TELKOM
NIKA
, Vol.11, No
.3, March 2
0
1
3
, pp. 1707
~
1713
ISSN: 2302-4
046
1707
Re
cei
v
ed
Jun
e
17, 2012; Revi
sed Septe
m
ber
10, 201
2; Acce
pted
Octob
e
r 19, 2
012
Efficient RFID Data Cleaning Method
Li Xing, Fu
Wen
-
Xiu
Schoo
l of Elect
r
onic a
nd Infor
m
ation En
gi
ne
erin
g, Beiji
ng Ji
aoton
g Un
ivers
i
t
y
.
Beiji
ng, 10
00
4
4
, Chin
a
e-mail: li
xin
gbjt
u
@1
63.com
A
b
st
r
a
ct
RF
ID (Rad
io
F
r
equ
ency
Identific
atio
n) t
e
chn
o
lo
gy tra
n
sfers d
a
ta
b
e
tw
een
mova
ble
tag
g
e
d
obj
ects and re
aders w
i
thout li
ne of si
ght, an
d the captur
ed
data ten
d
s to
be noisy. The in
here
n
t unrel
ia
b
ilit
y
mak
e
s the d
a
t
a
unr
eli
abl
e to
appl
icati
on. N
o
w
adays, t
he
ma
in so
luti
on i
s
to use sli
d
in
g w
i
ndow
, but it is
difficult to deci
de the w
i
nd
ow
si
z
e
, es
pec
ial
l
y
w
hen the
tag mov
e
s frequ
en
tly or w
i
th high false pos
itive. T
o
solve th
e
me
ntion
ed
prob
le
ms, SW
KF
(Slidi
ng W
i
nd
ow
base
d
o
n
K
a
l
m
a
n
F
ilter
Pre-proc
essin
g
)
i
s
prop
osed. It pr
eproc
esses th
e
RF
ID data to
mak
e
the
rea
d
rate close to the real o
ne, d
e
tects and
filters
th
e
mo
bil
e
ta
gs. T
hen, t
he
pre
p
r
o
cesse
d
data
i
s
smooth
e
d
to
further i
m
prove
accur
a
cy. At t
he s
a
me
ti
me,
the
mi
d-w
i
nd
ow
sli
de p
o
int r
educ
es the o
u
tput.
T
h
roug
h t
he c
o
mbi
natio
n of
Kal
m
a
n
F
ilter
and s
lid
in
g w
i
n
dow
,
SW
KF
provide
s
accurate RF
ID data to ap
pli
c
ation.
Ke
y
w
ords
:
ra
dio freq
ue
ncy i
dentific
atio
n, Kal
m
a
n
filter, preproc
ess, slidi
ng w
i
ndow
1. Introduc
tion
Combi
ned
wi
th Internet and telecomm
unication, RF
ID techn
o
log
y
can achieve global
scale items tracking a
nd in
formation
sha
r
ing [1,2
]. With the increa
si
ng numb
e
r of
RFID data a
nd
the inherent
unrelia
bility [3], it is nece
s
sary
to
clea
n the co
llected d
a
ta to satisfy RFID
appli
c
ation [4
-8]. The e
r
rors o
c
curring i
n
the pro
c
e
s
s
of data capture oft
en in
clud
e false
neg
ative
and false positive. The observed r
ead ra
te in real-wo
r
l
d
RFID de
plo
y
ments is often in the 60
%-
70% ran
ge [3
,9], that is to
say, over 30
% of the tag readin
g
s a
r
e routinely dro
p
ped.
Many schol
ars have
wo
rke
d
on the p
r
ob
le
ms m
ention
ed above. T
h
e method
ba
sed
on
slidin
g wi
ndo
w [4,5,10,1
1
]
is the
typical
approa
ch, b
u
t
it is h
a
rd
to
deci
de th
e wi
ndo
w
size. With
lots of interfe
r
en
ce, espe
ci
ally tags in m
obile
conditio
n
that is hard
to be detect
ed, it increa
ses
the erro
r d
a
ta. Method
s
prop
osed i
n
[7,12] ca
n
correct th
e re
ad rate dyn
a
m
ically, but t
hey
prod
uce large
amounts of o
u
tput data.
In this pape
r, SWKF is presented. We
su
mmari
ze
ou
r contri
bution
s
as
follo
ws:
·
The combi
nation of Kalman Filter a
nd slidin
g wi
ndo
w. The Kalman Filter
pre
-
processin
g
corre
c
ts the
read rate to
so
me extent, removes t
he int
e
rferen
ce, de
tects an
d filters mo
bile tag
s
.
·
Smoothin
g
the
pre-pro
c
e
s
sed dat
a. It
avoids the
wi
ndo
w jitterin
g
phe
nome
non
. This i
m
proves
the effect of the slidi
ng wi
ndo
w processing. Th
e mi
d-wi
ndo
w
s
lid
e point de
cre
a
se
s the
storage
spa
c
e.
·
Experime
n
ts. They sh
ow that the accu
racy
an
d effectivene
ss of the pro
p
o
s
ed
algorith
m
.
The
re
st of t
he p
ape
r i
s
orga
nized
as fo
llows. Se
ction 2
prese
n
ts the
relat
ed
wo
rk.
Section 3 pre
s
ent
s the structure
of RFID data. Secti
on 4 pre
s
ent
s a detailed treatme
nt of the
clea
ning
met
hod. Se
ction
5 re
po
rts ex
perim
ental
a
nd pe
rforman
c
e
re
sults.
We co
ncl
ude
o
u
r
study in Secti
on 6.
2. Related Work
For
RIFD dat
a cle
anin
g
, Smooth i
s
first pro
p
o
s
ed
in
the EPC Gl
o
bal Read
er P
r
otocol
[10], the origi
nal purpo
se i
s
to rep
r
e
s
en
t a lar
ge nu
mber of tag
strea
m
event
s with the lo
gical
meanin
g
ful e
v
ents. But it actually play
s the rule
of
smoothing
eve
n
t strea
m
to clea
n the false
negative. Its wind
ow
size is fixed.
Jeffery p
r
op
ose
s
a
stati
s
tical
smo
o
thing al
gorith
m
SMURF [5], which m
odel
s the
unreli
ability of RFID readi
ngs
by viewi
ng RFI
D
st
reams as a
statistical
sampl
e
of tags in t
h
e
physi
cal
wo
rl
d. It adapt
s th
e wi
ndo
w
si
ze to p
r
ovide
a
c
curate
RFID data to
ap
pli
c
ation. But
if the
tags move
ra
pidly, it can’t addr
ess the
wind
ow
size. To solve thi
s
probl
em, Ling
yong Men
g
et
al.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 3, March 2
013 : 1707 –
1713
1708
[11] prop
ose
s
a n
e
w
a
nd imp
r
oved
algorith
m
,
whi
c
h
con
s
i
der
s f
a
ct
o
r
su
ch a
s
r
e
a
der
comm
uni
cati
on
rang
e, vel
o
city of tag
m
o
vement, a
n
d
re
ading
fre
q
uen
cy in
dete
rrin
g
the
si
ze
of
wind
ow.
J
e
ffrey
et al.
[4] present
s ESP, a declarative query-
based framework. ESP consi
s
ts of a
prog
ram
m
abl
e pipeline of
decl
a
rative q
uery-ba
s
ed
stage
s. These stage
s
segm
ent the clea
ning
proc
es
s into f
i
ve tas
k
s
.
Thes
e five approaches
have
inc
r
eas
i
ng levels
of func
tionality. ESP us
es
wind
owed proce
s
sing to
grou
p re
adin
g
within
a g
r
anule, cl
ean
s data based
on tempo
r
al
and
spatial
corre
c
tion, but it is
n’t related to how to
set it. It mainly solves false ne
g
a
tive and false
positive.
G
o
nz
a
l
ez
et
al.
[7] presents a ne
w
clea
ning met
hod ba
se
d o
n
Dynami
c
Bayesian
Networks
(DBNs). It corre
cts the tag re
ad rate
dyna
mically, and consi
der
s
the observation
s and
estimate
s, bu
t they are obtained by the
hist
ori
c
al d
a
ta, and ca
n’t be update
d
dynamically.
W
h
en
th
e mis
s
e
d
r
e
ad
r
a
te
is h
i
gh
, es
pe
c
i
a
lly
w
h
en
ta
g
s
mo
ve
r
apid
l
y, th
e
e
ffe
ct o
f
th
e
mentione
d al
gorithm
s ab
o
v
e is bad.
Wan
g
Yan
et al.
[12] propo
se
s a cl
eani
n
g
method tha
t
used Kalm
a
n
Filter, whi
c
h solve
s
false ne
gative and false p
o
sitive from si
ngle re
ade
rs, but need
s mu
ch spa
c
e to store the tag
s
.
PSCleani
ng based on pseudo
event
i
s
p
r
o
posed
in [13]. It re
d
u
ce
s th
e tim
e
del
ay of
data outp
u
t b
y
introdu
cin
g
the notio
n of
pseudo
eve
n
t into sli
d
ing
wind
ow,
and
de
cre
a
ses the
volume of o
u
t
put by han
dl
ing false p
o
sitive and d
u
p
licate
rea
d
ing
s
at the
sam
e
time. But it
doe
sn’t imp
r
ove the a
c
cu
racy. Th
ere, we p
r
e
s
ent
SWKF, whi
c
h co
mbine
s
Kalman Filte
r
with
sliding window.
3. RFID Data
RFID system
gene
rate
s
a stream
of data
that
re
sults f
r
om interrog
ation cy
cle
s
o
c
curri
ng
at recurring ti
me interval
s at each
Rea
der. The
rea
d
ing data g
e
nerate
d
in ea
ch inte
rro
gati
on
cycle i
s
usua
lly a set of tu
ples
of the fo
rm
(EP
C
, Re
ader, time
).
The tupl
e ca
n be
add
ed
with
extra informat
ion, such as the tag type (class
0, class 1, generation
2,
etc), the antenna u
s
ed
by
the Read
er,
o
r
the
po
we
r l
e
vel of th
e int
e
rrogatio
n
sig
nal. From th
e
appli
c
atio
n p
e
rspe
ctive, it
is
necessa
ry to look at multipl
e
in
terro
gatio
n cycle
s
as a
singl
e unit kn
own a
s
a rea
d
cycle. In su
ch
ca
se, we get
tuples of the form (EPC, Reade
r, time, respon
ses), whe
r
e resp
on
se
s is the
numbe
r of interrogatio
n cy
cle
s
wh
en the
tag was
read
.
Definition 1 (epo
ch) Th
at is, the read cy
cl
e mentio
ne
d above, a si
ngle unit. We
assume
that 10 interrogation
cycle
s
as a
singl
e unit.
Definition 2 (
p
i
) Read rate of the tag i in a epo
ch,
/1
0
i
r
e
sp
on
se
p
.
4. Cleaning
Metho
d
4.1. Kalman Filter Model
Kalman Filter con
s
i
s
ts of two p
a
rts: ti
m
e
update
and
measuremen
t update. Tim
e
updat
e
process esti
mates
the current
state utilizing
the optimal
valu
e of the time on a
state.
Measurement
update pro
c
e
ss uses
o
b
se
rvation
s
on
the curre
n
t
status of
amend
ment
s
to
update th
e estimates obtai
ned fro
m
the
previou
s
time
to get more
accurate e
s
timates, recy
cl
ing
to approxim
ate the true value.
Linea
r differe
ntial equatio
n
s
of Kalman
Filter:
Predi
ction eq
uation:
()
(
1
)
(
)
(
)
X
kA
X
k
B
U
k
W
k
Observation equatio
n:
()
()
()
Z
kH
X
k
V
k
()
X
k
is t
h
e
sy
st
e
m
st
at
e
at
t
i
me
k
, an
d
()
Uk
is the
system
control
at time
k
.
A
and
B
are the sy
stem para
m
et
ers
(
A
=
B
=1).
()
Z
k
is the measu
r
ed valu
e
at time
k
,
H
is the
measurement
syst
em para
m
eters
(
H
=1).
()
Wk
and
()
Vk
denote
the pro
c
e
s
s
and me
asure
m
ent
noise. They
are a
s
sum
e
d
to be Ga
ussian white noi
se, zero
mea
n
, and vari
an
ce, re
sp
ectiv
e
ly
are
Q
,
R
.
Firs
t,
we predic
t the read
rate
of RFID
in
the
next
cycle
by utili
zin
g
the
process mod
e
l.
The mo
st wid
e
ly used nu
m
e
rical predicti
on met
hod i
s
regressio
n
analysi
s
. We u
t
ilize the most
comm
only used method in
the reg
r
e
ssi
o
n
analysi
s
lea
s
t squ
a
re
s fit.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Efficient RFID Data
Clea
n
i
ng Method
(L
i Xing)
1709
Rea
d
rate fo
r n cycl
es u
s
e
s
line
a
r reg
r
e
ssi
on a
nalysi
s
of fitting a straig
ht line y
=
a
+
b
k
,
and b is a
s
fo
llows:
1
2
1
()
(
)
()
n
ii
i
n
i
i
x
xy
y
b
xx
(1)
(|
1
)
(
1
|
1
)
X
kk
X
k
k
b
t
(2)
In equatio
n 2
,
(|
1
)
Xk
k
rep
r
e
s
ent
s the optim
al
value at
k
-1,
t
re
pre
s
e
n
ts
a re
ad
cycle.
Rea
d
rate pre
d
ictio
n
s
have
been
update
d
, but
covari
an
ce of
(|
1
)
Xk
k
is not u
pdat
ed.
P
denote
s
the covarian
ce:
(|
1
)
(
1
|
1
)
Pk
k
P
k
k
A
Q
(3)
In equ
ation
3,
(|
1
)
Pk
k
is t
h
e
cov
a
rian
ce
of
(|
1
)
Xk
k
,
(1
|
1
)
Pk
k
is t
h
e
cov
a
ri
an
ce of
(1
|
1
)
Xk
k
,
A
is the tran
spo
s
e
d
matri
x
of
A
(
A
=1),
Q
is the cova
rian
ce of the
system p
r
o
c
e
ss. Equ
a
tion
2, 3 forecast
the sy
stem.
With RFI
D
re
ad rate p
r
edi
ction
s
, and th
en
colle
ct the o
b
s
erve
d rate o
f
the RFID
re
ader. T
hen,
we
can
get t
he optimal
e
s
timate of cu
rrent
(k) RFI
D
rea
d
rate from the
predi
ction an
d measureme
n
t.
(|)
(
|
1
)
(
)
(
(
)
(|
1
)
)
X
k
k
X
kk
K
g
k
Z
k
X
kk
(4)
K
g
is the Kalma
n
gain
:
(
)
(
|
1)
/
(
(
|
1)
)
K
gk
P
k
k
P
k
k
H
R
(5)
No
w, we
ha
s got the o
p
timal RFI
D
re
ad rate
(|)
X
kk
at time
k
.
Ho
wev
e
r, in o
r
de
r to
make th
e Ka
lman Filter
consta
ntly run
n
ing u
n
t
il the
end of the
system
pro
c
e
ss,
we n
eed
to
update the
co
varian
ce of
(|)
X
kk
at
k
.
(|)
(
(
)
)
(
|
1
)
Pk
k
I
K
g
k
H
P
k
k
(6)
Whe
n
t
h
e
sy
st
em st
ep
s
i
n
t
o
k
+1
state,
(|
)
Pk
k
is
the formula (3)
(1
|
1
)
Pk
k
. In this
way, the algo
rithm co
uld g
o
autore
g
ressive operato
r
.
4.2. Sliding
Windo
w
Mo
del
It uses ad
ap
ted slidin
g wi
ndo
w, the ke
y is the deci
s
ion of the slid
ing win
d
o
w
si
ze. For
each tag, it views e
a
ch e
poch as a
n
i
ndep
ende
nt Berno
u
lli trail
(|
|
,
)
i
BW
p
, where |
W
| is
the
wind
ow si
ze.
If the
tag
i
is
read, an
d ap
pears in
W
, then tag
i
meets:
iS
and
SW
, wh
ere
W
denote
s
th
e
set,
av
g
p
den
otes the e
s
timation
of
i
p
cal
c
ulate
d
ba
se
d o
n
al
l the e
p
o
c
h
report
s
i
n
the wind
ow, that is
()
||
i
av
g
iS
p
p
S
(7)
The value |
S
| follows Bernoulli distribution,
||
(
|
|
,
)
av
g
SB
W
p
.
Based o
n
Bernoulli mod
e
l, if the average
read rate is
av
g
p
for each tag in each epo
ch
in
W
, the
n
the
prob
ability th
at we
miss
a
readi
ng from
tag
i
ove
r
W
is
(1
)
av
g
W
p
. Setting the leas
t
probability for tag been observ
ed in the
window to
be
, then the p
r
o
bability to en
sure that tag
i
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 3, March 2
013 : 1707 –
1713
1710
been ob
serv
ed
is
1
, that is
(1
)
av
g
W
p
. So the lea
s
t size
of win
dow i
s
*
1
ln
(
)
/
av
g
wp
,
whi
c
h gu
ara
n
t
ees complet
ene
ss.
Becau
s
e the
mobile tag
s
are dete
c
ted
in the prep
ro
ce
ssi
ng, we j
u
st nee
d to focu
s on
compl
e
tene
ss. In ad
dition,
*
w
is
s
e
t to infinite when
0
avg
p
. This is not p
e
rm
itted, so
we
keep
the win
d
o
w
size u
n
chan
ge
d. It is worth
y
to m
ention
that it adopt
s a mid
-
wi
ndo
w
slide
point.
In
other wo
rd
s, it
prod
uces
readi
n
g
s
with
an e
p
o
c
h v
a
lue
co
rre
sp
ondin
g
to the
midpoi
nt of
the
wind
ow. Thi
s
greatly re
du
ces the outp
u
t data.
4.3. SWKF
The Kalm
an
Filter an
d
sliding
win
d
o
w
a
r
e
introd
uce
d
a
bove.
In the
Kalm
an Filte
r
model,
whe
n
the misse
d
re
ad rate i
s
hi
g
h
, the p
r
edi
cti
on a
c
curacy
will b
e
g
r
eatly
red
u
ced. T
h
en
just relying
o
n
the Kalman
Filter doe
s n
o
t corre
c
tly reflect the re
al
data (see fro
m
figure 2
)
, the
clea
ning effe
ct is bad. Be
sides, it prod
u
c
e
s
larg
e am
ounts of outp
u
t data.
For the sli
d
in
g wind
ow, th
e tags that are det
ecte
d far away from t
he rea
d
e
r
wit
h
a low-
probability wi
ll force it to
use
larger
wi
ndow to gua
rantee compl
e
teness.
But
they can
cause
probl
em
s
in
environ
ment whe
r
e
tag
s
a
r
e
m
obile.
F
o
r
i
n
the multi-tag ca
se, a similar re
adi
ng
results in
an
overly l
a
rg
e
co
ntrib
u
tion
to the
ove
r
all count
esti
mate, an
d th
us
a l
a
rg
e o
v
er-
estimation
error. Also,
wh
e
n
the mi
ssed
read
rate
is
hi
gh, the
clea
ni
ng effect i
s
b
ad. Algorith
m
1
is the pseud
o
-
co
de.
Algorithm 1 S
W
KF
Require: Co
mpleteness confidence
δ
1:
1
w
2: While (
()
g
et
Nex
t
Epoch
) do
3:
_
()
kal
p
r
o
c
e
ss
4:
()
processWi
n
dow
W
5:
*
(,
)
av
g
w
c
om
pl
et
eSi
z
e
p
6: if (
*
ww
) then
7:
*
ma
x
{
mi
n
{
2,
}
,
1
}
ww
w
8: end if
9: end
w
h
ile
SWKF ru
ns
a slidin
g-win
dow a
g
g
r
ega
te for each o
b
se
rved tag i
.
The wind
o
w
si
ze i
s
initially set to
one
epo
ch,
and the
n
a
d
j
u
sted
dynam
i
c
ally ba
se
d o
n
ob
se
rved
readin
g
s.
Du
ri
ng
each epo
ch, for tag i, it utilize
s
Kalman F
ilter for preproce
s
sing (
_
()
ka
l
p
r
o
ce
s
s
), whi
c
h re
mov
e
s
interferen
ce t
o
som
e
exte
nt, detects
a
nd filt
ers th
e
mobile tag
s
.
After prep
ro
ce
ssi
ng, SWKF
pro
c
e
s
ses th
e rea
d
ing of
tag i insid
e
the win
d
o
w
W. The p
r
o
c
essing in
clu
d
e
s e
s
timating
the
requi
re
d mod
e
l para
m
eters for tag i (e.
g
.,
av
g
p
) as
well
as emitting a
n
output re
ad
ing for tag i if
there exits
at
least rea
d
ing
within
th
e win
dow. Then,
S
W
KF con
s
ult
s
it
bi
nomial
-
sampli
ng mo
del
to determi
ne
the num
ber
o
f
epochs
nee
ded to g
u
a
r
a
n
tee complet
ene
ss
(
(,
)
av
g
c
o
mple
te
Size
p
).
If the re
qui
re
d
*
w
exce
eds t
he
cu
rre
nt windo
w
size
||
wW
,
S
W
K
F
inc
r
e
a
se
s t
h
e
cu
r
r
ent
wind
ow size
(
*
m
a
x{m
i
n{
2
,
},
1
}
ww
).
5. Results a
nd Analy
s
is
5.1. Experimental Setup
In our expe
riments, the
hard
w
a
r
e e
n
v
ir
onme
n
t: 2.61G Athlon
dual-co
r
e
CPU, 1G
Memory, 32
0
G
Ha
rd Di
sk. The software envir
o
n
me
nt: Windo
ws
XP operatin
g
system, Ora
c
le
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Efficient RFID Data
Clea
n
i
ng Method
(L
i Xing)
1711
11g Ente
rp
rise Edition, Ma
tlab and P
L
/SQL la
n
guag
e. Experime
n
t
s are de
sig
n
ed to me
asure
the algorith
m
s SMURF, KAL_RFI
D pro
posed in
[6] and SWKF pro
posed in this
pape
r.
In orde
r to be more
conv
incin
g
, we collect data in
the physical
environm
ent
(Tige
r
RF10
01 UHF
Read
er) for 150 re
ad cy
cl
es.
5.2. Ev
aluation Model
In this paper,
we asse
ss t
he sp
ace
co
st and accuracy, and give th
e stand
ard d
e
finition
of accura
cy.
Definition
3 (Accu
ra
cy)
Gi
ven two
data
set
s
, the rea
l
data
set
r
D
and
clea
ned
dat
a set
c
D
. In a time period T, the accura
cy of the data can b
e
expre
s
sed a
s
:
(
)
()
()
/
(
)
Ar
c
r
PT
D
T
DT
DT
(8
)
5.3. Ev
aluation Model
This se
ction analyses
the rea
s
on
able
n
e
ss
a
nd
t
he
cleanin
g
effect of SWKF. Firstly, the
rea
s
on
able
n
e
ss i
s
prese
n
ted. Lea
st squ
a
re
s es
tim
a
tion is u
s
ed in
Kalman Filter predi
ction.
Figure1 Ideal,
Kalman filter
Figure2 Pre
p
roce
ss insta
n
ce in the real
-worl
d
environ
ment
Figure 1
sho
w
s, in th
e ide
a
l ca
se, the l
east
squ
a
re
s
method m
a
ke
s the d
a
ta clo
s
e to the
real. In
ord
e
r
to be
clo
s
e
r
to re
ality, we
prep
ro
ce
ss th
e collecte
d
d
a
ta in
real
-wo
r
ld e
n
viron
m
e
n
t
mentione
d a
bove. Figu
re
2 sho
w
s th
at the
lea
s
t
sq
uares m
e
thod
can
correct
rea
d
rate
(re
sp
on
se/10
)
to
a
ce
rtain
extent, but th
e read
lea
k
a
ge
rate i
s
still
large, that i
s
to
say, the
d
a
ta
need
s to
be
furthe
r p
r
o
c
e
s
sed.
The
Kal
m
an Filte
r
p
r
epro
c
e
s
sing
can
effectivel
y detect
mob
ile
tags, and
slidi
ng wi
ndow i
s
difficult
to do
this. However, the ability to
detect m
obil
e
tags is
a very
importa
nt fact
or that affe
cts the
accu
ra
cy of slidin
g win
dow. So the
combinatio
n of
Kalman Filte
r
and sli
d
ing wi
ndo
w is a workabl
e option,
whi
c
h indi
cat
e
s that SWK
F
is rea
s
o
nab
le.
In orde
r to verify the accura
cy of SWKF
, it
is com
pare
d
with KAL_RFI
D an
d SMURF.
Figure 3
sho
w
s th
at the
accuracy of
SWKF is
hig
her tha
n
that
of SMURF
and KAL_
R
F
I
D.
Although the
read le
akage
rate incre
a
se
s,
the accu
ra
cy of SWKF is improved.
To sh
ow th
e
accuracy of t
he algo
rithm
more
effectively, we cle
a
n
the data coll
ected i
n
real
appli
c
ati
on e
n
viron
m
ent with
SWKF. The
re
su
lts a
r
e
sh
own
in Fi
gure
4.
In the
10th
cycle,
SMURF
ha
s false
negative
.
In the 60th and 10
0th cy
cl
e, tag
s
mov
e
dynami
c
ally
, the read
rat
e
is
redu
ce
d to 0, SMURF in
creases the
windo
w si
ze
, and ca
uses fal
s
e po
sitive. During 1
30 to 140
cycle
s
, the
wi
ndo
w si
ze
of SMURF
chan
ges f
r
eq
uent
l
y
, and ca
use
s
la
rge
errors. In the first 1
30
cycle
s
, KAL_
R
FID
han
dle
s
bette
r, but i
t
has fal
s
e
p
o
sitive an
d n
egative du
rin
g
130
-1
40
cycle
s
whe
n
the tags move fre
quently (read
rate is
abo
ut 0.3). Because of the prep
ro
ce
ssin
g of
Kalman Filter, the read ra
te is clo
s
e to
the tr
ue in the first 13
0 cycle
s
for S
W
KF. After the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 3, March 2
013 : 1707 –
1713
1712
smoothi
ng
of slidin
g
wind
ow, it works
well. But
it al
so
ha
s some
false
du
ring
130-140
cy
cl
es.
Ho
wever, co
mpared to KAL_RFI
D, it has so
me impro
v
ement.
Figure 3. Accura
cy
Figure 4. Cle
aning in
stan
ce in the real
-worl
d
environ
ment
5.4. Space Cost Analy
s
is
This sectio
n analyzes the
spa
c
e
co
st of t
he algorith
m
. Figure 5 s
hows that the number
of data o
u
tp
ut and i
nput
is alm
o
st th
e
sam
e
for KAL_RFI
D in
150
rea
d
cy
cles. Be
cau
s
e
the
prep
ro
ce
ssin
g filters the
mobile tag
s
, and then e
ffe
ctively preve
n
ts the su
dd
en increa
se
of the
slidin
g wi
ndo
w, so
the
sp
ace
co
st for
SWKF is
larg
er tha
n
SMURF. When th
e num
ber of
data
increa
se
s or t
he tag moves frequently, SWKF will ne
e
d
much m
o
re
spa
c
e tha
n
SMURF.
Figure 5. Space Perfo
r
ma
n
c
e
6. Conclusio
n
RFID te
chn
o
l
ogy has
a
wide a
ppli
c
at
ion, and al
so rai
s
ed n
e
w
challen
g
e
s
to data
clea
ning, wh
ich ha
s be
come an im
portant fi
eld
of RFID rese
arch. According to the
cha
r
a
c
teri
stics of RFI
D
da
ta, we propo
se a
dat
a cl
e
aning al
go
rith
m, which utili
ze
s the Kalm
an
Filter to
pre
p
ro
ce
ss the
RFID data,
and th
en
wit
h
sli
d
ing
wi
n
dow for furt
her processi
ng.
Experiment
s sho
w
that it has a go
od cl
e
aning effe
ct in compl
e
x en
vironme
n
t.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Efficient RFID Data
Clea
n
i
ng Method
(L
i Xing)
1713
Referen
ces
[1]
W
ang Sha
o
-h
ui, Liu Su
jua
n
. Efficient Passive F
u
ll-
di
sclosur
e Attack on RF
ID Light-
w
e
i
g
h
t
Authentic
atio
n
Protocols L
M
AP+
+
AND SUAP.
T
E
LKOMNIKA Indones
ian J
ourn
a
l of Electric
a
l
Engi
neer
in
g
. 2012; 10(
6): 145
8-14
64.
[2]
Is
w
a
nj
ono,
Ba
gio
B, Ka
lamu
llah
R. A
n
gl
g
o
rithm for
pr
e
d
ictin
g
th
e sp
eed
of traffic
l
i
ght v
i
ol
ators.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2011; 9(1): 5
5
-
64.
[3]
Barna
b
y
J. Fed
e
r. Despite W
a
l-Mart’s Edict, Radi
o T
ags W
ill T
a
ke
T
i
me. N
e
w
Y
o
rk T
i
mes. 2004.
[4]
Jeffre
y
SR, Al
onso G, F
r
anklin MJ, Hong W
,
W
i
dom J.
A pipel
in
ed fra
m
e
w
ork for on line clea
nin
g
of
sensor
data
strea
m
s
. In: L
i
u
L
,
Reuter
A, et a
l
., eds. Proc.
of the
22n
d Int’
l
Conf. o
n
D
a
ta
Engi
neer
in
g
.
Atlanta: IEEE Computer
Societ
y
,
2006. 140
−
142
[5] Jeffery
SR,
Garofalak
i
s M, Franklin MJ.
Ad
a
p
tive cl
ea
nin
g
for RF
ID data
s
t
reams
. In: Day
a
l U, Whang
KY, et al., eds.
Proc. of the 32
nd Int’l Co
nf. on
Ver
y
Larg
e
D
a
ta Bases. Seo
u
l: ACM, 2006.
163
−
17
4.
[6]
Bai Y, W
a
n
g
F
S
, Liu PY. E
fficientl
y
F
ilter
i
ng RF
ID data streams. In: Proc. of the 1st Int’l VLDB
W
o
rkshop o
n
Clea
n
Data
bas
es. Seoul: Mor
gan Ka
ufman
n
Publ
ishers, 20
06. 50
−
57.
[7]
Gonzal
ez H, H
an J, Sh
en
X.
Cost-Co
n
scio
u
s
clea
ni
ng
of
massive
RFID data sets
. In: Proc. of the
23r
d
Int’l Conf. on D
a
ta Engi
ne
erin
g. Ist
anbul: IEEE Computer S
o
ciet
y
,
200
7. 1
268
−
12
72
.
[8]
Khouss
a
in
ova
N, Balazi
nska
M, Suciu D.
T
o
w
a
rds correct
ing i
n
p
u
t data
errors pro
b
a
b
il
i
s
tically us
in
g
integr
ity constr
aints
. In: Chr
y
s
anthis
RK. Pro
c
. of the 5th
A
C
M Int’l W
o
rks
hop
on
Data
E
ngi
neer
in
g for
W
i
reless a
nd
Mobil
e
Access.
Chica
go: ACM
,
2006. 43
−
50
.
[9]
S. R. Jeffery
,
G. Alonso, M.
J.
F
r
anklin, W
.
Hon
g
, an
d J.
W
i
dom. Dec
l
ar
ative Su
pp
ort for Sens
or D
a
ta
Clea
n
i
ng. In Pervasiv
e, 200
6.
[10]
EPCGlob
al Re
ader Protoc
ol
Standar
d
1.1. http://
w
w
w
.
EP
Cglo
ba
linc.or
g.
2006.
[11]
Lin
g
y
on
g Me
n
g
, F
engq
i Yu.
RF
ID Data
Clea
n
i
ng B
a
se
d on A
d
a
p
tive
W
i
ndow
. Pro
c
of the 2n
d
Internatio
na
l C
onfere
n
ce o
n
F
u
ture Com
pute
r
and Comm
un
icatio
n. W
uhan
, China, 20
10,
746-
749.
[12] W
ang
Y
a
n
,
S
ong
Bao-
ya
n.
Clea
n
i
ng M
e
th
od
of RF
ID
Da
ta Stream B
a
s
ed
on K
a
lma
n
F
ilter.
Journ
a
l
of Chin
ese Co
mp
uter Syste
m
s
. 2011; 32(
9): 179
4-17
99.
[13] W
ang
Y
a
n
,
S
h
i
Xin, S
ong
B
ao-
ya
n. RF
ID
Data
Cle
an
ing
Method
Bas
e
d
on Ps
eu
do
Ev
ent. Jour
na
l of
Comp
uter Res
earch a
nd D
e
v
e
lo
pment. 20
0
9
; 46(Sup
p
l): 2
70-2
74.
Evaluation Warning : The document was created with Spire.PDF for Python.