TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 14, No. 3, June 20
15, pp. 543 ~ 5
5
6
DOI: 10.115
9
1
/telkomni
ka.
v
14i3.806
9
543
Re
cei
v
ed Fe
brua
ry 24, 20
15; Re
vised
Ma
y 13, 20
15
; Accepte
d
May 28, 20
15
The Adaptive F
e
deral Unscented Particle Filt
er
Algorithm with Applications in All-Attitude Integrated
BDS/INS Navigation System
Ping Luo, Li-jie Yuan*, Liang-xue Hua
ng, Xian-fei
Li
Ke
y
Lab
orator
y of Industrial In
ternet of T
h
ings& Net
w
o
r
k Co
ntrol, MOE, Chong
qin
g
Un
iver
sit
y
of Posts an
d
T
e
lecommunic
a
tions, Ch
on
gq
ing, Ch
ina, 4
0
0
065
*Corres
p
o
ndi
n
g
suthor, e-mai
l
:
y
u
anl
iji
e20
15
@16
3
.com
A
b
st
r
a
ct
In order to
i
m
pr
ove th
e at
titude acc
u
rac
y
, the t
hesis
establ
ishes t
h
e al
l-attitud
e
i
n
tegrate
d
BDS/INS nav
ig
ation
n
onl
ine
a
r
system
mo
del
bas
ed
on
th
e
positi
on, v
e
loc
i
ty, attitude
by
add
ing
the
BD
S
’
s
attitude meas
u
r
ement infor
m
ation
int
o
th
e me
asur
e
m
ent
equ
atio
n of
th
e trad
ition
a
l
B
D
S/INS inte
gr
ated
navigation nonlinea
r system
m
o
del. Considering
the pr
oblem that
the dy
nam
i
c
nav
igation syst
em m
o
del
is
difficult to
acc
u
rately
descr
ib
e the c
o
mp
lex
navi
gat
i
on
en
viron
m
e
n
t, the
thesis
i
m
prov
es the
dyn
a
m
i
c
character
i
stics of the infor
m
ati
on di
stri
buti
on
of the feder
al
fi
lter alg
o
rith
m w
h
ich co
ul
d time
ly chan
ge
base
d
on the
ei
genv
a
l
ues r
a
tio of
ea
ch subsyste
m
’s error
vari
anc
e matrix. Then,
the
ad
aptiv
e feder
al u
n
sce
nte
d
particl
e filter (AF
U
PF
) is propose
d
. T
he si
mu
lati
on sh
ow
s that the pro
pose
d
al
gorith
m
co
uld effecti
v
el
y
w
eaken th
e i
m
p
a
ct o
n
the
system
accura
cy of the
inac
curate hig
h
-dy
n
a
m
ic mod
e
l, and i
m
pr
ove
t
h
e
ada
ptab
ility, the fault tolera
nc
e and th
e a
ccu
racy, especi
a
ll
y the attitude a
ccuracy.
Ke
y
w
ords
:
BDS/INS integ
r
ated nav
ig
ati
on syste
m
, attit
ude u
pdate,
adaptiv
e info
rmati
on d
i
strib
u
tion
,
AFUPF
Copy
right
©
2015 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 BDS/INS integrate
d
navigation
system ha
s b
e
com
e
a
n
i
m
porta
nt dev
elopme
n
t
dire
ction of the navigat
io
n
system, which could
overcome th
e sh
ortco
m
ing
s
of
each
other,
and
have hig
her navigation
accuracy
th
a
n
ea
ch
syst
em to work
alone.
With
the continu
ous
developm
ent
of satellite att
i
tude dete
r
mi
nation te
chn
o
l
ogy, BDS no
t only provid
es the
po
sition
informatio
n a
nd velo
city i
n
formatio
n, b
u
t also give
s accu
ra
cy at
titude inform
ation [1]
Error!
Refere
nce s
o
urce not
found.
for the carri
er. The f
u
lly integr
ate
d
BDS/INS navigation sy
stem
whi
c
h
refle
c
ted the
BDS
attitude information in
the
integrated
n
a
vigation
system mea
s
u
r
e
m
ent
model
ha
s b
e
com
e
the f
easi
b
le
soluti
ons to
imp
r
o
v
e the a
c
curacy of attitud
e
, whi
c
h fu
rther
increa
se
s the
obse
r
vability of the model [2].
In pra
c
tical
appli
c
ation
s
,
the integrat
ed navi
gatio
n system typically u
s
e
s
indire
ctly
kalma
n
filter (KF) algo
rithm to simplif
y proce
d
u
r
e
s
and improv
e perfo
rman
ce, be
cau
s
e
the
system’
s
e
rro
r is
studie
d
a
nd mod
e
led i
n
advan
c
e
. A
t
pres
ent, the domes
tic
and
fo
r
e
ign
e
x
pe
r
t
s
also
study the appli
c
ation
s
of the indirectly KF
in the BDS/INS integrate
d
na
vigation syst
em.
Ho
wever, the
BDS’s pse
u
d
o
rang
e-e
r
ror
is wo
rst
and t
he model isn’
t accu
rate so
that it does no
t
meet the ba
sic p
r
emi
s
e
of the i
ndirect KF, thus affects the filt
er accuracy of
the INS’s data
integratio
n.
The the
s
i
s
gets th
e no
nlinea
r m
o
d
e
l of the
all
-
attitude inte
grated
BDS/
INS
navigation
system [3-5] by the
dire
ct m
e
thod [6]: est
ablishe
s
the
state equ
atio
n based o
n
the
mech
ani
cs
chore
o
g
r
aphy
equation
s
and attitude
erro
r equ
ation in INS; establi
s
he
s the
measurement
equatio
n ba
sed
on the
p
o
sition m
e
ss
age, velo
city messag
e of t
he BDS a
nd
the
attitude error
messag
e of the BDS/INS.
The i
n
form
ation fu
sion
h
a
s t
w
o
ways:
centralize
d
filter
and
distrib
u
ted
filter. Th
e
distrib
u
ted fe
deral
kalman
filter (FKF
) based
o
n
t
he KF a
nd
the inform
ation di
stributio
n
techn
o
logy h
a
s be
en ta
ke
n se
riou
sly b
e
ca
use of
the parallel dat
a pro
c
e
s
sing,
flexible desi
gn,
small am
ount
of calcul
ation
,
better fault-tolera
nt perfo
rmance.
Ho
wever, the
use of the su
bfilter of the
FKF must mee
t
certain ide
a
l
conditio
n
s:
1. Dynamic m
odel of the sy
stem is a
c
curate and line
a
r;
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 14, No. 3, June 20
15 : 543 – 55
6
544
2. Noise is un
related to the
white noi
se
a
nd ha
s the kn
own
statistica
l prope
rtie
s;
3. Ch
oo
se th
e ap
pro
p
riate
initial value
of the
state v
a
riabl
es an
d
the noi
se
variance mat
r
ix i
n
orde
r to en
su
re the unbi
ased estimatio
n
re
sult
s and p
r
event the filter divergen
ce
.
Therefore, th
e un
scented
particl
e filter
(UPF
) [7, 8] is u
s
ed in
su
bfilter and th
e federal
unsce
nted pa
rticle filter (F
UPF) [9] is
constitu
ted in t
he thesi
s
. Th
e UPF is the
non-li
nea
r no
n-
gau
ssi
an filter algo
rithm an
d has b
een u
s
ed in ma
ny
fields
su
ch a
s
navigation
system and ta
rg
et
tracking. Th
e
UPF dire
ctl
y
estimates t
he pa
ra
mete
rs of the n
a
v
igation syst
em, avoide
s the
lineari
z
atio
n of the nonlin
ear
state equ
ation and e
n
s
ures th
e hig
h
pre
c
i
s
ion o
f
the navigation
sy
st
em.
After analy
s
is, it is not
difficult to find
th
at the info
rm
ation di
strib
u
tion ha
s
alwa
ys be
en
the key to d
e
sig
n
ing a
n
d
achievin
g th
e FKF, the
d
i
fferent value
s
have diffe
rent federal filter
cha
r
a
c
teri
stics. In a
c
tual
appli
c
ation, d
ue to
th
e
ch
ange
of the
syst
em
environment
or ot
her
factors, the
perfo
rman
ce
of each su
bsyste
m of
the integ
r
ated
navigation
system may
also
cha
nge. The
pred
efined inf
o
rmatio
n distribution
co
efficient rem
a
ini
ng unchan
ge
d durin
g filtering
doe
s not accurately refle
c
t
the actual sit
uation an
d affects the e
s
ti
mated ac
cu
ra
cy
of
t
he sy
st
em.
Mean
while, t
a
kin
g
into a
c
count that
each comp
o
nent of the
state vecto
r
has th
e diffe
rent
observability and co
nverg
ence, t
he feedba
ck con
s
i
derin
g the st
ate variable
s
as a whol
e
i
s
extremely unreasona
ble.
B
e
cau
s
e
t
h
e
v
a
rian
ce
m
a
t
r
ix
of
ea
ch
su
bf
ilter of
the FKF
could
refle
c
t the
state
estimation a
c
curacy in re
al
-time. Based
on the
point, a new ad
apti
v
e federal un
scented pa
rti
c
le
filter (AFUPF
) is propo
se
d
in the thesi
s
, w
hose information dist
rib
u
tion co
effici
en co
uld
timely
cha
ng ba
se
d
on the eige
nval
ue
s ratio
of each
sub
s
ystem’
s error va
rian
ce m
a
trix
a
nd bett
e
r
satisf
y the d
y
nam
ic state
chan
ge
s u
nder
hi
gh
-d
ynam
ic envi
r
o
n
m
ent so that
improve the
estimation a
c
curacy of the
federal filter.
The rest
of th
e thesi
s
i
s
o
r
gani
zed
as fo
llows
: In Sect
ion 2, the
est
ablishment
of the all-
attitude integ
r
ated BDS/INS navigation
system m
ode
l is de
scribe
d. In Section
3, the AFUPF
algorith
m
is d
i
scusse
d. In Section 4, experim
ent
s results and an
al
ysis are put. Finally, in Section
5, con
c
lu
sion
s of the wo
rk
are ma
de.
2. Establishment of the All-atti
tude In
te
grate
d
BDS/INS Nav
i
gation Sy
stem
Model
2.1. The Co
mposition of the BDS/INS Integra
t
ed Nav
i
gation Sy
stem
In the thesi
s
, the com
p
o
s
ition of the BDS/IN
S integra
t
ed navigatio
n system i
s
shown in
Figure 1. Th
e
informatio
n fusio
n
sy
stem
has t
w
o
navigation
sen
s
o
r
s which a
r
e B
D
S an
d INS
as
the com
m
on
referen
c
e system, whi
c
h
would
co
ns
t
i
tute three
subfilters, n
a
m
ely: the po
sition
subfilter, the
velocity subfil
ter and the a
ttitude s
ubfilter [10-1
2
] an
d the main filter. The three
subfilters
use
the UPF
alg
o
rithm,
an
d t
heir i
n
formati
on is se
nt to
the main filt
er to h
a
ve the
fusion
in
ord
e
r
to
get the
n
a
vigation
parameters
and
the be
st attitu
de e
s
timation
as the
output
of
the navigatio
n system.
1
,
g
g
PX
2
,
g
g
PX
3
,
g
g
PX
11
,X
P
33
,X
P
22
,X
P
,
m
m
PX
,
gg
PX
,
f
a
Figure 1. Structure of all
-
attitude
integrated BDS/INS navigation
system
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
The Adapti
v
e
Federal Un
scente
d
Particl
e
Filter
Algori
t
hm
with Application
s
in… (Ping Luo
)
545
2.2. The Mod
e
l of the BDS/INS Nav
i
gation Sy
stem
The all-attitude integrated
BDS/INS navigati
on syst
em model e
s
tablish
ed by the dire
ct
method is a
d
opted in the thesi
s
, whi
c
h
con
s
id
ers
the
navigation p
a
ram
e
ters of the INS as the
state vari
able
,
establi
s
he
s
the mea
s
u
r
e
m
ent equ
atio
n ba
sed
on t
he po
sition m
e
ssag
e, velo
city
messag
e of the BDS an
d the attitude error
me
ssage of the BDS/INS, and deal
s with
the
modified AF
UPF in o
r
de
r to get the
navigati
on
para
m
eter
e
s
timate
s a
s
the output of
the
BDS/INS integrated n
a
viga
tion system.
2.2.1. The State Equa
tion
of the Integr
ated
Nav
i
gation Sy
stem
The thesi
s
con
s
id
ers the navigation
c
oordinate
system to
be the ge
ogra
phi
cal
coo
r
din
a
te system, estab
lishe
s the
state
equatio
n
base
d
on t
he me
cha
n
ics ch
oreog
ra
phy
equatio
ns
an
d the INS’
s a
ttitude error
e
quation
co
ns
i
derin
g that th
e existen
c
e
o
f
the inclin
ation
error of the platform coul
d
result in oth
e
r
axial gene
rating spe
c
ific force com
p
onent
s acco
rding
to the literatu
r
e [6], and e
s
tabli
s
he
s th
e ran
dom
co
nstant mo
del
for the accel
e
rom
e
ter’
s of
fset
and gyro
scop
e’s offset. Th
erefo
r
e, the state variable
s
used in
simul
a
tion are:
T
en
u
e
n
u
e
n
u
e
n
u
x
h
a
a
add
d
(1)
Whe
r
e, the
st
ates a
r
e th
e
velocity, posit
ion, pla
tform
attitude error,
accele
rom
e
ter’s bia
s
e
s
a
nd
gyroscope’
s range drift.
The state e
q
u
a
tion of the BDS/INS
integrated navigati
on syste
m
is:
,
x
tF
x
t
u
t
(2)
Discretize fo
rmula (2
) with
the fourth-ord
er Ru
nge
-Kut
ta method:
k
k
k
u
x
f
x
,
1
`
(3)
Whe
r
e,
f
is th
e no
nlinea
r f
unctio
n
;
k
u
is the n
o
ise
vari
able
s
of
syst
em; varia
n
ce
matrix of
the system n
o
ise i
s
k
Q
.
2.2.2. The Observ
a
tion Equation o
f
th
e Integra
t
ed
Nav
i
gation Sy
stem
The velo
city for the ea
st (x), north (y
),
and altitude
(z), po
sition
s
of the BDS a
nd the
differen
c
e of
the attitude b
e
twee
n the INS and B
D
S
are
sele
cted
as the o
b
se
rved varia
b
le
s,
namely:
[]
T
kE
B
N
B
U
B
B
B
B
yl
h
The an
gle-error of the
st
at
e equatio
n
of the integrated
navig
ation syste
m
is the
mathemati
c
al
platform mi
salig
nment a
ngle wh
ich descri
b
e
s
the relation
shi
p
betwe
en the
mathemati
c
al
platform co
ordin
a
te syst
em
an
d the
geog
rap
h
ical
coo
r
di
nate
system, b
u
t
the
differen
c
e
of
the
observ
ed attitude
betwe
en th
e
INS a
nd B
D
S d
e
scri
be
s the
relatio
n
shi
p
betwe
en
the
carrie
r coo
r
di
nate system and
th
e
ge
og
raphi
c co
ordi
nate system. In
ord
e
r
to
m
a
ke
them unified i
n
the phy
sica
l sen
s
e, the
r
e sh
oul
d
be
essentially th
e relatio
n
ship
betwe
en the
m
,
namely
bb
t
pt
P
CC
C
. Where,
p
is the pl
atform coo
r
di
nate sy
stem,
b
is the
ca
rri
e
r
coordinate
sy
st
em,
t
is th
e geo
graphi
cal co
ordinate
system. Ba
se
d
on th
e relati
onship, we
could o
b
tain a
n
attitude subfil
ter [11, 12] of the federal
filter, who
s
e ob
servatio
n mat
r
ix is:
3
cos
c
os
c
o
s
s
i
n
0
1
sin
c
os
0
cos
sin
s
in
sin
c
os
c
o
s
h
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 14, No. 3, June 20
15 : 543 – 55
6
546
In the way, the attitude m
easure
m
ent i
n
fo
rmatio
n which
com
e
s f
r
om the multi-antenn
a
Beidou recei
v
er on the
basi
s
of carrier p
h
a
s
e d
i
fference technolo
g
y is e
n
tered i
n
to the
observation
e
quation of th
e tradition
al BDS/IN
S integrated
navig
ation nonli
n
e
a
r sy
stem m
odel,
the purp
o
se of which i
s
to corre
c
t the mathem
atical model error of the BDS/INS integrated
navigation
system and im
p
r
ov
e the attitude accu
ra
cy.
The ob
se
rvation equ
ation o
f
the in
tegrate
d
navigation
system i
s
:
13
3
3
3
3
6
11
1
3
3
2
3
3
3
6
1
1
33
33
3
3
6
000
,0
0
0
00
0
kk
k
k
k
h
yh
x
h
x
h
(5)
Whe
r
e,
1
(
111
)
hd
i
a
g
;
2
11
1
hd
i
a
g
;
1
k
y
is the ob
se
rved variabl
e;
h
is linear
function;
k
x
is the state vector value at the
k
time;
1
k
v
is obse
r
vation
noise, the varian
ce
matrix of the measurement
noise i
s
k
R
.
3. The AF
UPF Algorithm
3.1. The Uns
cented Parti
c
le Filter
In the B
D
S/INS inte
grate
d
navig
ation
system, th
e p
a
rticle
filter (PF) [13]
coul
d solve th
e
nonlin
ear
no
ngau
ssian p
r
oblem. Th
e
PF relie
s on
the impo
rtan
ce
sampli
ng
and the
pro
p
o
sal
distrib
u
tion
which
could
ap
proximate
the
po
sterio
r
dist
ribution
rea
s
o
nably. Becau
s
e it i
s
ha
rd t
o
desi
gn
su
ch
prop
osal di
stribution
so tha
t
the PF
gen
erally
sele
cte
s
the p
r
io
ri di
stributio
n a
s
t
he
prop
osal di
stribution,
whi
c
h woul
d ign
o
r
e the ob
se
rvation inform
ation of the
curre
n
t mom
ent,
and ma
ke th
e state e
s
tim
a
tion serio
u
sl
y depen
d on
the syste
m
’
s
mod
e
l. If the mod
e
l is
not
accurate o
r
t
he me
asure
m
ent noi
se
in
cre
a
ses s
udd
enly, the way co
uld n
o
t effectively re
pre
s
en
t
the re
al di
stri
bution of th
e
pro
bability d
ensity
fun
c
tio
n
. Mean
while
, unde
r the d
i
stributio
n, th
e
weig
hts a
r
e
calcul
ated
with
out co
nsi
d
e
r
ing the i
n
fluen
ce of th
e mo
del noi
se,
wh
ich
woul
d aff
e
ct
the filtering a
c
cura
cy.
Therefore,
R. Merwe
pro
poses the
UPF
[7] which
uses the
UKF app
roxim
a
tion to
gene
rate the
propo
sal di
stribution for the PF. Bu
t the UPF doe
s not satisfy the fault-tolerant
requi
rem
ent of BDS/INS system.
3.2. The Principle of the
Federal
Kal
m
an Filter
The fed
e
ral kalman filte
r
[1
4] ba
sed
on t
he the
o
ry
of t
he d
e
centrali
zed
filter i
s
a
kind
of
two-stage
filter stru
ctu
r
e, who
s
e
ba
si
c i
dea i
s
to
pa
ra
llel de
al with
each
subfilter and
obtai
n th
eir
estimation
i
x
and varian
ce
matrix
1,
,
i
Pi
n
n
, then sen
d
them int
o
the main filter to get
the peri
odi
c fusio
n
processing a
nd obt
ain the glob
a
l
optimal esti
mation
g
x
and the varian
ce
matrix
g
P
, and t
hen
se
nd th
e
g
x
and
the
enla
r
ged va
rian
ce
matrix
1
ig
P
into ea
ch su
bfilter
to
reset each su
bfilter, that is:
1
1
1
1
ii
g
ii
g
ig
n
i
i
P
kP
k
Qk
Q
k
xk
x
k
(6)
Each
subfilt
er corre
c
tes the time and the ob
se
rvation
1
i
yk
and gets the
estimation
1
i
xk
and the varian
ce matrix
1
i
Pk
. Accordi
ng to the following fo
rmula, we
can o
b
tain th
e global o
p
timal estimatio
n
and
the vari
ance matrix o
f
the main filter.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
The Adapti
v
e
Federal Un
scente
d
Particl
e
Filter
Algori
t
hm
with Application
s
in… (Ping Luo
)
547
11
1
1
1
11
11
k
+
1
1
n
gi
i
n
gg
i
i
i
Pk
P
k
xk
P
k
P
x
k
(7)
Whe
r
e,
x
k
is th
e state va
riab
les of
system
;
1
yk
is the o
b
se
rved vari
able;
Pk
is the
mean squa
re
erro
r matrix (MSE) for the
k
time;
Qk
is the noise matrix for the
k
time;
i
is
the informatio
n distrib
u
tion
coeffici
ent;
i
mean
s the
ith
subfilter;
g
means
the main filter.
3.3. The Info
rmation Dis
t
ribution Me
t
hod of th
e F
e
deral Filter
Algorithm
The
key to
resea
r
ch an
d de
sign
the
federal
filter is to
determine the i
n
formatio
n
distrib
u
tion coefficient [14
,
15] which
dire
ctly
affects the pre
c
i
s
ion and the
fault toleran
c
e
perfo
rman
ce.
After
a
lot o
f
simulation, the global
estimation of th
e federal filter has the be
st
accuracy whe
n
the informat
ion distri
butio
n coeffici
ent is
0,
1
,
1
,
2,
,
mi
nn
i
n
n
.
But in the
a
c
tual
high
d
y
namic navi
gation
enviro
n
ment, the
perfo
rman
ce
and
the
estimated
qu
ality of each
subfilter is
co
nstantly c
han
ging. Th
erefo
r
e, it is h
ope
d that the ov
eral
l
perfo
rman
ce
of the filter can alway
s
b
e
clo
s
e
to the optimal su
bsyste
m, in other words,
the
informatio
n
d
i
stributio
n co
efficient can
follow
th
e
pe
rforma
nce of
the
subfilter to ma
ke tim
e
ly
cha
nge
s [15].
Acco
rdi
ng to
the inform
atio
n dist
ributio
n
way,
the e
ssence of the
feder
al filter
al
gorithm
[14] is to magnify the m
ean squa
re
error mat
r
ix (MSE)
1
i
times, and a
ssi
gn t
hem to the
different su
bfilters. The KF
could autom
atically
make
the use of different weig
hts acco
rdin
g to
the me
rits
of
the qu
ality of inform
ation,
the hig
her th
e information
quality of th
e subfilter i
s
,
the
smalle
r the in
formation di
st
ribution
coeffi
cient is,
an
d then the hig
h
e
r
the utilizatio
n of the fusio
n
informatio
n of the subfilter i
s
.
We
coul
d det
ermin
e
the inf
o
rmatio
n dist
ributi
on coefficient of each subfilter a
c
cording to
the varia
n
ce
matrix which
contai
ns the
estimate
d
e
r
ror m
e
ssa
ge.
Acco
rdi
ng to
this m
e
thod, t
he
highe
r the
preci
s
ion
of the
filter is, the
smaller th
e inf
o
rmatio
n di
stribution
coeffi
cient i
s
, whi
c
h is
equivalent to
expandin
g
the role
of se
nso
r
with
hi
g
her a
c
curacy
. The adapti
v
e federal filter
algorith
m
pro
posed in the thesi
s
is a
s
fol
l
ows:
Schu
r de
com
positio
n of
i
P
.
T
ii
PU
U
(8)
Therefore, th
e informatio
n di
strib
u
tion coefficient of the
i
subfilter i
s
:
1
,,
1
2
12
11
1
,
,n
n
;
1
,
,m
m
,,
,
nn
ij
i
ij
i
j
i
i
ii
i
i
m
m
im
m
ij
di
a
g
(9)
Whe
r
e,
,1
,
2
,
,,
,
ii
i
i
m
m
diag
,
,1
,
2
,
,,
,
ii
i
m
m
is
the eig
enval
ues of
i
P
;
U
is the
unitary matrix;
nn
is the total
numbe
r of the su
bfilters,
mm
is the dime
nsio
n of the varian
ce
matrix of the subfilter,
i
is the assig
n
me
nt matrix, which is d
e
term
ined by the prop
ortio
n
of
each re
spe
c
ti
ve comp
onen
t
,
ij
on the diago
nal of the varian
ce ma
t
r
ix of each
subf
ilter to the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 14, No. 3, June 20
15 : 543 – 55
6
548
sum of e
a
ch
respe
c
tive comp
one
nt
,
1
nn
ij
i
on
the diago
na
l of the variance matrix
of each
subfilter,
whi
c
h i
s
used t
o
ma
ke th
e
j
compon
ent of
the
i
subfilter
have its o
w
n
inform
atio
n
distrib
u
tion coefficient
ij
, avoid treatin
g
the feature
s
,
the estim
a
tion accu
ra
cy and the
conve
r
ge
nce
velocity of
each
comp
o
nent in
the
same
way, there
b
y am
pli
f
ying the
sa
me
multiples
so a
s
to re
sult in the pathol
ogi
ca
l ca
se
s, and
improve calculation a
c
curacy.
Therefore:
1
11
1
2
1
1
1
21
2
2
2
1
2
1
12
11
1
11
1
1
1
2
11
11
1
22
1
2
2
2
2
2
11
1
12
gg
g
m
m
i
gg
g
m
m
i
ii
g
g
m
m
g
m
m
g
mmmm
im
m
ig
ig
ig
m
m
ig
ig
ig
m
m
im
m
g
m
m
im
m
g
m
m
im
m
g
PP
P
PP
P
Pk
P
k
PP
P
PP
P
PP
P
PP
P
mmm
m
(10)
Whe
r
e, kn
own that
11
12
ii
and
=
T
gg
Pk
P
k
, we coul
d kn
ow
T
ii
Pk
P
k
, whic
h
mean
s that the varian
ce
matrix doe
s not satisf
y the requi
rem
e
n
t
of the symmetry, which coul
d
destroy the converg
e
n
c
e
and sta
b
ilit
y of the filter, resultin
g in lo
wer e
s
timatio
n
accuracy, e
v
e
n
diverge
n
t, losing the true
meanin
g
of t
he alg
o
rithm.
Therefore, in
orde
r to e
n
sure the
sym
m
etry
of the varian
ce matrix of
the
subfilte
r
of the FKF, the pap
er
imp
r
oves i
n
form
ation dist
ribut
ion
method de
scribed ab
ove.
Firs
t, s
p
lit the information dis
t
ribution c
o
effic
i
ent. Set up:
1
2
i
i
i
im
m
B
(11)
Therefore:
11
1
2
22
ii
i
i
ii
ii
i
im
m
im
m
i
m
m
BB
(12)
Get the new f
o
rmul
a, that is:
11
11
1
22
=
ii
ii
ii
g
i
g
im
m
i
m
m
Pk
B
P
k
B
Pk
(13)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
The Adapti
v
e
Federal Un
scente
d
Particl
e
Filter
Algori
t
hm
with Application
s
in… (Ping Luo
)
549
Thus:
1
11
1
11
1
1
12
1
2
1
1
21
2
1
2
2
2
2
2
2
11
2
2
ii
i
g
i
gi
i
g
i
i
g
m
m
i
i
m
m
gi
i
g
i
i
g
m
m
i
i
m
m
g
m
m
i
mm
i
g
m
m
i
m
m
i
g
mmm
m
i
mm
i
m
m
Pk
P
k
B
P
k
B
PP
P
PP
P
PP
P
(14)
Whe
r
e,
T
ii
Pk
P
k
, which mean
s th
a
t
the varian
ce matrix satisfies th
e req
u
irem
ent of
symmetry. T
herefo
r
,
we
can g
e
t the
a
daptive info
rmation di
stri
b
u
tion
coe
ffici
ent expresse
d in
vec
t
or form. With the adaptive alloc
a
tion s
t
rategy, we could
re
set the state e
s
timation
i
x
k
,
the varian
ce
matrix
i
Pk
, and the varian
ce m
a
trix of the proce
s
s noi
se
i
Qk
for ea
ch subfil
ter
in accordan
ce with the formula (1
5).
11
11
1
1
ii
g
i
ii
g
i
ig
n
i
i
Pk
Pk
Qk
Qk
xk
x
k
(15)
3.4. The Improv
ed Adaptiv
e
Federal UPF Algorith
m
In this work,
the AFUPF is pro
p
o
s
ed to ov
ercome
the FKF’s informatio
n dist
ribution
limitation an
d
sati
sfy the fa
ult-tolerant re
quire
ment of
the BDS/INS
system. Fi
gure 1
sho
w
s th
e
stru
cture of t
he AF
UPF a
l
gorithm. Ba
sed o
n
th
e
al
l-attitude inte
grated
BDS/I
N
S navig
atio
n
system comp
leted with the
nonlinea
r st
ate E
quation
(3) a
nd linea
r obse
r
vation
equatio
n (5)
o
f
the all-attitud
e
integrate
d
BDS/INS navigati
on syste
m
, the impro
v
ed AFUPF algorith
m
can
be
adopte
d
to so
lve the proble
m
of informati
on fu
sio
n
, whi
c
h could b
e
concl
ude
d as f
o
llows:
(1) Initiali
ze
the state
pa
ramete
rs:
00
0
0
,,
,
g
gg
n
xP
Q
,and dist
ribute t
hem for ea
ch su
bfilter
according to the formul
a (6
).
For
1,
,
,
nn
n
For
1,
,
N
,
i
Sample the
i
th particl
e fro
m
the prior
0
p
x
in the
n
th subfilte
r
and set up:
0|0
00
00
ˆ
[0
0
]
,
0
0
00
k
TT
kk
k
P
xx
P
Q
R
(16)
Comp
ute
21
x
n
sig
m
a points a
n
d
their wei
ght
s:
0|
0
00
0
|
0
,
x
xP
P
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 14, No. 3, June 20
15 : 543 – 55
6
550
00
0
0
0
,,
ii
i
kx
x
j
j
xx
x
n
P
x
n
P
(17)
0
2
0
,0
1,
0
12
,
1
,
2
,
2
1
m
x
c
x
cm
jj
j
x
Wj
n
Wj
n
WW
W
n
j
n
(18)
Whe
r
e,
nn
is the numbe
r of the su
bfilters,
and
N
is the numbe
r of the sampl
ed pa
rti
c
le
s,
x
n
is the dimen
s
i
on of state vector.
(2)
Cal
c
ulate
d
usin
g the UPF [9] for each subfilter at t
he
TT
k
,
,
1
time.
Importan
c
e
sampling
step
of the UPF in the
n
th subfilte
r
at the
TT
k
,
,
1
time.
For
1,
,
,
kT
T
For
1,
,
,
nn
n
For
1,
,
N
,
i
update the parti
cle
s
with the UKF [16-18] in the
n
th s
ubfilter in time
k
.
(a) Calculati
ng sigm
a poi
nts
11
1
1
1
1
ˆˆ
ˆ
,,
ii
i
i
i
i
kk
k
k
k
k
j
j
xx
x
n
P
x
n
P
(19)
(b) Updatin
g time
,1
,
1
ii
jk
k
j
k
xf
x
(20)
2
1
,1
0
2
11
|
1
,1
,1
1
0
n
i
mi
kk
jj
k
k
j
n
T
ii
ci
i
kk
kk
kk
j
j
k
k
j
k
k
k
j
xW
x
PW
x
x
x
x
Q
(21)
,1
,
|
1
ii
jk
k
j
k
k
zh
x
(22)
2
1
,1
0
n
i
mi
kk
j
jk
k
j
zW
z
(23)
(c)
Upd
a
ting
measurement
with latest ob
servatio
n
2
11
,1
,1
0
2
1
1
,1
,1
0
kk
kk
n
T
ii
ci
i
kk
kk
z
zj
j
k
k
j
k
k
k
j
n
T
ii
ci
i
kk
kk
xz
j
j
k
k
j
k
k
j
P
W
zz
zz
R
PW
x
x
z
z
(24)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
The Adapti
v
e
Federal Un
scente
d
Particl
e
Filter
Algori
t
hm
with Application
s
in… (Ping Luo
)
551
1
k
k
k
k
z
z
z
x
k
P
P
K
(25)
1
1
ˆ
ii
i
kk
kk
kk
k
xx
K
z
z
(26)
T
k
z
z
k
i
k
k
i
k
K
P
K
P
P
k
k
1
|
(27)
(d) Sam
p
ling
Sample:
1
ˆ
,,
ii
ii
kk
k
k
k
qx
x
z
N
x
P
(28)
For
1,
,
N
,
i
evaluate and no
rmali
z
e the importa
nce
weig
hts:
1
1
0
1
,
,
ii
i
n
kk
k
k
ii
ii
i
kk
kk
k
ii
i
kk
k
pz
x
p
x
x
ww
ww
w
qx
x
z
(29)
(3)
Re
sampli
ng of the subf
ilter
(a) Evaluatin
g
eff
N
2
1
1
eff
N
i
k
i
Nw
(30)
(b) Re
sidu
al
resam
p
ling wh
en
ef
f
t
h
NN
, for
1,
,
iN
, s
e
t:
11
,,
1
N
N
ii
i
kk
k
ii
xx
N
(31)
(c) Subfilter’
s
output
The po
steri
o
r distributio
n is approxim
ate
d
as follo
ws.
i
k
k
N
i
i
k
k
k
x
x
z
x
p
1
:
1
(32)
One obtai
ne
s the following
estimate a
s
:
1
ˆ
N
ii
kk
k
i
x
x
(33)
(4) Ada
p
tive informatio
n di
stributio
n mat
r
ix and total output
Fusin
g
re
sult
s of all subfilters to gen
era
t
e the adaptive information
distrib
u
tion m
a
trix
of each
subfil
ter acco
rdin
g to the formul
a (9) a
nd the
final total estimation re
sult
according to
the formula (7).
4. Res Simulation res
u
lts
In this se
ctio
n the nume
r
i
c
al results o
b
tai
ned by a
pplying si
mul
a
ted data to
the filters
are
presente
d
. The
r
e
are
thre
e different cases u
s
ed in the t
e
s
t
ing in
order to verify
the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 14, No. 3, June 20
15 : 543 – 55
6
552
perfo
rman
ce
of the imp
r
oved AF
UPF algo
rithm
with the
a
ll-attitude int
egrate
d
BDS/INS
navigation
system, whi
c
h a
r
e:
Ca
se 1: th
e
traditional
FUPF algo
rithm
based
on th
e BDS/INS i
n
tegrate
d
n
a
v
igation
system mo
de
l establi
s
he
d by the direct
method a
c
cording to the literatu
r
e [3];
Ca
se 2: the
traditional
FUPF algorith
m
bas
ed o
n
th
e all-attitude
integrate
d
B
D
S/INS
navigation
system mod
e
l e
s
tabli
s
he
d by
the dire
ct method in the the
s
is;
Ca
se 3: the improve
d
AFUPF algo
rith
m base
d
on
the all-attitud
e
integrate
d
BDS/INS
navigation
system mod
e
l e
s
tabli
s
he
d by
the dire
ct method in the the
s
is.
Simulation compute
r
is: Pentium (R) D
ual-Co
r
e
E5200 CP
U;
2.50GHz freque
ncy;
2.00GB RAM
;
Windo
ws X
P
Profession
al OS. The
Matlab 7.1 si
mulator
run
s
in a Monte Carlo
fashio
n. In the sim
u
lation,
the sa
mpling
freque
ncy
of I
M
U i
s
10
Hz, the up
date
cycle of the
BDS’s
data is 1
s
, and the sim
u
lation time is 500s. Set u
p
: the rand
o
m
con
s
tant
of gyroscop
e
is
0.
1
/
h
, the driving white noi
se o
f
the first-
ord
e
r Markov proce
s
s of gyroscope is
0.
1
/
h
, the
white n
o
ise of
gyrosco
pe i
s
0.
05
/
h
, the relate
d
time is 60
s; t
he drivin
g whi
t
e noise of th
e firs
t-
orde
r Ma
rkov process of g
y
roscop
e is
g
4
10
2
,
the related ti
me is 60
s; the position, ve
locity
and attitude p
r
eci
s
io
n of the BDS re
ceivers
whi
c
h
is u
s
ed in the
system is 10m, 0
.
2 m/s and 1’.
The re
sult
s from ea
ch case are
comp
ared in term
s o
f
trajectory, velocity, positi
on and
attitude error.
The co
mpa
r
i
s
on
s of traje
c
tory unde
r
different filter al
gorithm a
r
e
showed in Fig
u
re
2. The e
r
ror
curve
s
of the
positio
n e
s
ti
mation in
th
e
three
ca
se
s
are
sh
owed i
n
Figu
re
3 to
5.
The e
r
ror curves of th
e vel
o
city e
s
timation in
the th
re
e cases are
showed i
n
Fig
u
re
6 to
8. T
h
e
error curve
s
of the attitude estimation in
the
three cases are sh
owe
d
in Figure 9 to 10.
(a)
Comp
ari
s
on of estimat
ed and real
trajecto
ry
(b)
Comp
ari
s
on of trajecto
ry in height-lat
itude
plane
(c) Com
p
a
r
ison of trajecto
ry in height-lo
ngitude
plane
(d)
Comp
ari
s
on of trajecto
ry in latitude-
longitud
e
pla
n
e
Figure 2. Co
mpari
s
o
n
of trajecto
ry unde
r different filter algo
rithm
46.
78
46.
8
46.
82
46.
84
46.
86
46.
88
46
.
9
8
8.
2
8.
4
8.
6
3000
3200
3400
3600
3800
4000
4200
4400
4600
4800
L
o
n
g
i
t
ude
(
)
L
a
ti
tu
d
e
(
)
He
i
g
h
t
(
m
)
T
h
e t
r
u
e
t
r
aj
ec
t
o
r
y
Th
e
c
a
s
e
1
Th
e
c
a
s
e
2
Th
e
c
a
s
e
3
8.
0
5
8.
1
8.
1
5
8.
2
8.
2
5
8.
3
8.
3
5
8.
4
8.
4
5
8.
5
300
0
320
0
340
0
360
0
380
0
400
0
420
0
440
0
460
0
480
0
L
ong
i
t
ud
e
(
)
H
e
ight
(
m
)
T
h
e t
r
u
e
t
r
aj
ec
t
o
r
y
Th
e
c
a
se
1
Th
e
c
a
se
2
Th
e
c
a
se
3
8.
1
6
8.
1
7
8.
1
8
8.
19
41
00
41
50
42
00
42
50
8.
34
8.
35
8.
3
6
8.
37
8.
38
3
350
3
400
3
450
46.
78
46.
8
46.
82
46.
84
46.
86
46.
88
46.
9
3000
3200
3400
3600
3800
4000
4200
4400
4600
4800
L
ongi
t
ude
(
)
)
He
i
g
ht
(
m
46.
815
46.
82
46.
825
46.
83
4100
4150
4200
4250
46.
865
46.
87
46.
875
3350
3400
3450
T
h
e
tr
ue
tr
a
j
e
c
t
o
r
y
Th
e
c
a
s
e
1
Th
e
c
a
s
e
2
Th
e
c
a
s
e
3
46
.
7
8
46
.
8
46
.
8
2
46.
84
46
.
8
6
46
.
8
8
46
.
9
8.
05
8.
1
8.
15
8.
2
8.
25
8.
3
8.
35
8.
4
8.
45
8.
5
Lo
n
g
i
t
u
d
e
(
)
L
a
ti
tud
e
(
)
46
.
8
1
4
6
46
.
8
1
4
8
46
.
8
1
5
8.
1
4
7
8.
1
4
8
8.
1
4
9
46
.
8
79
8
46
.
8
8
46.
88
02
8.
3
9
8.
3
9
1
8.
3
9
2
46
.
8
5
9
6
46
.
8
59
8
46
.
8
6
46
.
8
60
2
8.
31
6
8
.
31
65
8.
31
7
8
.
31
75
8.
31
8
T
h
e
tr
u
e
tr
a
j
e
c
tor
y
Th
e
c
a
se
1
Th
e
c
a
se
2
Th
e
c
a
se
3
Evaluation Warning : The document was created with Spire.PDF for Python.