TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.6, Jun
e
201
4, pp. 4692 ~ 4
7
0
4
DOI: 10.115
9
1
/telkomni
ka.
v
12i6.545
3
4692
Re
cei
v
ed
De
cem
ber 2
9
, 2013; Re
vi
sed
March 8, 201
4; Acce
pted
March 20, 20
14
Robust Weighted Measurement Fusion Kalman
Predictors with Uncertain Noise Variances
Wen
-
juan Qi,
Peng Zhang
, Gui-huan Nie, 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
For the multis
ensor system
with uncertain
noise
varianc
es, using the
m
i
ni
m
a
x r
o
bust estimatio
n
princi
pl
e, the l
o
cal a
nd w
e
i
ght
ed
me
asur
eme
n
t fusion
rob
u
s
t
time-v
aryin
g
Kal
m
a
n
pre
d
ict
o
rs are
pres
ent
ed
based on the
worst-case co
nservative syst
em
with the conservative
upper bound of no
is
e varianc
es. The
actual
pred
icti
on err
o
r vari
a
n
ces ar
e gu
ar
antee
d to h
a
v
e
a
min
i
mal
upp
er bo
un
d for all
ad
miss
i
b
le
uncerta
inties
o
f
noise
vari
anc
es. A Lya
p
u
n
o
v
appr
oac
h is
prop
osed
for t
he ro
bustn
ess
ana
lysis a
n
d
thei
r
robust accur
a
c
y
relatio
n
s are
proved. It is prove
d
that the robust accur
a
cy of
w
e
ighte
d
me
asur
e
m
e
n
t
robust fus
e
r
is
hig
her t
han
tha
t
of e
a
ch
loc
a
l
robus
t K
a
l
m
an
pred
ictor. Sp
ec
ially,
the
corres
pon
din
g
ste
ady
-
state robust l
o
cal a
nd w
e
ig
hted meas
ure
m
e
n
t fusion
K
a
l
m
a
n
pred
ictors are also
prop
osed a
n
d
the
conver
genc
e i
n
a re
ali
z
at
ion
betw
een ti
me-
v
aryin
g
an
d
steady-state K
a
l
m
a
n
pr
edictors
is prove
d
by t
h
e
dynamic error
system
analysis
(DESA)
m
e
thod. A M
onte-Carlo
s
i
m
u
lation exam
ple
s
hows the
effectiveness
o
f
the robustnes
s and accur
a
cy
relatio
n
s.
Ke
y
w
ords
:
mu
ltise
n
sor inf
o
rmatio
n
fu
sio
n
, w
e
ighted
me
asur
e
m
ent
fusion, unc
ertain n
o
ise var
i
ance,
mi
ni
max ro
bust
Kalman pr
edic
t
or
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
Multisen
so
r i
n
formatio
n fu
sion i
s
widely
applie
d to m
any filelds i
n
cluding
defen
ce, target
tracking,
GP
S po
sition
a
nd
so
on [1,
2]. Its aim
i
s
to
co
mbin
e
the lo
cal
e
s
timators o
r
l
o
cal
measurement
s to
obtain t
he fu
sed
esti
mators of
th
e sy
stem
sta
t
e, who
s
e
accuracy i
s
hig
her
than that
of e
a
ch
lo
cal
esti
mator. Kalm
a
n
filter
in
g a
p
p
roa
c
h
i
s
the
ba
sic tool
of
the i
n
form
ation
fusion
with th
e assu
mption
that the mo
d
e
l pa
ra
mete
rs an
d noi
se
varian
ce
s
are
exactly kn
own.
W
h
en
th
er
e
e
x
is
t u
n
c
e
r
t
a
i
n
t
ie
s
,
th
e pe
rfo
r
ma
nc
e
of t
he Kalm
an fil
t
er
can
be
ve
ry po
or [3], a
n
d
an inexact m
odel may ca
u
s
e the diverg
ent filter. In
o
r
de
r to handl
e this pro
b
le
m, various st
udie
s
on desi
gnin
g
of the robust Kalman filte
r
s have be
en
reporte
d [4-6]. The robu
st Kalman filters
guarantee
to
have a
minim
a
l up
per b
oun
d of th
e a
c
tu
a
l
filtering
erro
r varia
n
ces for all a
d
mi
ssi
bl
e
uncertaintie
s
.
For the
syst
ems
with the
model p
a
ra
meters
un
ce
rtainties, there are t
w
o i
m
porta
nt
approa
che
s
f
o
r
de
signin
g
t
he
robu
st Kal
m
an filter
s
su
ch th
at the
Ri
ccati eq
uatio
n ap
pro
a
ch [
4
],
[7-8] and the
linear mat
r
ix inequality (L
MI) app
ro
a
c
h
[5-6], [9]. The disa
dvanta
ge of these two
approa
che
s
i
s
that o
n
ly model p
a
ra
met
e
rs are
un
ce
rtain whil
e the
noise vari
an
ce
s a
r
e a
s
su
med
to be exactly kno
w
n. The
robu
st Kalman filter
ing p
r
oblem
s for systems
with uncertain n
o
i
s
e
varian
ce
s are
seldom
con
s
idere
d
[10, 11], and t
he robu
st inform
ation fusio
n
Kalman filter
are
also
seld
om rese
arche
d
[12, 13].
For informati
on fusio
n
based on the Kal
m
an filt
ering,
there exist two methodol
og
ies [14,
15], the state and mea
s
urem
ent fusi
on method
s,
the former
method can
give a fused
state
estimato
r by combi
ng o
r
weighting the l
o
cal
state e
s
timators, whil
e the later fu
sion m
e
thod i
s
to
weig
ht all the
local
mea
s
u
r
ement to o
b
tain
a fu
se
d
measurement
equatio
n, an
d then to
obt
ain
global o
p
tima
l state estima
tor based on
a singl
e Kalm
an filter.
In this
pape
r, usin
g the
minimax rob
u
st e
s
timatio
n
pri
n
ci
ple, t
he lo
cal
and
wei
ghted
measurement
fusio
n
rob
u
s
t time-va
r
yi
ng a
nd
st
ea
dy-state
Kal
m
an
pre
d
icto
rs are
pre
s
e
n
ted
based o
n
th
e wo
rst
-
case
con
s
e
r
vative system
wit
h
the con
s
ervative upper boun
d of n
o
ise
varian
ce
s. Th
e co
nverg
e
n
c
e in a realization
between t
he time-va
r
ying and
stea
d
y
-state Kalm
an
predi
ctors i
s
rigorously proved
by the
dynamic error sy
stem an
alysis (DESA) method
[16, 17].
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Weig
hted Mea
s
u
r
e
m
ent Fusion
Kalm
an Predi
ctors with
Un
certai
n… (We
n
-jua
n Qi)
4693
Furthe
rmo
r
e, a Lyapunov e
quation
ap
proach is prese
n
ted for the
robu
stne
ss a
n
a
lysis, which is
different fro
m
the Ri
ccati e
quation
app
roach an
d
the
LMI app
ro
ach. The
con
c
e
p
t of the rob
u
st
accuracy
is
given a
nd th
e robu
st a
ccura
cy rela
tio
n
s
are p
r
ove
d
, it is prove
d
that the
ro
bust
accuracy of t
he ro
bu
st we
ighed m
e
a
s
u
r
eme
n
t fusi
o
n
Kalman p
r
edicto
r
is hi
g
her tha
n
that
o
f
each local ro
bust Kalma
n
predi
cto
r
.
The
remai
n
d
e
r of thi
s
p
aper is orga
nize
d a
s
foll
ows. Sectio
n
2 give
s p
r
oblem
formulatio
n. The robu
st
weig
hted me
asu
r
em
ent
fusio
n
time-v
arying Kalm
an p
r
edi
ctors are
pre
s
ente
d
in
Section
3.
The
rob
u
st
local
an
d f
u
se
d
steady-state Kalm
a
n
predi
ctors are
pre
s
ente
d
in
Section 4.
The robu
st a
c
cura
cy an
al
ysis i
s
given
in Section 5
.
The simul
a
tion
example is gi
ven in Sectio
n 6. The co
nclusio
n
is prop
ose
d
in Secti
on 7.
2. Problem Formulation
Con
s
id
er m
u
i
l
tisen
s
or line
a
r di
sceret time-v
aryin
g
system with
u
n
ce
rtain n
o
ise varain
ce
and ide
n
tical
measurement
matrix.
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 dis
c
rete time,
n
x
tR
is the state,
m
i
y
tR
is the
measurement
of the
th
i
s
u
bsys
tem,
r
wt
R
is the input noi
se,
m
tR
is the co
mmon
disturban
ce noise,
m
i
tR
is the
measureme
n
t noise of the
th
i
sub
s
y
s
t
e
m
,
t
,
t
an
d
H
t
are kno
w
n time-varyin
g
m
a
trice
s
with a
ppro
p
ri
ate 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 mea
n
s, the
con
s
e
r
vative and
actual va
rian
ce
s
and cro
s
s-co
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:
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: 4692 – 4
704
4694
ii
vv
Rt
Rt
,
1,
,
iL
,
t
(8)
3. Robus
t Weighted M
e
a
s
uremen
t Fu
sion Time-v
a
r
y
i
ng Kalman Predictor
s
Introdu
ce the
centralized fu
sion me
asure
m
ent equatio
n.
cc
c
c
yt
H
t
x
t
v
t
e
H
t
x
t
v
t
(9)
With the defin
itions:
T
TT
1
,,
cL
y
t
yt
yt
,
T
TT
,,
c
H
tH
t
H
t
,
T
TT
1
,,
cL
v
t
vt
vt
T
,,
mm
eI
I
(10)
And the fuse
d noise
c
vt
resp
ectively has t
he co
nserva
ti
ve and actu
al
variances a
s
:
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)
Applying the
weig
hted l
e
a
s
t
squa
re
me
thod [18
], fro
m
(9
),
we
ha
ve the
co
nse
r
vative
weig
hted fusi
on mea
s
u
r
em
ent equatio
n.
MM
y
tH
t
x
t
v
t
(12)
Whe
r
e
M
y
t
is the
con
s
e
r
vative weig
hted
fusion m
e
asu
r
em
ent
and
M
vt
is the
con
s
e
r
vative fused me
asurement white
noise with co
nse
r
vative varian
ce
M
Rt
, s
u
c
h
t
hat:
T1
1
T
1
(
(
)
)
()
()
Mc
c
c
y
t
e
Rt
ee
Rt
y
t
(13)
T1
1
T
1
((
)
)
(
)
(
)
Mc
c
c
v
t
e
R
te
e
R
tv
t
(14)
T1
1
()
(
(
)
)
Mc
Rt
e
R
t
e
(15)
Based
on the
worst-ca
se
conservative system (1
) an
d (12
)
with A
s
sumption
s 1
-
3 an
d
con
s
e
r
vative uppe
r
b
oun
d
s
Qt
and
i
v
Rt
, the conservative o
p
timal weight
ed me
asure
m
ent
fused time
-va
r
ying Kalman
predi
cto
r
s
ˆ
|
M
x
tN
t
,
1
N
are given as.
Whe
n
1
N
, the one-ste
p
predi
ctor is given a
s
:
ˆˆ
1|
|
1
MM
M
M
M
x
tt
t
x
t
t
K
t
y
t
(16)
=
MM
tt
K
t
H
t
(17)
1
TT
=|
1
|
1
MM
M
M
K
tt
P
t
t
H
t
H
t
P
t
t
H
t
R
t
(18)
TT
T
1|
|
1
MM
M
M
M
M
M
Pt
t
t
Pt
t
t
t
Q
t
t
K
t
R
t
K
t
(19)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Weig
hted Mea
s
u
r
e
m
ent Fusion
Kalm
an Predi
ctors with
Un
certai
n… (We
n
-jua
n Qi)
4695
With
ˆ
1|
0
M
x
,
1|
0
0
|
0
M
PP
, and
1|
M
Pt
t
satisfies the
Ri
ccati equatio
n.
1
TT
TT
1
|
|1
|1
|1
|1
MM
M
M
M
M
Pt
t
t
P
t
t
P
t
t
H
t
H
t
Pt
t
H
t
R
t
Pt
t
t
t
Q
t
t
(20
)
Whe
n
2
N
, the multi-step p
r
e
d
i
c
tor i
s
given by:
ˆˆ
|,
1
1
|
MM
x
tN
t
t
N
t
x
t
t
,
2
N
(21)
With the defin
ition
,1
2
,
,
n
ti
t
t
i
t
t
I
.
The co
nserva
tive
N
-step pre
d
iction e
r
ror varian
ce
|
M
Pt
N
t
is given by:
T
TT
2
|,
1
1
|
,
1
,1
1
1
,
MM
N
j
Pt
N
t
t
N
t
P
t
t
t
N
t
tN
t
j
t
j
Q
t
j
t
j
t
N
t
j
(22)
Substituting t
he actu
al me
asu
r
em
ent
i
y
t
into the co
nserv
a
tive weighte
d
mea
s
ureme
n
t
fusion
Kalam
n
predi
ctors
(16)
and
(2
1),
we
obtain th
e
actu
al on
e-st
ep a
nd
N
-ste
p time-varying
Kalman predi
ctors.
The act
ual predictio
n errors are give
n a
s
:
ˆ
1|
1
1
|
1
|
MM
M
x
tt
x
t
x
t
t
t
x
t
t
t
w
t
(23)
ˆ
||
|
1
MM
n
M
M
M
M
x
t
t
x
t
xt
t
I
K
t
H
t
xt
t
K
t
v
t
(24)
Substituting (24) into (23)
yields:
1|
|
1
MM
M
M
M
x
t
t
tx
tt
t
w
t
K
t
v
t
(25)
The actu
al weighted me
asurem
ent fuse
d one-st
ep p
r
edictio
n error
varian
ce satisfies the
Lyapun
ov eq
uation.
TT
T
1|
|
1
MM
M
M
M
M
M
Pt
t
t
Pt
t
t
t
Q
t
t
K
t
R
t
K
t
(26)
With the initial value
1|
0
0
|
0
M
PP
, and from (4),
we ha
ve:
1|
0
1
|
0
MM
PP
(27)
Whe
r
e
M
Rt
is the actual vari
an
ce of
M
vt
, and fro
m
(14)
and (1
5) we h
a
ve:
T1
1
T
1
1
T1
1
(
(
)
)
()
()
()
(
(
)
)
Mc
c
c
c
c
R
t
e
R
te
e
R
t
R
t
R
te
e
R
te
(28)
MM
Rt
Rt
(29)
Iterating (1
), we have the
non-re
cu
rsive
formula:
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: 4692 – 4
704
4696
2
,1
1
,
1
1
tN
it
xt
N
t
N
t
x
t
t
N
i
i
w
i
,
2
N
(30)
The act
ual predictio
n errors are give
n a
s
:
2
ˆ
||
,1
1
|
,
1
1
MM
tN
M
it
xt
N
t
x
t
N
x
t
N
t
t
N
tx
tt
t
N
i
i
w
i
,
2
N
(31)
So we have the actu
al
N
ste
p
weig
hted m
easure
m
ent fuse
d pre
d
icti
on error va
ria
n
ce
s.
T
TT
2
|,
1
1
|
,
1
,1
1
1
,
MM
N
j
Pt
N
t
t
N
t
P
t
t
t
N
t
tN
t
j
t
j
Q
t
j
t
j
t
N
t
j
(32)
Theorem 1
.
For multi
s
en
sor un
ce
rtain
system
(1) a
nd
(1
2) with Assu
mption
s 1-3,
the
actual weight
ed mea
s
ure
m
ent fusion t
i
me-varyin
g
Kalman predi
ctors are
rob
u
st in the se
nse
that for all
admissibl
e a
c
tual va
rian
ces
,
i
v
Qt
R
t
and
1|
0
M
P
satis
f
y
i
ng (3)
an
d (4
), for
arbitrary time
t
, we have:
||
MM
P
tN
t
P
tN
t
,
1
N
(33)
And
|
M
P
tN
t
is the mi
nimal upp
er
boun
d of
|
M
P
tN
t
for
all admissibl
e uncertainti
es of
noise varia
n
c
e
s
. We cal
l
the a
c
tual
fused
Kal
m
an p
r
edi
ct
ors a
s
the
robu
st
weig
hted
measurement
fusion Kalma
n
predi
cto
r
s.
Proof
. When
1
N
, defining
1
|
1|
1|
MM
M
P
tt
P
t
t
P
tt
, s
u
btr
a
c
t
in
g (2
6)
fr
o
m
(19
)
yields th
e Lyapun
ov equation.
T
1|
|
1
MM
M
M
M
P
tt
t
P
t
t
t
t
(34)
TT
MM
M
M
M
t
t
Q
t
Q
t
t
K
tR
t
R
tK
t
(35)
Applying (3
), (29
)
and (35
)
yields that
0
M
t
, a
nd from (4) we have:
1|
0
1
|
0
1|
0
0
|
0
0
|
0
0
MM
M
PP
P
P
P
(36)
H
e
nc
e
fr
om (
3
4)
, w
e
h
a
v
e
2|
1
0
M
P
. Applying the mathem
atical ind
u
ctio
n
method yiel
ds
1|
0
M
Pt
t
, for all time
t
, i.e. the inequality (33) holds for
1
N
. When
2
N
, Defining
||
|
MM
M
Pt
N
t
P
t
N
t
Pt
N
t
, subtra
cting (32) fro
m
(22
)
yields:
T
TT
2
|,
1
1
|
1
|
,
1
,1
1
1
1
,
MM
M
N
j
Pt
N
t
t
N
t
P
t
t
Pt
t
t
N
t
t
N
tj
tj
Q
t
j
Q
tj
tj
t
N
tj
(37
)
Applying the
robu
stne
ss of the one-st
ep predi
cto
r
(33) a
nd (3
), we get
|0
M
Pt
N
t
,
therefo
r
e
(33
)
hol
ds fo
r
2
N
. Tak
i
ng
,
ii
vv
Q
t
Q
t
Rt
Rt
and
1|
0
1
|
0
MM
PP
, then
comp
ari
ng (1
9) with (26
)
and
(22)
wi
th
(3
7), we have
||
MM
P
tN
t
P
tN
t
,
1
N
. For
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Weig
hted Mea
s
u
r
e
m
ent Fusion
Kalm
an Predi
ctors with
Un
certai
n… (We
n
-jua
n Qi)
4697
arbitrary othe
r upp
er
bou
n
d
*
|
M
P
tN
t
, we have
*
||
|
MM
M
P
tN
t
P
tN
t
P
tN
t
whic
h
yields that
|
M
P
tN
t
is the minimal u
pper b
oun
d o
f
|
M
P
tN
t
. The pro
o
f is co
mpleted.
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
b
ound
s
Qt
and
i
v
Rt
, similar to the robu
st
weig
hted m
e
asu
r
em
ent fu
sion
time-varying
Kalman pre
d
i
c
tors,
the ro
b
u
st
lo
ca
l tim
e
-varying
Kal
m
an o
n
e
-
ste
p
an
d multi-ste
p
predi
cto
r
s a
r
e
given by:
ˆˆ
1|
|
1
ii
i
i
i
x
tt
t
x
t
t
K
t
y
t
,
1,
,
iL
(38)
=
ii
tt
K
t
H
t
,
T1
=|
1
ii
i
Kt
t
P
t
t
H
t
Q
t
(39)
T
|1
ii
i
Qt
H
t
P
t
t
H
t
R
t
(40)
With the initial value
ˆ
1|
0
,
1|
0
0
|
0
ii
xP
P
, and we have the Ri
ccati equatio
n.
1
TT
TT
1
|
|1
|1
|1
|1
ii
i
M
M
i
Pt
t
t
Pt
t
P
t
t
H
t
Ht
P
t
t
H
t
R
t
HtPt
t
t
t
Q
t
t
(41)
The
con
s
e
r
vative and th
e actu
al on
e-ste
p
p
r
edi
ction e
r
ror v
a
rian
ce
s
sati
sfy the
Lyapun
ov eq
uation
s
.
TT
T
1|
|
1
ii
i
i
i
i
i
Pt
t
t
Pt
t
t
t
Q
t
t
K
t
R
t
K
t
(42)
TT
T
1|
|
1
ii
i
i
i
i
i
Pt
t
t
Pt
t
t
t
Q
t
t
K
t
R
t
K
t
(43)
With the initial values
1|
0
0
|
0
i
PP
,
1|
0
0
|
0
i
PP
.
The co
nserva
tive local opti
m
al time-vary
i
ng Kalman m
u
lti-step p
r
e
d
i
c
tors are give
n by:
ˆˆ
|,
1
1
|
ii
x
tN
t
t
N
t
x
t
t
,
1,
,
iL
,
2
N
(44)
The co
nserva
tive optimal
N
step pre
d
ictio
n
error vari
an
ces
|
i
Pt
N
t
are given
by:
T
TT
2
|,
1
1
|
,
1
,1
1
1
,
ii
N
s
P
t
N
t
tN
t
P
t
t
tN
t
tN
t
s
t
s
Q
t
s
t
s
t
N
t
s
(45)
The act
ual
N
step predi
ction e
rro
r varia
n
ces are give by:
T
TT
2
|,
1
1
|
,
1
,1
1
1
,
ii
N
s
Pt
N
t
t
N
t
P
t
t
t
N
t
t
N
ts
ts
Q
t
s
t
s
t
N
t
s
(46)
Similarly, the
local time
-varying K
a
lma
n
on
e-step
a
nd multi
-
ste
p
pre
d
icto
rs a
r
e al
so
robus
t, i.e.,
1|
1|
ii
Pt
t
P
t
t
,
1,
,
iL
(47)
||
ii
P
t
Nt
P
t
Nt
,
2
N
,
1,
,
iL
(48)
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: 4692 – 4
704
4698
And
|
i
P
tN
t
is the mi
nimal upp
er b
ound of
|
i
P
tN
t
,
1
N
.
4. Robus
t Lo
cal and Fuse
d
Stead
y
-
sta
t
e Kalman Pr
edictor
s
Theorem 2
.
For multisensor un
ce
rtain time-inv
ariant sy
ste
m
(1) an
d (12
)
with
Assu
mption
s 1-3,
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
con
s
tant matrice
s
.
Assum
e
that
the measure
m
ents
i
yt
,
1,
,
iL
are b
ound
e
d
, then the
actual
wei
ght
ed me
asu
r
e
m
ent fusi
on
steady-state
Kalman
predi
cto
r
s a
r
e
given by:
ˆˆ
1|
|
1
ss
MM
M
M
M
x
tt
x
t
t
K
y
t
,
1
N
(49)
1
ˆˆ
|1
|
sN
s
MM
x
tN
t
x
t
t
,
2
N
(50)
=
Mn
M
IK
H
,
1
TT
=
MM
M
M
KH
H
H
R
(51)
TT
T
M
MM
M
M
M
M
QK
R
K
(52)
TT
T
M
MM
M
M
M
M
QK
R
K
(53)
Whe
r
e the su
perscript s de
notes “s
teady-state”, the initial value
ˆ
0|
0
s
M
x
ca
n arbitra
r
ily
be sel
e
ct
e
d
,
i
y
t
are the a
c
tual
measu
r
em
en
ts, and:
T1
1
T
1
()
(
)
Mc
c
c
y
te
R
e
e
R
y
t
(54)
T1
1
()
Mc
Re
R
e
,
T1
1
T
1
1
T1
1
()
()
Mc
c
c
c
c
Re
R
e
e
R
R
R
e
e
R
e
(55)
The
con
s
e
r
vative and a
c
tual
steady
-state
pre
d
iction erro
r varian
ce
s
sati
sfy the
Lyapun
ov eq
uation
s
.
2
TT
11
T
0
N
NN
s
s
MM
s
PN
Q
,
2
N
(56)
2
TT
11
T
0
N
NN
s
s
MM
s
PN
Q
,
2
N
(57)
The act
ual st
eady-state Kalman p
r
edi
ct
ors a
r
e
robu
st, in the sense that:
M
M
,
MM
PN
PN
(58)
And
M
and
M
PN
are the minimal up
per bo
und
s of
M
and
M
PN
, res
p
ec
tively.
Proof.
As
t
, takin
g
the li
mit operation
s
for
(14
)
-(3
3
) with
,,
,
H
Q
and
i
R
ar
e
con
s
tant mat
r
ice
s
yield
s
(49)-(5
8). Ta
ki
ng
QQ
and
ii
vv
RR
, from (55), we hav
e
M
M
RR
, s
o
that from
(5
2
)
a
nd
(53
)
yi
elds
M
M
, hen
ce
from
(5
6)
and
(5
7),
we
hav
e
MM
PN
PN
. If
*
M
or
*
M
PN
is the a
r
bitra
r
y othe
r bou
nd of
M
or
M
PN
, we have
*
M
MM
or
*
MM
M
PN
PN
PN
, which yield
s
that
M
and
M
PN
are minimal. The
proof is
com
p
leted.
Similarly, the actual lo
cal
steady-state Kalman p
r
edi
ct
ors a
r
e give
n by:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Weig
hted Mea
s
u
r
e
m
ent Fusion
Kalm
an Predi
ctors with
Un
certai
n… (We
n
-jua
n Qi)
4699
ˆˆ
1|
|
1
ss
ii
i
i
i
x
tt
x
t
t
K
y
t
,
1,
,
iL
(59)
=
in
i
IK
H
,
1
TT
=
i
ii
i
v
KH
H
H
R
(60)
The co
nserva
tive and actu
al predi
ction
er
ror vari
an
ces satisfy Lyapunov eq
uati
on.
TT
T
ii
i
i
i
i
i
QK
R
K
,
1,
,
iL
(61)
TT
T
ii
i
i
i
i
i
QK
R
K
,
1,
,
iL
(62)
The act
ual st
eady-state fused Kalm
an
multi-ste
p
pre
d
ictor i
s
given
as:
1
ˆˆ
|1
|
sN
s
ii
x
tN
t
x
t
t
,
1,
,
iL
,
1
N
(63)
The con
s
e
r
vative and a
c
tual local ste
ady-state
N
s
t
ep
p
r
ed
ic
tion
e
r
ro
r
var
i
a
n
ce
s
ar
e
given as:
2
TT
11
T
0
N
NN
s
s
ii
s
PN
Q
,
2
N
,
1,
,
iL
(64)
2
TT
11
T
0
N
NN
s
s
ii
s
PN
Q
,
2
N
1,
,
iL
(65)
The
actu
al lo
cal
stea
dy-st
a
te Kalma
n
p
r
edi
ct
ors
(59
)
and
(63) a
r
e
ro
bust
in th
e
se
nse
that for all admissi
ble un
ce
rtainties of
Q
and
i
v
R
sat
i
sf
y
i
ng
,
ii
vv
QQ
R
R
, we have:
ii
,
ii
PN
P
N
,
1,
,
iL
(66)
And
i
and
i
PN
are the minimal up
per bo
und
s of
i
and
i
PN
, respe
c
tively. Hence t
hey are
calle
d the rob
u
st stea
dy-state Kalman p
r
edi
ctors.
Lemma 1
. [16, 17] Con
s
id
er a dynami
c
error sy
stem.
1
tF
t
t
u
t
(67)
Whe
r
e
0
t
,
n
tR
,
n
ut
R
, and
F
t
is u
n
iformly a
s
ymptotically
stable,
i.e., there
exist con
s
tant
s
01
and
0
c
s
u
c
h
that:
,,
0
ti
Ft
i
c
t
i
(68)
Whe
r
e
the
notation
denot
es th
e n
o
rm
of matrix,
,2
1
,
Ft
i
F
t
F
t
F
i
,
n
F
ii
I
. If
ut
is bo
und
e
d
, then
t
is bo
u
nded. If
0
ut
, then
0
t
, as
t
.
Theorem 3
.
Und
e
r th
e
co
ndition
s of T
heorem
2,
th
e ro
bu
st time
-varying
and
steady-
state Kalma
n
local
and
fused o
n
e
-
step a
nd m
u
lti-step
pre
d
ictors h
a
ve
each othe
r the
conve
r
ge
nce in a reali
z
atio
n, such that:
ˆˆ
1|
1|
0
s
ii
xt
t
x
t
t
, as
t
, i.a.r
(69)
ˆˆ
||
0
s
ii
xt
N
t
x
t
N
t
, as
t
, i.a.r ,
2
N
(70)
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: 4692 – 4
704
4700
ˆˆ
1|
1|
0
s
MM
xt
t
x
t
t
, as
t
, i.a.r
(71)
ˆˆ
||
0
s
MM
xt
N
t
xt
N
t
, as
t
, i.a.r ,
2
N
(72)
Whe
r
e the
n
o
tation “i.a.r”
denote
s
the
conve
r
ge
nce
in a re
alizatio
n [17], and
we have
the conve
r
ge
nce of varia
n
c
e
s
.
1|
ii
Pt
t
,
1|
M
M
Pt
t
, as
t
,
1,
,
iL
(73)
|
ii
P
tN
t
P
N
,
|
MM
P
tN
t
P
N
, as
t
,
1,
,
iL
(74)
Proof
. Acco
rding to Assu
mption 3, we
have [18]:
1|
ii
Pt
t
, as
t
,
1,
,
iL
(75)
Then fro
m
(1
9) and
(39
)
, we have:
ii
t
,
ii
Kt
K
,
1|
M
M
Pt
t
as
t
,
1,
,
iL
(76)
Similarly, we
can prove (74) hol
ds, Setting
ii
i
tt
,
ii
i
K
tK
K
t
in
(38
)
, applyin
g
(76
)
yields
0
i
t
0
i
Kt
, as
t
. Subtracting (59
)
fro
m
(38), an
d
defining
ˆˆ
||
s
ii
i
tx
t
N
t
x
t
N
t
, we have:
1
ii
i
i
tt
u
t
(77)
With
ˆ
|1
ii
i
i
i
ut
t
x
t
t
K
t
yt
. Noting
that
i
t
is u
n
iformly asymptot
ically
stable [19], a
n
d
ii
K
ty
t
is boun
ded
, applying Le
mma 1 to (3
8) yield
s
the
boun
dedn
ess of
ˆ
1|
i
x
tt
. Hence we
have
0
i
ut
. Applying Lemm
a
1 to (77), not
ing that
i
is a stable
matrix, so it i
s
also unifor
mly asymptotically stable,
hen
ce
0
i
t
, i.e. th
e conve
r
ge
nce (69)
hold
s
. The co
nverge
nce of (70
)
-(72
) ca
n be pr
ove
d
si
milarly. The p
r
oof is
compl
e
ted.
5. The Accurac
y
Anal
y
s
is
Defini
tion 1
. The t
r
ace
tr
|
P
tN
t
of the u
ppe
r b
ound
|
P
tN
t
of t
he a
c
tual
predi
ction
e
r
ror va
rian
ce
s
|
P
tN
t
for
all a
d
missi
b
le u
n
certaint
ies i
s
call
ed t
he
robu
st
accura
cy
or glo
bal a
c
curacy of a
robu
st Kalm
an predi
ctor,
and
tr
|
P
tN
t
is call
ed as its actual
accuracy. T
h
e smalle
r
tr
|
P
tN
t
or
tr
|
P
tN
t
mean
s the
hi
gher robu
st a
c
cura
cy o
r
a
c
tual
ac
cur
a
cy
.
Th
e ro
bu
st
ac
cu
racy
giv
e
s t
h
e low
e
st
b
o
u
nd of all
po
ssible a
c
tual a
c
curaci
es yield
e
d
from the un
ce
rtainties of no
ise varia
n
ces.
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 predi
ctors is given
by:
||
,
ii
P
tN
t
P
tN
t
1,
,
iL
,
1
N
(78)
||
|
,
MM
i
P
tN
t
P
tN
t
P
tN
t
1,
,
iL
,
1
N
(79)
tr
|
t
r
|
ii
P
tN
t
P
tN
t
,
1,
,
iL
,
1
N
(80)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Rob
u
st Weig
hted Mea
s
u
r
e
m
ent Fusion
Kalm
an Predi
ctors with
Un
certai
n… (We
n
-jua
n Qi)
4701
tr
|
t
r
|
t
r
|
MM
i
P
tN
t
P
tN
t
P
tN
t
,
1,
,
iL
,
1
N
(81)
,
ii
M
M
i
PN
PN
P
N
P
N
PN
,
1,
,
iL
,
1
N
(82)
With the defin
itions
1,
1,
1
,
1
ii
ii
M
M
M
M
PPP
P
.
tr
t
r
ii
PN
PN
,
tr
t
r
tr
MM
i
PN
PN
P
N
,
1,
,
,
iL
1
N
(83)
Proof
. A
c
c
o
rding to the
robu
stne
ss (3
3), (47
)
a
nd
(48
)
, we h
a
ve (78) an
d t
he first
inequ
ality of (79
)
. Since t
he co
nservat
i
ve weighte
d
measure
m
e
n
t fuser i
s
e
quivalent to the
con
s
e
r
vative
centralized f
u
se
r [20], th
e second
in
equality of
(79)
ha
s b
e
e
n
proven i
n
[21].
Takin
g
the
trace
op
eratio
ns fo
r
(78
)
a
nd (79
)
yield
s
the
ineq
uali
t
ies (80
)
an
d
(81
)
. As
t
,
taking the limi
t
operation
s
for (7
8)-(8
1) yi
elds (82
)
and
(83
)
. The pro
o
f is com
p
leted.
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
(84)
2
0
0
0
1
0.
5
,
01
T
T
,
2
H
I
(85)
Whe
r
e
0
0.35
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 th
e po
siti
on an
d velo
ci
ty of target at
time
0
tT
.
wt
,
t
and
i
t
are inde
pen
den
t Gau
ssio
n
white noi
se
s with ze
ro me
an and un
kn
o
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
,
d
i
a
g
(
1
.5
,
2
.5
)
R
,
di
a
g
(
1
,
2
)
R
,
1
di
a
g
(
3
.6
,
2
.
5
)
R
,
1
d
i
a
g
(
3
,1
.
8
)
R
,
2
dia
g
(
8
,
0
.36)
R
,
2
d
i
a
g
(
6
,
0
.2
5)
R
,
3
diag
(
0
.5
,
2
.8
)
R
,
3
di
a
g
(0.
3
8
,
2)
R
,
1,
N
2
N
. 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
ons
of the
predictio
n e
rro
r varian
ce
mat
r
ice
s
and th
ei
r tra
c
e
s
of th
e ro
bu
st
steady-state l
o
cal an
d wei
ghted mea
s
u
r
eme
n
t fusion
Kalman pred
ictors are
sh
own in Tabl
e
1-
Table 3. The
s
e matrices a
n
d
their tra
c
e
s
verify the accura
cy relatio
n
s
(82
)
an
d (8
3).
The tra
c
e
s
of
the co
nserva
tive and a
c
tu
al ro
b
u
st
one
-step
an
d two
-
step
predi
ction e
rro
r
varian
ce
s are
compa
r
ed in
Figure 1 and
Figure 2. We
see that the trace
s
of the local an
d fuse
d
robu
st time
-v
arying
Kalma
n
on
e-step
a
nd two-st
ep
predi
cto
r
s qui
ckly
co
nverg
e
to th
ese of
the
corre
s
p
ondin
g
stea
dy-stat
e
Kalman p
r
edicto
r
s,
wh
i
c
h
sho
w
the
robu
st a
c
curacy rel
a
tion
s
(80
)
,
(81
)
and (83
)
hold.
Table 1. The
Con
s
e
r
vative and Actual A
c
cura
cy Co
m
pari
s
on of On
e-ste
p
Predi
ction Erro
r
V
a
rian
ce
s Ma
t
r
ice
s
i
and
i
,
1,
2
,
3
,
iM
1
2
3
M
1.4
931
0
.
653
8
0
.
653
8
0
.631
4
1.799
5
0.6200
0.6200
0.5833
0.85
58
0.4877
0.4877
0.5592
0
.
73
15
0.
409
8
0
.
40
98
0.
499
5
1
2
3
M
1.1
667
0.5
123
0
.
51
23
0.4
989
1.36
98
0
.
48
36
0
.
48
36
0.4
617
0
.
620
2
0
.367
2
0
.
367
2
0
.434
6
0
.
536
5
0
.3
134
0
.
31
34
0.3
922
Evaluation Warning : The document was created with Spire.PDF for Python.