TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5324 ~ 53
3
0
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.566
0
5324
Re
cei
v
ed
Jan
uary 23, 201
4
;
Revi
sed Ma
rch 1
8
, 2014;
Acce
pted April 4, 2014
Sensor Fusion via Brain Emotional Learning for Ground
Vehicle
Alv
a
ro Varg
as-Clara, Sa
ngram Re
dk
ar*
Dep
a
rtment of Engi
neer
in
g an
d Comp
uti
ng S
y
stems, Arizo
n
a
State Univer
sit
y
700
1 E. W
illia
ms Field Ro
ad,
Me
sa AZ 8521
2, United State
s
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: sredkar@
a
su
.edu
A
b
st
r
a
ct
In this w
o
rk, the an
alysis
of a
filter cons
istin
g
of
the
Br
ai
n Emoti
o
n
a
l Lear
nin
g
(BEL) alg
o
rith
m i
s
prese
n
ted. T
h
e inn
e
r w
o
rkin
gs of
the BEL
filter are bas
e
d
on e
m
otio
na
l lear
nin
g
mod
e
l in
ma
mmali
an
s
brai
n. T
h
e
BEL
filter
is i
m
ple
m
e
n
ted
in
si
mulati
on f
o
r
the
purp
o
se
of se
n
s
or fusi
on
in
a
grou
nd v
e
h
i
cle
.
I
n
simulati
on, th
e
sign
als fro
m
a
Globa
l P
o
sitio
n
in
g S
yste
m
(
G
PS) and
an
Inertia
l
N
a
vig
a
ti
on Syste
m
(IN
S
)
are i
n
tegr
ated,
in or
der t
o
acc
u
rately tr
ack th
e traject
o
ry of
a gro
u
n
d
ve
hic
l
e ar
ou
nd
a tra
ck. T
he BEL fi
l
t
er
is provid
ed w
i
th some sens
or
y signa
l and r
e
w
a
rd sign
al
, subse
q
u
ently the filter
seeks
to dimi
nish
n
o
is
e
from both
sens
ing un
its,
thus eli
m
i
nati
ng trac
king
error. A
p
e
rformanc
e
co
mp
ariso
n
betw
een
the BE
L fil
t
er
,
and th
e
mor
e
commonly
util
i
z
e
d
Ka
l
m
a
n
fil
t
er is pr
es
ente
d
. T
he BEL fi
lter de
monstrat
ed ro
bustn
ess
to
uncerta
inties
from the s
ens
in
g u
n
its, it a
dap
ts quickl
y
w
i
th
dyna
mical
cha
nge
in
the
pl
a
n
t, and
has
s
m
a
l
l
computati
o
n
a
l
cost. T
he BEL filter de
mo
ns
trated to be effe
ctive in sens
or fusion.
Ke
y
w
ords
: Na
vigati
on; Emoti
ona
l decis
io
n mak
i
n
g
;
Senso
r
integrati
on; Kal
m
a
n
F
ilter
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
purpo
se
of a
navigat
ion
system
i
n
a ve
hi
cl
e i
s
to
dete
r
min
e
its
cu
rrent
locatio
n
,
velocity, and
dire
ction; in
other wo
rd
s determi
ne
t
he stat
e of t
he vehi
cle. T
h
is info
rmatio
n is
usu
a
lly obtai
ned from mul
t
iple se
nsors
on the ve
hicl
e. The
sen
s
o
r
s
co
mmonly
use
d
a
r
e a
G
PS,
and an INS.
A GPS is a sen
s
or that
provide
s
p
o
sitioni
ng da
ta relative to an earth
-centere
d
coo
r
din
a
te
system. It uses at lea
s
t 4
o
r
mo
re
satelli
tes
with a
n
u
nob
stru
cted li
ne of
sight t
o
cal
c
ulate p
o
sition, time, a
nd velocity. GPS re
ceive
r
s can obtai
n
signal
s from
GPS satellites
unde
r any weather
con
d
itions, an
d an
ywhere on
E
a
rth. GPS are available for civilian a
n
d
military appli
c
ation
s
. They
are highly a
c
curate
in th
ree-di
men
s
ion
a
l positio
ning
. GPS position
errors are bo
unde
d and a
r
e depe
nde
nt on the availa
bility of GPS
satellite
s
[1].
An
IN
S se
nso
r
us
es
a
cce
le
r
a
tion
, a
n
d
r
o
ta
tio
nal sen
s
o
r
s
to continuo
usly calcul
ate
positio
n, ori
e
ntation, an
d
velocity. Although, it
s p
r
i
m
ary o
u
tput i
s
p
o
sitio
n
rel
a
tive to an
e
a
rth
-
cente
r
ed
co
o
r
dinate
syste
m
. In contra
st to a
GPS
sen
s
o
r
, the
INS positio
n errors a
r
e
not
boun
ded, an
d gro
w
with time. In addition, the erro
rs are depe
nde
nt on the qua
lity of its inertial
s
e
ns
or
s
[1
].
The integ
r
ati
on of GPS and INS are i
n
efforts to combat ea
ch
of the sen
s
in
g unit’s
wea
k
n
e
sse
s
. For exam
ple,
INS are initi
a
lly given
po
sition an
d vel
o
city inform
ation from a
not
her
sou
r
ce, an
d
sub
s
e
que
ntly it gene
rate
s its o
w
n
u
p
d
a
ted po
sition
and velo
city by integrating
informatio
n receive
d
from
its inertial
sensors.
Ho
wever, any small erro
rs
whi
c
h a
r
ise i
n
the
measurement
are integ
r
ate
d
into grad
ual
ly larger
e
r
rors. By integrating the INS wi
th a GPS, the
GPS capabili
ty for online
calibration and error
estim
a
tion will hel
p mitigate the INS integration
drift. Conversely, in the event that there is an ob
struction to the line of
sight bet
ween vehicle
and
satellite
s
, an
d the GPS is unable to p
e
rform. Th
e INS ca
n perfo
rm as th
e sh
ort-te
rm ba
ckup
whe
n
GPS sign
als a
r
e
unavailabl
e. Theref
ore, a
s
GPS and
INS have compl
e
me
nta
r
y
cha
r
a
c
teri
stics, their imple
m
entation is
con
s
id
ere
d
in
an integrate
d
appro
a
ch [2].
As a re
sult, the navigation
system utilizes
the outp
u
t signal
s from
these sen
s
o
r
s an
d
integrate
s
th
em to
obtain
more p
r
e
c
i
s
e info
rmat
ion
abo
ut the
vehicl
e’s state
.
This p
r
ocess of
integratio
n is
comm
only ref
e
rred to a
s
sensor fu
sion.
There are nu
mero
us m
e
th
ods to fu
se I
N
S
and
GPS, such
a
s
, loo
s
ely co
upled
or tightly
co
upled
integ
r
a
t
ion. In the
majority of t
hese
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Senso
r
Fu
sio
n
via Brai
n Em
otional Learning
for G
r
ou
nd Vehicl
e (A
lvaro Varg
as-Clara)
5325
desi
g
n
s
GPS and INS integration filter i
s
usually so
me form of a Kalman filter [1], [3-6].
In
most
ca
se
s, an ex
tended Kal
m
an filter is i
m
pleme
n
ted
wi
th ine
r
tial
errors a
s
its state to obt
ain
satisfa
c
to
ry p
e
rform
a
n
c
e.
Kalman
fi
lter equ
ations a
r
e o
p
timal
when
se
nso
r
observation
s are
unbia
s
e
d
wit
h
white
noi
se. Also, the
r
e i
s
a h
e
a
vy computati
onal
co
st in
Kalman filter
impleme
n
tation, due to co
nstant up
dati
ng of Kalman
gains.
In this pap
er, we p
r
e
s
ent a BEL filter integration app
roa
c
h to achie
v
e lower
comp
utationa
l effort but with competi
t
ive
perform
ance mea
s
u
r
es
co
mpa
r
e
d
to the m
o
re
comm
only used Kalman filter.
The pa
per i
s
orga
nized a
s
follows. Th
e sen
s
o
r
inte
gration BEL
filter is discu
s
sed in
se
ction
2. Implementatio
n
of the BEL
filter an
d
si
mulation
set
up i
s
di
scussed
in
se
ctio
n 3.
Simulation
s result
s are p
r
e
s
ente
d
in se
ction 3.
Lastly, con
c
lu
sion
s
are ma
de in the se
ction 4.
2. Proposed
Metho
d
The p
r
op
ose
d
filter utilize
d
for
sen
s
o
r
fusion
con
s
ists of th
e B
E
L model. B
E
L is a
netwo
rk
mod
e
l whi
c
h
sim
u
lates th
e b
r
ain e
m
otion
a
l learning
p
r
ocess of m
a
mmalia
n was
develop
ed b
y
Balkeniu
s
& Moren [
7
, 8]. It is
a com
putatio
nal mod
e
l o
f
the amygdala,
Orbitofrontal
Cortex (OF
C
), thalamu
s
, and se
nsory input cort
ex, which a
r
e kno
w
n to
be
respon
sibl
e for emotio
nal l
earni
ng an
d p
r
ocessin
g
.
Re
sea
r
che
r
s
in cont
rol ha
ve taken inte
rest in utili
zin
g
this BEL m
odel a
s
a fe
edba
ck
controlle
r. Th
is is
motivate
d by the fa
ct that
re
sea
r
ch
in psy
c
hol
og
y, AI, and co
gnitive scien
c
e
identify the
re
ciprocal influ
ences
of e
m
o
t
ion an
d
cog
n
i
tion [8]. Thi
s
is m
o
tivated
by the fa
ct th
at
resea
r
ch in psych
o
logy, AI, and cognitiv
e
scie
n
ce
ide
n
tify the recip
r
ocal influen
ces of emotio
n
,
cog
n
ition an
d
de
ci
sion
makin
g
[8]. Therefore,
L
u
ca
s
et al. [
9
], first intro
duced th
e B
r
ain
Emotional Le
arnin
g
Ba
se
d
Intelligent Controlle
r (
BEL
BIC)
whi
c
h
consi
s
ted
of th
e BEL mo
del
but
utilized
as
dire
ct ada
ptive feedba
ck contro
l. BELBIC ha
s
been im
ple
m
ented in
many
engin
eeri
ng systems ap
pli
c
ation
s
, su
ch
as, power
system [10], aero
s
pa
ce la
u
n
ch vehi
cle [
11],
queu
e man
a
gement [12],
flight simula
tion se
rvo
sy
stem [13], a
nd othe
r un
certain n
onlin
ear
system
s [14]. In all applications the B
E
L model
de
monst
r
ated robu
stne
ss to
unce
r
taintie
s
, on-
line a
dapta
b
i
lity, and sm
all computati
onal
co
st. Howeve
r, the
r
e is not
an
y re
sea
r
ch i
n
impleme
n
ting
BEL as a filter for sen
s
or
fusion.
The in
ner wo
rkin
g of BEL
is a
n
a
c
tion
g
eneration
system found
ed
on
sen
s
o
r
y in
put an
d
reward si
gnal
[14]. The em
otional le
arni
ng o
c
curs
p
r
i
m
arily in the
amygdala. T
h
e learning
of the
amygdala i
s
given in the followin
g
equ
a
t
ion:
ma
x
0
,R
e
ai
GS
w
A
(
1
)
Whe
r
e
G
a
i
s
t
he amygd
a
la
gain,
is the a
m
ygdala le
arning rate,
S
i
is the
sen
s
o
r
y input,
Re
w
is
the re
wa
rd
si
gnal, an
d
A
is the a
m
ygda
la output. Th
e
ma
x
term i
s
for ma
king t
he lea
r
nin
g
in
the
amygdala m
o
notoni
c, implying that learni
ng in amygda
la sho
u
ld be
perm
ane
nt.
Similarly, the learni
ng rule in OFC i
s
sh
o
w
n in the follo
wing e
quatio
n:
Re
oi
GS
M
O
w
(
2
)
Whe
r
e
G
o
is t
he OF
C gain,
is
the OF
C
l
earni
ng rate, and
MO
is th
e mod
e
l outp
u
t, calculated
as in Equatio
n (3):
MO
A
O
(
3
)
In whic
h,
O
i
s
the
o
u
tput
of the
OF
C.
The m
odel
first receives th
e sen
s
o
r
y in
put,
S
i
, then
the
model calcul
ates the inte
rnal sig
nal
s of the am
ygdal
a and OF
C, these sig
nal
s are
calculate
d
as
in Equation (4) and
(5):
ai
A
GS
(
4
)
oi
OG
S
(
5
)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5324 – 53
30
5326
The amyg
dal
a lea
r
n
s
to p
r
edi
ct an
d re
act to give
a
n
emotio
nal
sign
al. Whil
e
the OF
C
system
detect
s
the differen
c
e be
tween the e
x
pected
sy
st
em’s p
r
edi
cti
on and the
actual received
emotional sig
nal
[15].
Controlle
rs b
a
se
d o
n
the
BEL model
d
e
mon
s
tr
ate
d
robu
stne
ss t
o
un
ce
rtainti
e
s,
while
being
simple
and havin
g lo
w co
mputatio
nal co
st. To utilize this ve
rsio
n of the BEL model a
s
a
filter, it is im
p
o
rtant to
und
ersta
nd th
at
BEL m
odel
in
esse
nce
con
v
erts t
w
o
set
s
of inp
u
ts
(
S
i
and
Re
w
) into a
d
e
ci
sion
sign
al
as its out
put. Therefo
r
e, it is impo
rtant to impleme
n
t this BEL mod
e
l
in an app
ro
priate ma
nn
er so that
input sig
nals and output
signal
s ha
ve the prop
er
interp
retation
s for the probl
em at hand.
3. Rese
arch
Metho
d
In this paper, a sim
u
lation of a
ground
vehi
cle
around a t
r
ack i
s
utilized
to draw
perfo
rman
ce
comp
ari
s
o
n
b
e
twee
n Kalm
an Filter
a
n
d
the BEL filter. The pe
rformance of the
s
e
two filter i
s
based on their ability reduce noi
se
f
r
om
GPS as the vehicle trajectory i
s
tracked.
Two t
r
a
c
ks
are sim
u
lated,
a ci
rcular an
d
figure
-
8
t
r
a
c
k. The
vehi
cl
e is
mod
e
led
as traveling
a
t
a
velocity of 5 m/s. The traje
c
tory of the vehicl
e on the
track is given
by the followi
ng equ
ation [1]:
3s
i
n
2s
i
n
c
o
s
1/
2
c
o
s
po
s
Nor
t
hi
ng
S
t
Eas
t
i
n
g
S
t
t
Dow
n
h
t
δ
(
6
)
Whe
r
e
S
is the tra
c
k
scali
ng p
a
ra
meter,
h
is the
cr
os
sover
height,
is m
ean
an
gular spee
d,
and
is an arbitrary pha
se an
gle. T
h
is mod
e
l is implemente
d
in MATLAB, which al
so
cal
c
ulate
s
ve
hicle
velo
city, accel
e
ratio
n
attitude, an
d
attitude rate
s. Th
e traj
ect
o
rie
s
simul
a
ted
can b
e
se
en i
n
Figure 1. Both simulate
d
trac
ks h
a
ve cha
nge
s in el
evation of 10 meters.
Figure 1. Figure
-
8 Track (l
eft)
and Ci
rcu
l
ar Tra
c
k (rig
ht)
The vehi
cle
d
y
namic
mod
e
l
co
nsi
s
t of
a
Type2 T
r
a
cki
ng Mo
del. T
h
i
s
tra
c
king
mo
del
can
estimate po
sition, velocity in three di
mens
i
o
n
s
, given the app
rop
r
iate mea
s
ureme
n
ts. T
he
tracker utilizes a
host
vehicl
e dynamic model
with zero
-mean white noi
se accelerati
on,
unbo
unde
d
steady-state mean squa
re
d
velocity
an
d unbo
und
ed
steady-state
mean squa
red
positio
n vari
ations. T
he
full trackin
g
model i
s
im
plemente
d
,
whi
c
h in
clu
d
e
thre
e po
si
tion
comp
one
nts
and th
ree vel
o
city compo
n
ents. Th
e
n
e
c
e
s
sary Kalm
an filter
com
pone
nts fo
r a
3-
dimen
s
ion Ty
pe2 tra
cki
ng filter are the fo
llowing:
2
2
2
2
,
2
,
2
,
00
0
0
0
00
0
0
0
00
0
0
0
00
0
0
0
00
0
0
0
00
0
0
0
N
E
D
vN
vE
vD
0
P
(
7
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Senso
r
Fu
sio
n
via Brai
n Em
otional Learning
for G
r
ou
nd Vehicl
e (A
lvaro Varg
as-Clara)
5327
10
0
0
0
01
0
0
0
00
1
0
0
000
1
0
0
000
0
1
0
000
0
0
1
t
t
t
Φ
(
8
)
22
22
22
000
0
0
0
000
0
0
0
000
0
0
0
000
0
0
000
0
0
000
0
0
ac
c
ac
c
ac
c
t
t
t
Q
(
9
)
Whe
r
e
P
0
is t
he estim
a
tion
unce
r
tainty covarian
ce m
a
trix,
Φ
is the st
ate-tra
n
sitio
n
matrix, and
Q
is the cova
rian
ce of dy
n
a
mic di
sturba
nce n
o
ise.
The Kalman f
ilter utilized for the perform
anc
e com
p
ari
s
on i
s
of the following form
:
TT
KP
H
H
P
H
R
(
1
0
)
Whe
r
e
K
i
s
t
he Kalm
an
g
a
in,
H
is me
asu
r
em
ent
sensitivity matrix, and
R
is
th
e
se
ns
o
r
no
is
e
c
o
varianc
e
matrix.
11
1
xx
K
z
H
x
(
1
1
)
Whe
r
e
z
is
measurement
vector, which is
comp
os
ed of the
co
mputed p
o
siti
on, velocity
and
clo
ck e
r
r
o
r
s
f
r
om t
he GP
S
.
P
P
KHP
(
1
2
)
The implem
e
n
tation follo
ws the a
bov
e equatio
ns
in chronol
ogi
cal orde
r. First, the
Kalman gai
n
is co
mpute
d
by Equatio
n (10
)
; Follo
wed by the
corre
c
ted
sta
t
e estimation
in
Equation (11); lastly, the correcte
d cova
rian
ce matrix
is com
puted
by Equation (12). To finali
z
e
the Kalman
filter imple
m
entation, t
he temp
oral
update
s
a
r
e co
mputed
by the foll
owin
g
equatio
ns:
11
x
Φ
x
(
1
3
)
T
P
Φ
P
Φ
Q
(
1
4
)
The implem
e
n
tation of BEL model a
s
a
filter is cho
s
en to be in si
milar man
n
e
r
as the
Kalman filte
r
i
m
pleme
n
tatio
n
. Thi
s
d
one
i
n
effo
rt
s to
draw
an
accu
ra
te pe
rform
a
n
c
e
co
mpa
r
iso
n
betwe
en BEL
filter and Kal
m
an filter. Ho
wever,
slight
differen
c
e
s
arise du
e to the fact that BEL
model i
s
ori
g
inally de
sig
n
ed fo
r d
e
scri
ptive pur
po
se with
no
en
igeeri
ng
appl
ication
in
mind.
Therefore, it
is upto the d
e
sig
ner to a
p
p
rop
r
ia
tely select the
sen
s
ory in
put si
gnal an
d re
ward
sign
al in acco
rdan
ce to en
g
i
neeri
ng ap
pli
c
ation.
For the imple
m
entation of the BEL filter in
this study, we sele
cted the se
nsory in
put (
S
i
)
to be of the form (15
)
:
i
S
1
zx
(
1
5
)
W
h
er
e
x
1
i
s
the vehi
cle
states o
b
taine
d
from
the
vehicl
e traj
ect
o
ry mo
del.
GPS
data
can
b
e
obtain from
a num
be
r of
satellite
s
,
rangin
g
fro
m
4 to 2
9
. In
addition,
GPS noise
can
be
simulate
d to be of different
noise di
strib
u
tions.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5324 – 53
30
5328
The re
wa
rd
function (
Re
w
) is
sele
ct
ed with obj
e
c
tive of minimzing the di
fference
betwe
en
GPS
and
M
e
a
s
ured
.
Thi
s
fu
nction
plays
an imp
o
rtant
role
in BEL
filter. The fil
t
er
attempts to increa
se the rewa
rd while minimizi
ng the sen
s
o
r
y input. The implemented reward
function i
s
given in Equatio
n (16
)
:
12
Re
i
wK
S
K
(
1
6
)
Whe
r
e
K
1
a
nd
K
2
are
g
a
ins. Th
e re
ward fun
c
tio
n
gain
s
a
r
e
positive real
numbe
rs. F
r
om
Equation
(16
)
, it can be
se
en that BEL filter obt
ain
s
m
a
ximum re
wa
rd when th
e sensory inp
u
t i
s
zero. Cl
osely notici
ng Eq
u
a
tion (15
)
, th
e sen
s
ory i
n
p
u
t is i
n
e
s
sen
c
e, a
n
e
r
ror
signal. Th
e B
E
L
filter tries
to diminis
h
the the error.
To
carry o
u
t the
simul
a
tion a
nu
mbe
r
of p
a
ra
mete
rs ha
d to
be
sel
e
cte
d
. Fi
rst, the
learni
ng rate
s
for
the am
ygdala
a
nd OFC we
re selecte
d
to
b
e
16
e
, and
14
e
,
respe
c
tively. The OF
C le
arning rate was cho
s
e
n
to be
slightly large
r
to ma
ke the
OFC l
earn th
e
error in the
amygdala q
u
i
cker than th
e amygdala
itself to eliminate the error. The othe
r
para
m
eters were
the
gain
s
in th
e
Re
w f
unctio
n
, which were
sele
ct
ed to
be
K
1
= 0.001
a
nd
K
2
=
1. Th
ese
pa
ramete
rs a
n
d
lea
r
ning
rate
s
we
re
sel
e
ct
ed th
rou
gh tri
a
l an
d e
r
ror to imp
r
ove B
E
L
filter perform
ance.
All simulations are carried
out in MATLab. The
number of
satellites for
GPS is
varied. In
addition, GP
S noise dist
ri
bution
s
are varied.
Perfo
r
mance mea
s
ure
s
for both
Kalman and
BEL
filter are ave
r
age RMS error for p
o
stio
n
s
, velo
city, and average
Central
Processing Unit
(CP
U
)
time.
4. Results a
nd Analy
s
is
The first sce
n
a
rio
simul
a
te
d is
with
a ci
rcula
r
tra
c
k. T
he
simulatio
n
time is
sel
e
ct
ed to b
e
0.2 hou
rs. T
h
e first 1
00
se
con
d
s
of the
simulatio
n
da
ta is not
sam
p
led to all
o
w
settling time.
The
simulation i
s
executed
100 iterations. The num
ber of
satellites for th
i
s
scenario is 29.
Perform
a
n
c
e
of Kalman an
d BEL filter are obtaine
d, re
sults a
r
e
sho
w
n in Tabl
e 1
.
Table 1. Perf
orma
nce Co
mpari
s
o
n
for Circula
r
Track Simulation
The a
bove t
able d
e
mon
s
trates th
at BEL filter wa
s sup
e
rio
r
in
diminishing
p
o
sition
al
errors. Thi
s
t
r
end
was m
a
intained th
ro
ugh all
GPS
noise di
stribu
tions. In
som
e
ca
se
s, it e
v
en
perfo
rmed
be
tter than Kal
m
an filter in
redu
cin
g
velo
city errors. A signifi
cant re
sult obtai
ned
is
that BEL pe
rforme
d bette
r
in re
du
cing th
e computat
io
nal cost
ac
ro
ss
all
n
o
ise di
stributio
n
ca
ses.
In the wors
t cas
e
, BEL CP
U time was
half
of the Kalman filter be
st CPU time.
The
se
cond
scena
rio
sim
u
lated i
s
with
a fi
gure-8
track. Thi
s
si
m
u
lation wa
s
condu
cted
in simil
a
r fashion
as the fi
rst
scen
ario.
The figu
re
-8 t
r
ack sim
u
late
d a
more d
e
m
andin
g
tracking
trajecto
ry. Ta
ble 2 illustrates the result
s obtained fro
m
the se
cond
simulation
scenari
o
.
Po
si
t
i
on
N
[
m
]
P
o
s
it
io
n
E
[
m
]
P
o
s
it
io
n
D
[m
]
V
e
l
o
c
i
t
y
N
[m
/
s
]
V
e
l
o
c
i
t
y
E
[m
/
s
]
V
e
l
o
c
i
t
y
D
[m
/
s
]
N
(
0,
2
)
1.
1
4
5
1
0.
316
7.
7
7
9
5
.
373
3.
2
0
6
3
.
3
33
0
.
1
0
1
U(
‐
1
,
1)
1.
2
6
8
8
.
1
31
6.
0
4
5
4
.
772
3.
0
3
1
3
.
2
28
0
.
0
9
2
E
x
p
(
2
)
1.
1
6
5
1
0.
819
8.
3
0
5
8
.
245
3.
1
9
2
3
.
3
35
0
.
1
5
7
Tri
(
‐
1,
0
,
1)
1.
1
3
3
7
.
8
54
5.
7
2
1
4
.
749
3.
0
2
7
3
.
2
12
0
.
0
8
9
W
e
i
(
1,
2)
1.
1
7
4
8
.
2
58
5.
9
5
1
5
.
750
3.
0
2
6
3
.
2
19
0
.
1
0
5
Po
si
t
i
on
N
[
m
]
P
o
s
it
io
n
E
[
m
]
P
o
s
it
io
n
D
[m
]
V
e
l
o
c
i
t
y
N
[m
/
s
]
V
e
l
o
c
i
t
y
E
[m
/
s
]
V
e
l
o
c
i
t
y
D
[m
/
s
]
N
(
0,
2
)
0.
5
8
6
4
.
9
15
5.
0
1
7
4
.
934
4.
9
0
8
4
.
9
98
4
.
9
1
9
U(
‐
1
,
1)
0.
5
7
5
2
.
2
76
2.
8
6
1
2
.
714
2.
7
6
8
2
.
8
68
2
.
7
1
0
E
x
p
(
2
)
0.
5
9
3
4
.
9
25
4.
8
8
5
4
.
894
4.
9
2
7
4
.
9
35
4
.
9
1
1
Tri
(
‐
1,
0
,
1)
0.
5
4
0
1
.
9
37
2.
0
1
0
1
.
907
1.
9
4
4
2
.
0
14
1
.
9
0
7
W
e
i
(
1,
2)
0.
5
9
8
4
.
2
64
4.
0
1
8
4
.
169
4.
3
5
3
4
.
2
50
4
.
2
3
5
GP
S
No
i
s
e
D
i
s
t
r
i
but
i
o
n
Avg.
CPU
Ti
m
e
[s
e
c
]
Avg.
RM
S
Erro
r
Ka
l
m
a
n
F
ilt
e
r
GP
S
No
i
s
e
D
i
s
t
r
i
but
i
o
n
Avg.
CPU
Ti
m
e
[s
e
c
]
Avg.
RM
S
Erro
r
BE
L
F
ilt
e
r
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Senso
r
Fu
sio
n
via Brai
n Em
otional Learning
for G
r
ou
nd Vehicl
e (A
lvaro Varg
as-Clara)
5329
Table 2. Perf
orma
nce Co
mpari
s
o
n
for Figure-8 T
r
a
c
k Simulation
Re
sults o
b
tai
ned from the
Figure-8 track simu
l
a
tion
are si
milar to
the ones o
b
tained in
the previous
scena
rio, but
with
slightly
highe
r CP
U t
i
me and
po
si
tional erro
rs for b
o
th Kalm
an
filter and BEL filter implementation
s
. The figure-
8 tra
ck a
ppea
rs to be no more rigoro
u
s tha
n
the
circula
r
tra
c
k. For furth
e
r
perfo
rman
ce
comp
ar
i
s
o
n
betwe
en the
two filter imp
l
ementation
s
,
a
more inte
re
sting scen
ario i
s
analy
z
ed.
To co
ncl
ude,
the effects
of the numb
e
r o
f
sa
tellittes av
ailable i
s
an
al
yzed. As p
r
ev
iously
discussed, the num
ber of
satellites i
s
a determ
ental factor for
GPS to accurately calculate
positio
n and
velocity of a vehicle.
Th
erefore, for thi
s
last sce
nari
o
the num
be
r of satellite
s
is
varied from
4 to 29. Th
e
i
r effect
s on
the Ka
lman
and BEL filte
r
pe
rform
a
n
c
e are obtai
n
ed,
sho
w
n in Ta
b
l
e 3.
Table 3. Effects of Number of Satellites
on Kalman and BEL Filter Implementati
o
n
In the majorit
y
of the case
s the RMS e
rro
r for po
siti
on and velo
city increa
sed
as the
number
of satellites decreased fo
r both filter impl
em
entations.
However,
the i
n
crement
s in the
BEL implem
entation were small i
n
compa
r
ison to
the Kalman
filter. In the Kalman fil
t
er
impleme
n
tation, the RMS error for p
o
sit
i
on and velo
city appear to gro
w
expon
e
n
tially when t
he
satellites decreased f
r
om
14 to 4.
The result
s dem
onstrate
that the BEL filter is less sensitive to
the effects
of the numb
e
r of satellite
s
availabl
e. In
addition, the
CPU time i
n
cre
a
sed a
s
t
he
numbe
r
of satellites i
n
cre
a
se
d fo
r b
o
th filter im
ple
m
entation
s
.
Although, thi
s
effe
ct was
more
Posi
t
i
o
n
N
[m
]
P
o
s
i
t
i
o
n
E
[
m
]
P
o
s
it
io
n
D
[m
]
V
e
l
o
c
i
t
y
N
[m
/
s
]
V
e
l
o
c
i
t
y
E
[m
/
s
]
V
e
l
o
c
i
t
y
D
[m
/
s
]
N
(
0,
2)
1.
1
9
5
11.
259
8.
2
4
1
5
.
4
8
1
1
.
42
1
1
.
2
3
3
0.
100
U(
‐
1
,
1)
1.
1
9
0
9
.
1
13
6.
4
1
3
4
.
9
3
2
1
.
03
2
0
.
8
7
9
0.
090
E
x
p
(
2)
1.
1
9
8
11.
429
8.
4
5
1
8
.
4
7
0
1
.
45
8
1
.
2
1
1
0.
152
Tr
i
(
‐
1,
0,
1)
1.
1
5
8
8
.
6
74
6.
1
9
6
5
.
0
3
4
1
.
00
2
0
.
8
5
8
0.
089
W
e
i
(
1,
2)
1.
2
0
2
8
.
8
29
6.
2
0
1
5
.
8
3
7
0
.
99
8
0
.
8
6
4
0.
109
Posi
t
i
o
n
N
[m
]
P
o
s
i
t
i
o
n
E
[
m
]
P
o
s
it
io
n
D
[m
]
V
e
l
o
c
i
t
y
N
[m
/
s
]
V
e
l
o
c
i
t
y
E
[m
/
s
]
V
e
l
o
c
i
t
y
D
[m
/
s
]
N
(
0,
2)
0.
5
8
8
4
.
9
36
5.
0
3
5
4
.
9
2
9
4
.
91
5
5
.
0
0
6
4.
932
U(
‐
1
,
1)
0.
5
8
1
2
.
8
23
2.
8
2
2
2
.
7
9
6
2
.
83
0
2
.
8
3
7
2.
799
E
x
p
(
2)
0.
5
9
1
4
.
9
39
4.
9
2
8
4
.
9
0
5
4
.
95
7
4
.
9
4
3
4.
854
Tr
i
(
‐
1,
0,
1)
0.
5
4
3
2
.
0
06
1.
9
7
2
1
.
9
9
2
2
.
00
9
1
.
9
9
5
1.
992
W
e
i
(
1,
2)
0.
5
9
4
4
.
2
64
4.
0
9
2
4
.
1
8
0
4
.
37
5
4
.
1
5
8
4.
267
GP
S
No
i
s
e
Di
s
t
r
i
b
u
t
i
o
n
Av
g
.
CP
U
Ti
m
e
[s
e
c
]
Av
g
.
RM
S
Err
o
r
Ka
l
m
a
n
Fi
l
t
e
r
GP
S
No
i
s
e
Di
s
t
r
i
b
u
t
i
o
n
Av
g
.
CP
U
Ti
m
e
[s
e
c
]
Av
g
.
RM
S
Err
o
r
BE
L
Fi
l
t
e
r
Po
si
t
i
o
n
N
[m
]
P
o
s
i
t
i
o
n
E
[
m
]
P
o
s
it
io
n
D
[m
]
V
e
l
o
c
i
t
y
N
[m
/
s
]
V
e
l
o
c
i
t
y
E
[m
/
s
]
V
e
l
o
c
i
t
y
D
[m
/
s
]
4
0
.
466
414.
316
2694
.
149
258.
8
8
0
2
.
096
6.
439
0.
075
9
0
.
620
29.
285
30.
007
29.
8
6
9
1
.
831
1.
695
0.
129
14
0.
767
14.
058
8.
5
1
5
6
.
5
8
2
1.
712
1.
258
0.
106
19
0.
920
12.
259
8.
5
4
6
6
.
0
7
7
1.
527
1.
258
0.
105
24
1.
011
12.
051
8.
2
9
4
5
.
7
7
4
1.
519
1.
242
0.
104
29
1.
195
11.
259
8.
2
4
1
5
.
4
8
1
1.
421
1.
233
0.
100
Po
si
t
i
o
n
N
[m
]
P
o
s
i
t
i
o
n
E
[
m
]
P
o
s
it
io
n
D
[m
]
V
e
l
o
c
i
t
y
N
[m
/
s
]
V
e
l
o
c
i
t
y
E
[m
/
s
]
V
e
l
o
c
i
t
y
D
[m
/
s
]
4
0
.
341
5.
0
5
4
4
.
9
50
4.
95
5
5
.
090
4.
960
4.
909
9
0
.
380
4.
9
7
7
4
.
9
47
4.
93
4
4
.
952
4.
981
4.
941
14
0.
437
4.
9
2
9
4
.
8
97
4.
91
6
4
.
937
4.
941
4.
914
19
0.
476
4.
9
3
2
4
.
9
21
4.
91
5
4
.
923
4.
929
4.
924
24
0.
539
4.
9
3
5
4
.
9
98
4.
92
4
4
.
920
5.
125
4.
930
29
0.
588
4.
9
3
6
5
.
0
35
4.
92
9
4
.
915
5.
006
4.
932
Ka
l
m
a
n
F
ilt
e
r
:
No
r
m
a
l
No
is
e
Di
s
t
r
i
b
u
t
i
o
n
No
.
Sa
t
s
Avg.
CP
U
Ti
m
e
[s
e
c
]
Avg.
RM
S
Erro
r
No
.
Sa
t
s
Avg.
CP
U
Ti
m
e
[s
e
c
]
Avg.
RM
S
Erro
r
BE
L
Fi
l
t
e
r
:
No
r
m
a
l
No
i
s
e
Di
s
t
r
i
b
u
t
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5324 – 53
30
5330
notice
able
for the Kalm
an
filter impl
eme
n
tation.
La
stl
y
, a simil
a
r trend
wa
s
obta
i
ned i
n
that
the
BEL filter wa
s su
pe
rior at
diminishing p
o
sition
al
erro
rs, while the
Kalman filter
wa
s su
peri
o
r
a
t
redu
cin
g
the
velocity erro
rs. An im
port
a
nt note
abo
ut this
scen
ario
, the effect
s o
n
the
numb
e
r of
satellites was carried out with a Gaussian GPS noise
distribution.
The
re
sults f
r
om this stu
d
y
demon
strat
ed t
he BEL
qualitie
s a
s
a
filter. It su
ccessfully
filtered the n
o
ise fro
m
GPS and wa
s able to accurate
ly follow the trajecto
ry of
a vehicle aro
und
a
tr
ack
.
It de
mo
ns
tr
a
t
ed r
o
b
u
s
t
n
e
s
s
to
a va
r
i
e
t
y of noi
se
distrib
u
tion
s, and
all thi
s
with
signifi
cantly less co
mputat
ional cost.
5. Conclusio
n
For
navigatio
n, one
of the
most im
po
rtant task
i
s
to
be a
b
le to
a
c
curately
obt
ain the
vehicle’
s
stat
e from
an
a
s
sortm
ent
of
sensors. In thi
s
p
ape
r a
ne
w filter i
s
dev
elope
d, which is
based on BE
L model i
s
in
vestigated. In
simulatio
n
,
the BEL mod
e
l is implem
e
n
ted as
a filter in
efforts to
re
d
u
ce
GPS
sen
s
or noi
se
an
d to a
c
cura
te
ly obtain vehi
cle’
s
stat
es a
s
it i
s
traveli
ng
arou
nd a tra
ck. The resu
lts from this study
demo
n
strate the
BEL qualities as a filter. It
perfo
rmed b
e
tter at redu
ci
ng po
sitional
RMS error
while having
significa
ntly less co
mputatio
nal
co
st than the
traditional K
a
lman filter i
m
pleme
n
tatio
n
. In addition,
re
sults
sh
ow that BEL filter is
less sensitive to the effect
s of the num
ber of
satellit
es avail
able t
o
accurately
obtain GPS data.
Ho
wever,
the
BEL filter
pe
rforman
c
e
is g
r
eatly a
ffe
cte
d
by the
sele
ction
of the
sensory in
put
and
reward sig
nal
. Further
re
search in the
cha
r
a
c
te
ri
zati
on of the se
n
s
ory in
put an
d re
ward si
g
nal
can furt
her e
nhan
ce the B
E
L filter perfo
rman
ce.
In con
c
lu
sio
n
,
the BEL filter can be
use
d
in
the
real t
i
me appli
c
ati
on for filteri
n
g se
nsor
noise on a
c
count of its rob
u
stne
ss to no
ise
un
ce
rtaint
y, and small computation
a
l co
st.
Referen
ces
[1]
Grew
al MS, W
e
ill LR,
Andrew
s
AP.
Glo
bal
Positi
oni
ng
S
ystems, Iner
tial Navigation,
a
n
d
Integr
atio
n.
2nd e
d
. Ne
w
J
e
rse
y
: J
ohn W
i
le
y
& S
ons. 20
07: 255-
25
6.
[2]
H Qi, JB Moore. Direct Kalman Filter
ing
Appro
a
ch
for
GPS/INS
Integration.
IEEE T
r
ans. Aer
o
sp.
Electron. Syst.
200
2; 38(2): 68
7–6
93.
[3]
W
e
i M, Sch
w
a
r
z KP. A disc
u
ssion
of mo
del
s fo
r GPS/INS integr
atio
n. In
Y. Bock a
nd
N. Lep
par
d
(Eds.), Global
Positio
n
in
g S
y
stems: An Overvi
e
w
. Ne
w
Y
o
r
k
: Springer-V
er
lag. 19
89: 31
6-
327.
[4]
Sch
w
arz KP,
Wei M, Gelderen MV.
A
i
de
d
versus
embe
d
ded: A
co
mp
a
r
ison
of tw
o a
ppro
a
ches
to
GPS/INS integration
. In Proce
edi
ngs of IEEE PLANS. 199
4; 314-3
21.
[5]
B Guangro
ng, L Hon
g
she
n
, H Ning
h
u
i
. Multisensor D
a
ta Processi
ng an
d F
u
sing Bas
e
d on Kalm
a
n
Filter.
T
E
LKOMNIKA Indone
sian Jo
urna
l of Electrical E
ngi
neer
ing
. 2
013;
11(3): 12
66-
12
77.
[6]
H Guo. Neur
al Net
w
ork Ai
ded Ka
lman
F
ilter
ing for Integrate
d
GPS/INS Navigat
ion S
y
stem.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2013; 1
1
(3): 1
221-
122
6.
[7]
J Moren, C B
a
lke
n
ius.
A C
o
mputati
o
n
a
l Mode
l
of
E
m
o
t
i
ona
l Le
arn
i
ng
in the
Amy
g
d
a
la
. in: From
Animals to An
i
m
als 6, Proce
edi
ngs of the 6
th
Internationa
l Confer
ence o
n
the Simulati
on
of Adaptive
Behav
ior. MIT
Press. Cambri
dge. 20
00.
[8]
Balke
n
ius
C,
Moren
J. Emot
ion
a
l
lear
ni
ng:
A Com
putati
o
n
a
l M
ode
l
of the
Am
ygd
a
la.
C
y
b
e
r
ne
ti
cs and
System
s
. 20
01
; 32(6): 611
–63
6.
[9]
Lucas
C, D S
hahm
irzad
i
, N
Sheik
hol
esl
a
mi
. Introduci
ng B
E
LBIC: Brai
n
Emotion
a
l
Le
a
r
nin
g
Bas
e
d
Intelli
gent C
ont
roller.
Int. J. Int
e
ll. Autom
.
Soft Comput
. 200
4
;
10(1): 11-22.
[10]
Mohamm
d
i MR, Lucas C, Najar AB.
A Novel C
ontro
ll
er for a Pow
e
r System bas
ed BELBIC
. In
Procee
din
g
s of
W
o
rld Automa
tion Co
ngress.
200
4; 409-
420.
[11]
Mehra
b
ia
n AR
, Lucas C, Ros
han
ian J. Ae
rospac
e lau
n
c
h
vehicl
e
cont
rol: An intell
ig
ent ada
ptive
appr
oach.
Aer
o
spac
e Scie
nc
e and T
e
ch
no
l
ogy
. 200
6; 149
-155.
[12]
Khara
a
jo
o
MJ.
App
licati
on
of brai
n
e
m
otio
n
a
l l
ear
nin
g
b
a
s
ed
i
n
tel
lig
ent c
ontrol
l
er (BE
L
BIC) to activ
e
que
ue
ma
nag
e
m
ent. Internati
ona
l Conf
erenc
e on
Com
putati
ona
l Scienc
e. 200
4; 662-
665.
[13]
Z
i
y
a
ng Z
,
Da
o
bo W
,
Z
h
ishe
n
g
W
.
Flight si
mu
lati
on serv
o
system
base
d
on l
earn
i
n
g
b
a
sed
intel
l
i
gent
control
l
er
. Pro
c
eed
ings
of the F
ourth Int
e
rnati
ona
l Co
n
f
erence
on Im
puls
e
an
d H
y
brid D
y
n
a
m
i
c
S
y
stems. 20
07
; 1372-1
3
7
5
.
[14]
Hua
ng G, Z
hen Z
,
W
ang D.
Brain e
m
otion
a
l
lear
nin
g
bas
e
d
intel
lig
ent co
ntroll
er for nonl
ine
a
r system
.
Paper
pres
ent
ed
at Sec
ond
i
n
ternati
o
n
a
l s
y
mposi
u
m
o
n
i
n
tellig
ent
inform
ation
tech
nol
o
g
y
a
ppl
icatio
n.
200
8; 660-
663.
[15]
Mehra
b
ia
n AR
, Lucas C, Ro
shan
ian J. De
si
gn of an
aer
ospac
e la
unch
vehicl
e auto
p
i
l
ot base
d
o
n
optimiz
ed emot
ion
a
l le
arni
ng a
l
gorit
hm.
Cyber
netics an
d Systems
. 2
008; 3
9
(
3
): 284-3
03.
Evaluation Warning : The document was created with Spire.PDF for Python.