the min
i
mal uppe
r bo
und of
i
P
.
Theorem 3.
Und
e
r th
e co
ndition
s of T
heorem 2,
an
d assu
me th
at the mea
s
u
r
eme
n
ts
i
yt
,
1,
,
iL
are bou
nde
d, then the robu
st time-v
ar
ying an
d steady-state
Kalman filters
ˆ
|
i
x
tt
and
ˆ
|
s
i
x
tt
,
ˆ
|
c
x
tt
and
ˆ
|
s
c
x
tt
given by
(27
)
an
d (42),
(1
3)
and
(3
5) hav
e ea
ch
othe
r
the conve
r
ge
nce in a reali
z
ation, such that:
ˆˆ
||
0
s
ii
xt
t
x
t
t
, as
t
, i.a.r
(48)
ˆˆ
||
0
s
cc
xt
t
x
t
t
, as
t
, i.a.r
(49)
Whe
r
e the
n
o
tation “i.a.r”
denote
s
the
conve
r
ge
nce
in a re
alizatio
n [15], and
we have
the conve
r
ge
nce of varia
n
c
e
s
.
|
ii
P
tt
P
,
|
ii
P
tt
P
, as
t
,
1,
,
iL
(50)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Centralize
d
Fusi
on
Kalm
an Filte
r
s
with Un
ce
rtain Noi
s
e Va
riances (We
n
-j
uan Qi)
4711
|
cc
P
tt
P
,
|
cc
P
tt
P
, as
t
(51)
Proof.
According to the compl
e
te ob
serva
b
ility and compl
e
te controllability
of each
sub
s
ystem, t
he time-varyi
ng local Kalm
an filt
ers (27)
have the co
n
v
ergen
ce that
[16]:
|1
ii
Pt
t
, as
t
,
1,
,
iL
(52)
From (28
)
an
d (31
)
, we ha
ve:
ii
t
,
ii
K
tK
,
|
ii
P
tt
P
, as
t
,
1,
,
iL
(53)
Setting
ii
i
tt
,
ii
i
K
tK
K
t
in (2
7), ap
plying
(53) yiel
ds
0
i
t
0
i
Kt
, as
t
. Subtra
cting
(42
)
from (2
7), a
n
d definin
g
ˆˆ
||
s
ii
i
tx
t
t
x
t
t
, we
have:
1
ii
i
i
tt
u
t
(54)
With
ˆ
1|
1
ii
i
i
i
ut
t
x
t
t
K
t
yt
. Noting that
i
t
is uniforml
y
asymptoticall
y
stable [1
7], and
ii
K
ty
t
is b
oun
ded, ap
plyin
g
Lemm
a
4
to (27
)
yield
s
the
boun
dedn
ess of
ˆ
|
i
x
tt
. Hen
c
e we have
0
i
ut
. Applying Lemma
4 to (54), noti
ng that
i
is
a stabl
e ma
trix, so it is also
uni
fo
rmly asympto
t
ically stabl
e
,
hence
0
i
t
, i.e.
the
conve
r
ge
nce (48
)
hold
s
. The co
nverg
e
n
c
e of (4
9) can
be proved
si
milarly.
From (33
)
an
d (46
)
, definin
g
|
ii
i
tP
t
t
P
yield the Lyapun
ov equat
ion.
T
1
ii
i
i
i
tt
U
t
(55)
T
TT
TT
T
T
1|
1
1|
1
i
i
in
i
n
i
i
v
i
ni
n
i
i
v
i
i
i
i
ii
i
i
i
Ut
t
Q
I
K
t
H
K
t
R
K
t
Q
I
KH
KR
K
P
t
t
t
tP
t
t
t
t
(56)
From (33),
noting that
i
t
is uniformly a
s
ymptotically
stable, appl
ying
ii
K
tK
,
0
i
t
and L
e
mma
3 yields
1|
1
i
Pt
t
is boun
ded. F
r
om
(56
)
yie
l
ds that
0
i
Ut
.
Applying Lem
ma 3 to (55
)
yields
0
i
t
, as
t
, i.e.,
|
ii
P
tt
P
holds. Sim
ilarly, we can
prove (51) h
o
l
ds. The p
r
oof
is com
p
leted
.
5. The Accurac
y
Anal
y
s
is
Defini
tion 1.
The t
r
ace
tr
|
P
tt
of the up
pe
r b
ound
|
P
tt
of the a
c
tual filteri
n
g
error
var
i
anc
e
s
|
P
tt
for all admi
ssi
ble
unce
r
taintie
s
is call
ed the
robu
st a
c
cura
cy or gl
obal a
c
cura
cy
of a robu
st Kalman filter, a
nd
tr
|
P
tt
is called a
s
its actu
al accuracy.
From this d
e
finition, the smaller
tr
|
P
tt
or
tr
|
P
tt
means the high
er robu
st accuracy
or actu
al accura
cy. The ro
bust a
c
cura
cy gives the lowe
st boun
d o
f
all possi
ble actual a
c
cura
cie
s
yielded from t
he un
certai
nties of noi
se varian
ce
s.
Theorem 4.
For m
u
ltise
n
s
or un
ce
rtain
system
(1
) and (2)
with Assu
mption
s
1-3,
the
accuracy co
mpari
s
o
n
of the local and f
u
se
d rob
u
st
Kalman filters is given by:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4705 – 4
716
4712
||
,
ii
P
tt
P
t
t
1,
,
iL
(57)
||
|
,
cc
i
P
tt
P
t
t
P
tt
1,
,
iL
(58)
tr
|
t
r
|
ii
P
tt
P
t
t
,
tr
|
t
r
|
t
r
|
cc
i
P
tt
P
t
t
P
tt
,
1,
,
iL
(59)
,,
1
,
,
,
ii
c
c
i
P
P
PPP
i
L
(60)
tr
t
r
ii
PP
,
1,
,
,
iL
tr
tr
tr
cc
i
PPP
(61)
Proof.
Acc
o
rding to the
ro
bustn
ess
(23
)
and
(3
4),
we
have
(57
)
a
n
d
the first in
e
quality
of (58).
The
se
con
d
in
equ
ality of (58
)
h
a
s
bee
n p
r
ov
en in
[18]. Ta
king
the trace
ope
ration
s f
o
r
(57
)
and
(58
)
yields the ine
qualitie
s (59
)
.
As
t
, taking th
e limit operati
ons fo
r (57
)
, (58) a
nd
(59
)
yields (6
0) and
(61
)
. The pro
o
f is co
mpleted.
From the ine
qualitie
s (59
)
, we can see
that all admissi
ble actu
al trace
s
tr
|
i
P
tt
and
tr
|
c
P
tt
are glob
ally controlled by
the upper b
ound
tr
|
i
P
tt
and
tr
|
c
P
tt
, res
p
ec
tively, and
the robu
st a
c
cura
cy
of th
e cent
ralized
ro
bu
st fuser is hig
her th
an that
of e
a
ch
lo
cal
ro
b
u
st
Kalman filter.
6. Similation Example
Con
s
id
er a th
ree
-
sen
s
or ti
me-inva
r
iant trac
kin
g
syste
m
with uncert
a
in noi
se vari
ances.
1
x
tx
t
w
t
,
,1
,
2
,
3
ii
yt
H
x
t
t
t
i
(62)
2
0
0
0
1
0.
5
,
01
T
T
,
2
H
I
(63)
Whe
r
e
0
0.25
T
is the sampl
ed pe
ri
od,
T
12
,
x
tx
t
x
t
is the sta
t
e,
1
x
t
and
2
x
t
are
the
po
sition a
nd vel
o
ci
ty of target
at
time
0
tT
.
wt
,
t
and
i
t
are ind
epe
nde
n
t
Gau
ssi
on
white noi
se
s with ze
ro me
an
a
nd un
kno
w
n un
ce
rtain actual
vari
an
ce
s
Q
,
R
and
i
R
r
e
spec
tively.
In the s
i
mulation, we tak
e
1
Q
,
0.
8
Q
,
dia
g
(
1
.5
,
2
.5
)
R
,
di
a
g
(
1
,
2
)
R
,
1
di
a
g
(
3
.6
,
2
.
5
)
R
,
1
di
a
g
(
3
,1
.
8
)
R
,
2
dia
g
(
8
,
0
.36
)
R
,
2
di
a
g
(6
,
0
.
2
5
)
R
,
3
diag
(
0
.5
,
2
.8
)
R
,
3
diag
(
0
.
3
8
,
2
)
R
, the
initial values
T
00
0
x
,
0
,
2
0
|
0
d
iag
(
1.
1
,
1.2
)
,
0
|
0
P
PI
.
The comp
ari
s
on
s of the fi
ltering e
r
ror v
a
rian
ce m
a
tri
c
e
s
an
d their trace
s
of the
robu
st
steady-state
local a
nd
ce
ntralized fu
si
on Kalman
fi
lters a
r
e
sho
w
n in T
able
1 and T
able
2.
These matri
c
es an
d their trace
s
veri
fy the accuracy re
lations (60
)
-(61).
The traces
of the con
s
erva
tive and a
c
tu
al ro
bu
st filtering e
rro
r
va
ri
ances are
co
mpared
in Figu
re
1.
We
se
e that
the tra
c
e
s
of
the lo
cal
an
d fused
rob
u
s
t time-va
r
yin
g
Kalma
n
filters
quickly
conv
erge
to th
ese of th
e
co
rresp
ondi
ng
st
eady-state K
a
lman
filters,
whi
c
h
sho
w
the
robu
st accu
ra
cy relation
s (59) an
d (6
1)
hold.
Table 1. The
Con
s
e
r
vative and Actual A
c
cura
cy Co
m
pari
s
on of
i
P
and
i
P
1,
2
,
3
,
ic
1
P
2
P
3
P
c
P
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Centralize
d
Fusi
on
Kalm
an Filte
r
s
with Un
ce
rtain Noi
s
e Va
riances (We
n
-j
uan Qi)
4713
0
10
20
30
40
50
60
70
80
90
100
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
t/
st
ep
2
tr
P
1
tr
P
2
tr
P
3
tr
P
1
tr
P
tr
c
P
3
tr
P
tr
c
P
0.8247
0.3416
0.34
16
0.3750
1.0554
0.3278
0.3278
0.3405
0.4360
0.2383
0.238
3
0.3233
0.3771
0.1
956
0.1956
0.2
805
1
P
2
P
3
P
c
P
0.6442
0.2
6
69
0.2669
0.2
9
56
0.7994
0.2
5
45
0.2545
0.26
89
0.3119
0.1
770
0.1770
0.2
495
0.2726
0.1
478
0.1478
0.2191
Table 2. The
Con
s
e
r
vative and Actual A
c
cura
cy Co
m
pari
s
on of
tr
i
P
,
tr
i
P
,
1,
2
,
3
,
ic
1
tr
P
,
1
tr
P
2
tr
P
,
2
tr
P
3
tr
P
,
3
tr
P
tr
c
P
,
tr
c
P
1.1998,0.939
8
1.3959,1.068
3
0.7593,0.561
3
0.6576,0.491
7
Figure 1. The
Traces of the
Con
s
ervative
and Actual L
o
cal a
nd Fu
sed Kalman Fil
t
ers
In orde
r to v
e
rify the abo
ve theoreti
c
al
accuracy
rel
a
tions, ta
king
20
0
Monte Carl
o
simulatio
n
ru
ns, the mean
squa
re erro
r (MSE) value
s
at time
t
of local or fu
sed robu
st Kalma
n
filters are defi
ned a
s
:
Τ
1
1
ˆ
ˆ
MS
E
|
|
jj
jj
j
t
x
tx
t
t
xt
xt
t
,
1,
2
,
3
,
c
(64
)
Whe
r
e
j
x
t
or
ˆ
|
j
x
tt
den
otes the
th
j
reali
z
ation of
x
t
or
ˆ
|
x
tt
.
Acco
rdi
ng to the erg
odi
city [19], we have
:
MS
E
t
r
tP
,
as
,
t
,
1,
2
,
3
,
c
(65)
The MSE curves of the local and fuse
d time-v
arying robu
st Kalman filters are
shown in
Figure 2, whi
c
h verify the accuracy rela
tions
(5
9) a
n
d
(61), an
d verify the ergodi
city (65).
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 6, June 20
14: 4705 – 4
716
4714
0.
4
0.
6
0.
8
1
1.
2
1.
4
1.
6
20
30
40
50
60
70
80
90
100
110
120
130
140
150
160
170
180
190
200
210
220
230
240
250
260
270
280
290
300
MSE1
MSE2
MSE3
MSEc
Figure 2. The
Compa
r
i
s
on
of
MS
E
t
and
tr
P
,
1,
2
,
3
,
c
7. Conclusio
n
For
multisen
sor system with
un
certai
n
noise vari
ances, u
s
ing
the minima
x robust
estimation
pri
n
cipl
e, the l
c
oal an
d centralize
d
fu
si
on
robu
st Kalm
a
n
time-va
r
yin
g
Kalman
filters
are
pre
s
e
n
te
d. Base
d on
the Lyapu
no
v equation
a
ppro
a
ch, thei
r robu
stne
ss
are
prove
d
a
n
d
their ro
bu
st a
c
cura
cy rel
a
tions a
r
e al
so
proved.
It is
proved th
at the ro
bu
st accru
a
ci
es
of the
centralized fu
sion K
a
lman
filters a
r
e
hig
her th
an
tho
s
e of the lo
cal
rob
u
st Kalm
an filters. Th
e
conve
r
ge
nce
probl
em
of th
e robu
st lo
ca
l and
c
entrali
zed
fusi
on ti
me-varyin
g
a
nd
steady
-sta
te
Kalman filters is proved by
the dynamic
error
sy
stem
analysi
s
(DE
SA) method
and the dyna
mic
varaince error system
anal
ysis (DVESA) method. Thi
s
extens
i
on of this paper t
o
system
s wi
th
uncertain n
o
i
s
e varia
n
ces
and mod
e
l pa
ramete
rs i
s
u
nder
study.
Ackn
o
w
l
e
dg
ements
This wo
rk i
s
sup
porte
d by
the
Natural
Scien
c
e
Fou
ndation
of
Ch
ina u
nde
r g
r
a
n
t NSF
C
-
6087
4063, the Innovatio
n and Scie
ntific Re
sea
r
ch Fou
ndati
on of grad
u
a
te stude
nt of
Heilo
ngjian
g
Province und
er grant YJS
C
X201
2-263
HL
J.
Referen
ces
[1]
Hall DL, Linas JL.
An introd
uc
tion to multise
n
s
or data fusio
n
, Proceedi
ng of
IEEE.1997; 85
(1): 6-23.
[2]
Li XR, Z
hu YM, Han CZ
.
Optimal li
near
esti
mation fu
sion, Part I:
Unifie
d fusio
n
rules,
IE
EE
T
r
ansactio
n
s o
n
Information T
heory
, 20
03; 4
9
(9): 219
2-2
2
3
0
.
[3]
Sri
y
an
and
a H.
A simple m
e
th
od for the co
nt
rol of div
e
rge
n
c
e in Ka
lman fi
lter alg
o
rithms,
Internatio
nal
Journ
a
l of Co
ntrol
. 197
2; 16(6)
: 1101-1
1
0
6
.
[4]
Le
w
i
s F
L
,
Xi
e LX, Po
pa D.
Optima
l an
d r
obust esti
mati
on.
Seco
nd E
d
itio
n. CRC Pr
ess, Ne
w
York
.
200
8.
[5]
Lu
X, Z
h
a
ng
H, W
ang W
.
Rob
u
st Kalm
a
n
filter
i
ng for
discrete-tim
e
s
y
stems
w
i
t
h
measur
emen
t
del
a
y
ed.
IEEE Transactions on Circuits and
System
s-II: Express Briefs.
2
007; 54(
6): 522
-626.
[6]
Xi
on
g K, W
e
i
CL, Li
u
LD. R
o
bust Ka
lman
fil
t
erin
g
for d
i
scr
ete-time
non
lin
ear s
y
stems
w
i
th par
ameter
uncerta
inties,
Aerosp
ace Sci
ence a
nd T
e
ch
nol
ogy
, 20
12; 18(1): 15-
24.
[7]
Jin
XB, Bao J, Z
hang J
L
.
Centrali
z
e
d fus
i
on estimation for unc
ertain m
u
ltisensor
system
based
on LM
I
meth
od,
Proce
edi
ng of the IEEE conferenc
e
on Mec
hatro
ni
cs and Autom
a
tion. 200
9; 238
3-23
87.
[8]
Yang F
,
Li Y.
Rob
u
st set-membershi
p
filteri
ng
for s
y
stems
w
i
th missi
ng
measur
ement: a line
a
r matri
x
ine
qua
lit
y
ap
pr
oach,
IET
Sign
al Process,
20
12; 6(4): 34
1-3
47.
[9]
Qu XM, Z
h
o
u
J.
T
he optima
l
robust fin
i
te-h
orizo
n
Kalm
an
filterin
g for m
u
ltipl
e
se
nsors
w
i
t
h
d
i
fferent
stochastic fail
u
r
e rates,
Appli
e
d Mathe
m
atics
Letters
, 201
3; 26(1): 80-
86.
1
tr
P
2
tr
P
1
tr
P
3
tr
P
tr
c
P
2
tr
P
3
tr
P
tr
c
P
t/s
te
p
MSE
Evaluation Warning : The document was created with Spire.PDF for Python.