TELKOM
NIKA
, Vol. 11, No. 10, Octobe
r 2013, pp. 5
911 ~ 5
916
ISSN: 2302-4
046
5911
Re
cei
v
ed Ap
ril 23, 2013; Revi
sed
Jul
y
6, 2013; Accept
ed Jul
y
18, 2
013
Feature Extraction of Turing Tool Wear
Based on J-EEMD
Hong
tao
Ch
en*, Pan Fu, Xiaohui Li
Institute of Mechan
ical En
gi
ne
erin
g of South
w
e
s
t Jiaoto
ng
Univers
i
t
y
, C
h
e
ngd
u 61
003
1, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: scd
y
c
ht@
1
6
3
.com
A
b
st
r
a
ct
In the mo
nitori
ng of cutting to
ol state, a larg
e nu
mb
er of re
dun
da
nt infor
m
at
ion is co
ntai
n
ed in th
e
sensor si
gna
l. T
herefore, it is obvi
ous
ly n
o
t cond
ucive to p
a
ttern recog
n
iti
o
n, and diffic
u
lt to classify the to
o
l
w
ear state correctly from the
avail
a
b
l
e se
ns
ors. T
he test platform that
ha
d real-ti
m
e info
rmati
on co
llecti
o
n
of the vibr
atio
n an
d ac
ousti
c emissio
n
si
g
nals
in turn
in
g
w
a
s built. Ob
served s
i
gn
als
w
e
re ada
ptiv
ely
process
ed usi
ng the metho
d
of ens
embl
e empiric
a
l
mod
e
deco
m
positi
on intro
duce
d
joint a
pprox
i
m
at
e
dia
gon
ali
z
a
t
io
n
of eig
e
n
m
atric
e
s (J-EEMD). T
h
is metho
d
is
base
d
o
n
the
character
i
stics
of the si
gna
l its
e
lf
deco
m
pose
d
i
n
to sever
a
l
intri
n
sic
mo
de
func
tions (IMF
), an
d the
n
transfor
m
s th
e e
ner
gy
ratio b
e
tw
een t
h
e
IMF
.
T
he w
h
ite no
ise
of e
a
ch
IMF
comp
on
en
t has
be
en
eli
m
i
nate
d
by
i
n
troduc
ing
JADE
alg
o
rith
m
duri
n
g
the sign
al d
e
co
mp
ositi
on. Co
mp
are
d
w
i
th the EEMD al
gorit
hm, the
deco
m
pos
iti
on effici
e
n
cy is sig
n
ifica
n
tl
y
improve
d
. T
h
e
exp
e
ri
me
nts
show
ed th
at the
met
hod
co
ul
d
id
entify th
e differ
ent stat
es of to
ol w
e
a
r
, if
app
lie
d to feature extractio
n
of
vibratio
n an
d acoustic e
m
i
ssion si
gna
l in
the cutting pr
o
c
ess.
Ke
y
w
ords
: En
sembl
e
e
m
pir
i
c
a
l mod
e
dec
o
m
p
o
siti
on; T
ool
w
ear; F
eature extraction; T
u
ri
ng
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
The em
piri
cal
mode
deco
m
positio
n (E
MD)
pro
p
o
s
e
d
by Hua
ng
e
t
al. (199
8) i
s
a time-
freque
ncy an
alysis meth
o
d
[1-2]. It
is base
d
on the time scale of the local sig
n
a
l cha
r
a
c
teri
stics,
and the
si
gn
al is
de
comp
ose
d
into th
e
su
perpo
si
tio
n
of a
numb
e
r of i
n
trin
sic mode
fun
c
tions
(IMF). Certai
n signifi
can
c
e band i
n
formation is
con
t
ained in e
a
ch IMF explod
ed. EMD met
hod
has be
en wi
d
e
ly used in many fields, but when ju
m
p
ing ch
ang
es exist in the time scale of the
sign
al, this
case th
ere will
be an
IMF compon
ent
co
ntains
differe
nt time scale
s
cha
r
a
c
teri
stic
comp
one
nts,
whe
n
the EM
D de
com
p
o
s
ition is
ca
rri
e
d
out on the
sig
nal. The m
a
jo
r drawba
cks
of
the
o
r
iginal
E
M
D are
the mode mixing probl
em whi
c
h is the
co
nseco
n
ce of
sig
nal inte
rmitte
nce
[3]. Because of this, so its appli
c
ation h
a
s be
en limited.
In orde
r to
solve the m
ode mixing
probl
em in t
he process
of traditional
EMD
decompo
sitio
n
method, the coll
ection
of
empirica
l mode de
compo
s
ition (EEMD) meth
od
prop
osed by Hua
ng
et al, based
o
n
n
u
m
ero
u
s
stu
d
i
e
s
statisti
cal
prop
ertie
s
of
the EMD of white
noise
re
sults.
Gau
s
sian
white noi
se
is
adde
d in
the
decompo
sitio
n
p
r
o
c
e
s
s in
EEMD, an
d t
h
e
combi
nation
of the sig
nal
and the
noi
se a
s
a
wh
o
l
e. Time-freq
uen
cy sp
ace
is divided i
n
to
different
scal
es i
ngredie
n
t
by filter ba
nk,
wh
en
the
ad
ditional
white
noise is unifo
rmly di
stribut
ed
in the entire t
i
me-fre
que
ncy space. Different
scal
e
si
gnal area is
automatically mapped to t
he
approp
riate scale ba
ckgro
und white
noi
se.
At
this
ti
me, White n
o
i
se of
zero m
ean i
s
ad
ded
to
each imf, and the noise will ca
ncel e
a
ch oth
e
r
af
ter many times the ave
r
a
ge cal
c
ul
atio
n
pro
c
e
ssi
ng. T
herefo
r
e, th
e
integrate
d
av
erag
e time
s
a
r
e m
o
re,
the
result
whi
c
h
o
b
tains is cl
oser
the prima
r
y si
gnal [4].
Modal alia
sin
g
probl
em
s o
f
traditional EMD is solved throu
gh E
E
MD, but in orde
r to
eliminate
th
e influen
ce of white noi
se a
dded
in
th
e d
e
com
p
o
s
ition
process
of th
e o
r
iginal
si
g
nal,
the integ
r
ate
d
average
at
lea
s
t 100
times i
s
re
quired. The
com
putational
efficien
cy i
s
cle
a
rly
redu
ce
d, so that it is not condu
cive to o
n
line
monito
ri
ng of tool we
ar. To solve t
h
is proble
m
, J-
EEMD de
co
mpositio
n al
gorithm
ha
s been
pro
p
o
s
ed.
Joint a
pproxim
ate d
i
agon
alizatio
n
of
eigenm
atri
ce
s (JADE
)
alg
o
rithm i
s
introdu
ced i
n
or
der to
elimin
ate the imp
a
c
t of white
no
ise in
IMF com
pon
ents, thu
s
weakenin
g
the
impa
ct ca
u
s
ed by the
la
ck of an
inte
grated
avera
g
e
number [5], [
6
].
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 591
1 –
5916
5912
2. Test Platform and Program
2.1. Test Platform
In this
study, tool we
ar
condition m
oni
tori
ng system
wa
s
built, which wa
s cap
able
of
monitori
ng cu
tting force, vibration a
nd a
c
ou
stic em
i
ssion sig
nal of CNC turni
ng
pro
c
e
ss in
re
al-
time. As
sho
w
n i
n
Fi
g.1, the te
st eq
uip
m
ent
a
nd i
n
st
rume
nts i
n
cl
u
ded:
CK614
3
/
100
CNC l
a
the,
Kistler9
257B
dynamom
eter, 8702B50M
1
K-Shear
cera
mic a
c
cele
rometer, 81
52
B12SP acou
stic
emission
sen
s
ors, DE
WE-3021
di
gital
acq
u
isitio
n
system, an
d
so on.
The
bl
ank mate
rial
and
blade
were re
spe
c
tively the austeniti
c st
ainle
ss
steel
304L a
nd Ke
nna KC5
010
turning in
se
rt. In
particula
r, three
types of bla
d
e
s
incl
uded
CNM
G12
040
4FP, CNM
G
120
408FP
and
CNM
G12
041
2FP we
re used.
Figure 1. Test platform
2.2. Test Pro
g
ram
The uniform desi
gn is one of space filling desig
ns and it seeks experimental poi
nts to b
e
uniformly sca
ttered on the
domain. Thi
s
met
hod
wa
s prop
osed by
Fang a
nd wa
ng in 197
8 a
nd
had bee
n
p
o
pularly
u
s
ed
sin
c
e 198
0
[7
].
Test of
cutting con
d
ition
s
we
re de
sig
ned acco
rdin
g
to
the uniform d
e
sig
n
method
, as sho
w
n in
table 1.
Table 1. CNC Turnin
g Test
Cutting Co
nd
itions
Grou
p
number
Cutting Speed
(m/mi
n
）
Cutting depth
(mm)
Feed ra
te
(mm/r)
Edge radius
(mm)
1 190
2
0.13
0.8
2 170
1.2
0.25
0.4
3 200
1
0.21
1.2
4 250
2
0.25
0.8
5 180
1.6
0.29
1.2
6 210
1.8
0.33
0.4
7 270
1.8
0.21
1.2
8 240
1.4
0.13
1.2
9 220
1.6
0.17
0.4
10 260
1.2
0.17
0.4
11 230
1
0.33
0.8
12 280
1.4
0.29
0.8
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion of Turi
n
g
Tool Wear
Based o
n
J-EEMD (Hongta
o
Che
n
)
5913
3. The JADE algorithm
JADE i
s
a
n
a
l
gorithm
that
use
s
signifi
ca
nt
eigen
pai
rs
of the
cumul
a
nt tensor to
find o
u
t
the estimate
d value
s
of
indep
ende
nt
comp
one
nts.
In this alg
o
r
ithm the te
nso
r
eig
enva
l
ue
decompo
sitio
n
is
co
nsi
dered a
s
mo
re
of a pr
ep
ro
ce
ssi
ng
step. E
i
genvalue
de
comp
ositio
n
can
be viewed a
s
diago
nali
z
ation. This
can find a
sep
a
ratin
g
matrix suited
to blind so
urce
s
e
pa
r
a
tion
.
Blind sep
a
ration of mixed model can be
expresse
d a
s
:
As
x
(1)
Whe
r
e:
x
: obs
ervation matrix,
T
n
x
x
x
x
,
,
,
2
1
A
:
n
n
unkno
wn n
onsi
ngula
r
mi
xing matrix
s
:
n
unkno
wn,
statistically indepe
ndent
so
urce sig
nal
s,
T
n
s
s
s
s
,
,
,
2
1
The gole of
the Blind Source Separat
i
on (
BSS) is c
o
ns
is
ts
of
recovering the s
e
t
of
sou
r
ce sig
nal
s S solely fro
m
the obse
r
v
ed (i
n
s
tantan
eou
s and lin
ear) ixtures
X, by estimating
either the mix
i
ng matrix A or its inverse
1
A
V
(silently assuming that A is invertible).
The ess
e
nc
e
of the BSS is to find a
s
eparating
matrix
W, mak
i
ng
the obs
e
rved
s
i
gnal x
by the W tran
sform, the out
put y as t
he estimate of the
source
sign
a
l
s.
Wx
y
(2)
Whe
r
e ea
ch
comp
one
nt of the
x
and
y
are mutually indepe
ndent.
First, a white
n
ing matrix should be
co
mput
ed acco
rding to the o
b
se
rvation m
a
trix.
VAs
Vx
z
(3)
Whe
r
e:
z
: whitened si
gnal matrix,
T
n
z
z
z
z
,
,
,
2
1
V
: whitening
matrix
The
JADE
al
gorithm
is ba
sed
on
the u
s
age
of t
he fo
rth ord
e
r cum
u
lant ten
s
o
r
; i
t
tries to
maximiz
e
[6], [8].
2
)
(
(
)
(
i
T
i
A
M
AF
diag
A
J
(4)
whe
r
e
A
is the
whiten
ed mi
xing matrix a
nd
i
M
are the
eigenm
atri
ce
s of the fou
r
t
h
orde
r cumul
a
nt tenso
r
(th
e
eigenval
ue
s of thos
e ei
genmat
rices
are the
ku
rto
s
is valu
es
of the
indep
ende
nt
comp
one
nt,
whi
c
h
ca
n b
e
cal
c
ul
ated
in
this way). T
he
startin
g
p
o
int of th
e
JADE
algorithm is
t
hat the
requirement of the
mos
t
BSS
algorithms
to
c
a
lc
ulate the
distributions
of t
he
indep
ende
nt
comp
one
nts,
can
be
fulfille
d by o
p
timizi
ng the
cumul
ant ap
proxim
ations of d
a
ta
[9].
The main a
d
vantage of the
fourth order
cumulant
s
is that they can
be optimized
by means of t
h
e
iterative Ja
co
bi algorith
m
[10].
In s
u
mmary, J
-
EEMD algorith
m process is as follows:
(1)
White noi
se of 0.4 stan
dard d
e
viatio
n is add
ed to the origin
al si
gnal
)
(
t
x
.
(2) T
w
o
signa
ls add
ed White noise sig
n
a
l
and White
n
o
ise
sign
al are Processe
d respe
c
tively by
EMD algo
rith
m, the result
s are re
co
rde
d
as IMF1 and
IMF2.
(3) Th
e o
b
servational m
a
tri
x
is
com
p
o
s
e
d
of th
e
co
rre
spo
ndin
g
co
mpone
nt
with
IMF1
and
IM
F2,
usin
g JADE algorith
m
to blind so
urce sep
a
rati
o
n
, thus, IMF
0
i
s
obtaine
d
sign
al compo
nen
t
extracted fro
m
the results.
(4) Ea
ch row of compo
n
e
n
ts of IMF0 is scale
d
in a
c
cord
an
ce wi
th each
ro
w of comp
onen
ts of
IMF1, obtaini
ng the ne
w IMF com
pone
nt.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 591
1 –
5916
5914
(5) Add
ed a
new
seq
uen
ce of white no
ise ea
ch
time
, repeating th
e above four
step
s, the final
decompo
sitio
n
result is ea
ch IM
F de
co
mpositio
n integratio
n average.
4. Featur
e Extra
c
tion
Ba
sed on J-EE
MD
4.1. Featur
e-Value De
ter
m
ination
As we
all kn
o
w
that re
dund
ant informatio
n is
contain
e
d in the ori
g
in
al sign
al colle
cted by
the se
nsor. A
nd having
a strong
ran
dom
ness, so it is
difficult to find a variatio
n
of the tool we
ar
state di
re
ctly
from the
raw
data. Th
erefo
r
e, it i
s
n
e
cessary
to a
naly
z
e th
e
ra
w d
a
ta u
s
ing
si
g
nal
pro
c
e
ssi
ng tech
niqu
es, a
nd to extract
more re
acti
ve characte
ri
stics of tool wea
r
. This t
e
st
extract the si
gnal en
ergy a
nd gravity fre
quen
cy as a
sign
al ch
ara
c
teristic valu
e.
Signal e
nerg
y
cha
r
a
c
teri
ze the d
egree
of si
g
nal
strength. And
reflec the
su
m of the
sign
al en
ergy in e
a
ch fre
q
uen
cy. Set
i
p
the
spe
c
tral
a
m
plitude
se
q
uen
ce of
origi
nal si
gnal
by
fast fourie
r transfo
rm (F
FT
), then the sig
nal ene
rgy
p
ca
n be cal
c
ul
ated as follo
ws:
n
i
i
p
p
2
(5)
The g
r
avity frequ
en
cy
c
f
respon
se of
a
chang
e of p
o
s
iti
on of
the spe
c
tral
cent
roid.
Different to
ol
wea
r
stat
e ca
use
d
chan
ge
s in the
vib
r
at
ion amplitu
d
e
of som
e
freq
uen
cy, thus t
h
e
positio
n of sp
ectral
ce
nter
of gravity are
affected
to a
large extent.
The calculation formul
a is
as
follows
:
n
i
i
n
i
i
i
c
p
p
f
f
1
1
(6)
4.2. Feature
Data Extr
acti
on
The process
of J-EEMD al
gorithm to extract
the
signa
l characte
ri
stics a
r
e a
s
follows:
(1) Int
r
od
uce
a white
noi
se
with 0.4 va
ria
n
ce, ta
ke the
numbe
r of int
egrate
d
ave
r
age
with10, a
n
d
sign
als of vibration an
d acousti
c emissi
on are p
r
o
c
e
s
sed
with J-EE
MD.
(2) F
FT is
a
pplied o
n
the
origin
al vibration si
gnal
s
and a
c
ou
stic emission
sig
nals, a
s
well
as
decompo
sitio
n
of each IMF
,
putting the experime
n
t data into Eq.(5) t
o
cal
c
ulate th
e energy.
(3)
set the o
r
i
g
inal si
gnal e
nergy valu
e a
s
x
p
, energy value of the ith IMF is
imfi
p
, thus get the
energy ratio o
f
each IMF as
imfi
x
i
p
p
Pp
/
.
(4)
Cal
c
ulate
the gravity frequen
cy to ea
ch IMF acco
rding to Eq.(6).
In our test, to
ol we
ar
state
wa
s divide
d a
c
cordi
ng to th
e the am
ount
of flan
k wea
r
. That
is
whi
c
h
state
1 in
dicates the tool
flan
k
wea
r
a
s
0mm
to 0.05
mm,
state 2
a
s
0.
05mm to
0.1
mm
and state 3 a
s
0.1mm to 0.15mm. J-EE
MD de
com
p
o
s
ition we
re completed u
n
d
e
r different to
ol
wea
r
state of
vibration sig
nals an
d aco
u
stic
emi
s
sio
n
signal
s in turnin
g. Figure 1 sho
w
s a
J-
EEMD deco
m
positio
n re
sults of vibration sig
nal
obt
ained u
nde
r the above thre
e different sta
t
es.
As ca
n be se
en from Fig.
2, the rang
es of
the freque
ncy dom
ain a
r
e not eq
ual. Such a
s
the first
comp
onent, the e
n
e
rgy i
s
con
c
e
n
trated in
the
above 2
000
Hz, the
se
co
nd compo
nen
t, in
the 10
00
Hz to 20
00
Hz, th
e third
comp
onent,
in t
h
e
500
Hz to
10
00Hz. T
hat i
s
to
say, the
more
the tool wea
r
is, the narro
wer the freq
ue
ncy ran
ge of IMF co
rre
sp
o
nding i
s
.
It also
can
b
e
lea
r
ned
fro
m
the test a
n
a
lysis,
that th
e sig
nal e
n
e
r
gy is
con
c
ent
rated i
n
the firs
t three of
IMF. At
th
e s
a
me time, there
a
r
e sig
n
i
ficant ch
ang
es in the thre
e IMF spect
r
u
m
Spectrum
of this th
ree
IMF
ch
ang
es wit
h
the
deg
ree
of tool
wea
r
.
Therefore,
it i
s
d
e
si
rabl
e th
at
energy ratio
after treatme
nt of the J-EEMD first th
ree
IMF, as well
as the g
r
avity frequen
cy were
sele
cted
as the
cha
r
a
c
teri
stic val
u
e
s
of the
vibratio
n si
gnal
me
asu
r
ed
expe
rimentally. It i
s
simila
rly
t
o
se
lect
t
he ch
ara
c
t
e
ri
st
ic v
a
lue
s
of the aco
u
s
tic emi
s
sion
sign
als.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Feature Extra
c
tion of Turi
n
g
Tool Wear
Based o
n
J-EEMD (Hongta
o
Che
n
)
5915
a) De
co
mpo
s
ition results o
f
state 1
b) De
co
mpo
s
ition results o
f
state 2
c)
Decomp
osi
t
ion results of
state 3
Figure 2. J-E
E
MD de
comp
osition results of vibration signal
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 10, Octobe
r 2013 : 591
1 –
5916
5916
5. Conclusio
n
EEMD alg
o
rit
h
m h
a
s solve
d
the
proble
m
of mo
de
m
i
xing proble
m
existing
in t
r
aditional
EMD, but ha
s the lo
we
r computation
a
l
efficien
cy,
no
t cond
ucive t
o
online
moni
toring tool
sta
t
e.
The blin
d so
urce sepa
rati
on JA
DE alg
o
rithm i
s
introdu
ced into
EEMD de
co
mpositio
n pro
c
e
ss,
that ca
n elim
inate the
white noi
se in
t
he IMF
com
pone
nts. Th
e
de
comp
ositi
on effici
en
cy is
signifi
cantly i
m
prove
d
co
mpared
with
EEMD al
go
rithm. Experi
m
ental
re
sult
s
sho
w
that
the
eigenvalu
e
e
x
tracted thro
ugh J-EEMD
algorith
m
ca
n
be well ide
n
tified the different states of the
tool
wear, o
n
the ba
sis of
t
he featu
r
e extra
c
tion
to cutting
vib
r
ation
and
a
c
ou
stic emi
s
sion
sign
als.
Ackn
o
w
l
e
dg
ment
This stu
d
y i
s
sup
p
o
r
ted
by the
Fu
ndame
n
tal
Re
sea
r
ch F
und
s fo
r th
e Central
Universitie
s
, Chin
a (SWJT
U
12
CX03
9).
Referen
ces
[1]
NE Hu
ang, Z
h
eng S
h
e
n
, SR
Lon
g.
T
he e
m
pirica
l
mo
de d
e
co
mp
ositio
n
and th
e Hi
lb
ert spectru
m
for
non
lin
ear a
nd
non-stati
o
n
a
ry time ser
i
es a
n
a
l
ysis
. Proc. Roy. Soc. Lon
do
n
454A, 19
98: 9
03-9
9
5
[2]
Z
H
W
u
and NE Huan
g. Ensembl
e
empir
i
c
a
l mo
d
e
deco
m
positi
on: a n
o
is
e-
assisted data
an
al
ysi
s
method.
Adv
a
nces in Adapti
ve Data Analysis
. 2009; 1(1): 1-
41.
[3]
Z
K
Peng, PW
T
s
e, EL Chu. An improv
ed H
i
l
bert-
Hu
ang tra
n
sform an
d its app
licat
i
on i
n
v
i
btatio
n sig
n
a
l
Anal
ys
is.
Joun
al of Soun
d an
d Vibrati
o
n
. 20
05; 286(
9): 187
-205.
[4]
Shao-
bai Z
h
a
ng, Da
n-da
n
Hua
ng. Electr
oenc
ep
hal
ogra
p
h
y
featur
e e
x
tractio
n
usi
n
g hig
h
time-
freque
nc
y res
o
lutio
n
a
nal
ys
is.
T
E
LKOMNIKA Indon
esi
a
Jo
u
r
nal
of el
ectric
al En
gi
neer
ing
.
201
2; 10(
6):
141
5-14
21.
[5]
Weilin
Li, Pa
n
Fu, Erqin Zha
n
g
. Appl
icati
on
of Fractal Dim
ensi
ons a
nd F
u
zz
y
Cl
usterin
g
to T
ool Wear
Monitori
ng.
T
E
LKOMNIKA Indon
esia
n Jour
nal
of Electric
al
Engin
eeri
n
g
. 2
013, 11(
1): 187
-194.
[6]
Card
oso JF
, Soul
oumi
a
c A.
Blin
d Bea
m
for
m
i
ng for no
n-Gaussi
an sig
n
a
l
s
.
IEEE Proceedi
ngs, Part F.
199
3; 140(
6): 362-3
70.
[7]
Y W
ang, KT
Fang. A n
o
te o
n
uniform
distrib
u
tion
an
d e
x
p
e
r
iments d
e
sig
n
.
Kexue T
o
n
g
b
ao (C
hin
e
se
Scienc
e Bull
eti
n
)
. 1981; 2
6
(6)
:
485.
[8]
Como
n P. T
e
n
s
or Dia
go
nal
iz
ation, a Us
eful
T
ool
in Si
gn
al
Processin
g
’, i
n
Blank
e M., Soderstrom M.
(Eds.)
IF
AC-SYSID 10th Symp
o-siu
m
on S
ystem Ide
n
tific
a
tion, De
n
m
ark
, 1994; 1: 77-8
2
.
[9]
Chalk
i
d
a
, Greece. Blin
d sig
nal pr
ocessi
ng
algorit
hms,
12th Int. W
o
rkshop o
n
Systems, Sign
als &
Imag
e Process
i
ng
. 20
05: 22-
2
4
.
[10]
yvar
in
en A, Karhun
en J, Oja
E. Indepe
nde
nt
Compo
nent A
nal
ysis.
Jo
hn
W
iley & Sons
, 200
1.
Evaluation Warning : The document was created with Spire.PDF for Python.