TELKOM
NIKA
, Vol. 11, No. 12, Decem
ber 20
13, pp.
7539
~75
4
7
e-ISSN: 2087
-278X
7539
Re
cei
v
ed
Jun
e
28, 2013; Revi
sed Aug
u
st
14, 2013; Accepted Sept
em
ber 2, 201
3
Ultrasonic Flaw Signal Classif
i
cation using Wavelet
Transform and Support Vector Machine
Yu Wang
Schoo
l of Softw
a
r
e En
gin
eer
i
ng, Cho
n
g
q
in
g Univers
i
t
y
of Arts and Scienc
e
s
, Chong
qi
ng 4
021
60, Ch
ina
e-mail: ll
yl
a
b
@
126.com
A
b
st
r
a
ct
T
h
is pap
er pre
s
ents a ultras
o
n
ic flaw
sign
al
cla
ssificati
on s
ystem by
usi
ng wavelet transform
and
supp
ort vector
mac
h
in
e (SV
M). A digital fl
aw
detecto
r is
first used to acqu
ire the si
gna
ls of defec
tive
carbo
n
fiber re
i
n
forced p
o
ly
mer (CF
R
P) spe
c
imen w
i
th voi
d
, dela
m
in
atio
n and
deb
on
di
ng. After that, the
time
do
main
b
a
sed
ultraso
n
i
c
sign
als ca
n
be pr
ocesse
d
by discrete
w
a
velet transf
o
rm (
D
W
T
), and
infor
m
ativ
e fe
atures ar
e ext
r
acted fro
m
D
W
T
coeffi
cient
s repres
entati
on of si
gn
als.
F
i
nal
ly, featu
r
e
vectors select
e
d
by PCA
meth
od are tak
en a
s
input to
trai
n the SVM classi
fier. F
u
rthermo
re, the selecti
o
n
of SVM para
m
eters and k
e
rn
el functio
n
has
been ex
a
m
i
n
e
d
in deta
ils. Ex
peri
m
e
n
tal res
u
lts vali
date th
at
the
mo
del
co
u
p
le
d w
i
th w
a
v
e
let tra
n
sfor
m
and
SVM is
a
pro
m
is
in
g too
l
to
dea
l w
i
th
classificati
on
for
ultraso
n
ic flaw
sign
als.
Ke
y
w
ords
:
ultr
ason
ic sign
al cl
assificati
on, su
pport vector
machi
ne, feature
extraction, w
a
velet transfor
m
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Ultra
s
oni
c m
e
thod
s are th
e most
su
cce
ssful
non
-de
s
tructive testin
g (NDT
) tech
nique
s
for q
uality asse
ssm
ent
an
d dete
c
tion
of
flaws in
en
gi
neeri
ng
mate
rials.
Co
nvent
ional
ultra
s
on
ic
testing techni
que
s, however, are
ba
sed
on manu
al
o
r
experi
ential
pattern ide
n
tification, whi
c
h
easily b
r
in
gs about
co
stl
y
, lengthy a
nd e
rra
ti
c a
nalysi
s
. Co
n
s
ide
r
abl
e a
d
v
ancem
ent a
nd
developm
ent
in the l
a
st fe
w d
e
cade
s
h
a
ve en
able
d
ultrasoni
c te
sting to
cha
n
g
e
from
a Bl
ack-
Smith profe
s
sion to
an a
d
vanced mul
t
idisci
plin
a
r
y
engin
eeri
ng profe
ssi
on. Mode
rn sign
al
pro
c
e
ssi
ng te
chni
que
s
and
artifici
al intell
igen
ce to
ols
can
be
integ
r
ated a
s
auto
m
atic
ultra
s
o
n
ic
sign
al cla
s
sif
i
cat
i
on
sy
st
em
s (A
US
CS
) [
1
]
.
I
n
A
U
S
C
S,
ultrasoni
c fla
w
sig
nal
s acq
u
ired in
a form
of digitized d
a
ta are p
r
ep
roce
ssed firstl
y, and in
form
ative feature
s
are extra
c
te
d usin
g vario
u
s
digital si
gnal
pro
c
e
ssi
ng
and p
a
ttern
recognitio
n
tech
niqu
es. F
i
nally, the set of sele
cte
d
feature
s
be
comes the b
a
s
is of flaw id
entificati
on b
y
training the
prope
r cl
assifier. Therefo
r
e,
extraction of f
eature
s
an
d d
e
sig
n
of cla
s
sifier play criti
c
al role
s in AUSCS.
The o
b
jectiv
e of this
con
t
ribution i
s
t
o
sh
ow th
at advan
ced
sig
nal processin
g
an
d
pattern re
cog
n
ition
techniq
ues ca
n
aid ultrasoni
c testing to co
rre
c
tly identify different fla
w
s
found in
carbon fibe
r rei
n
forced p
o
lymers (CFRP
s
).
The
re
st of this pap
er is org
ani
zed
a
s
follows. Secti
on 2 reviews
the previo
us
related
work.
Section 3
de
scrib
e
s th
e me
thodolo
g
ies o
f
wavelet tran
sform (WT
)
and sup
p
o
r
t vector m
a
chi
ne (SV
M
). Section
4 describe
s
the
experim
ental
pro
c
ed
ure an
d se
ction
5 a
nalyze
s
the
e
x
perime
n
tal result
s. Sectio
n 6 a
ddresse
s
t
he con
c
lu
sio
n
s.
2.
Rela
ted W
o
rk
The potential
of signal proce
s
sing an
d patte
rn re
cog
n
ition an
alysis on ult
r
ason
i
c
testing ha
s b
een inve
stiga
t
ed by severa
l authors.
Lee criti
c
ally reviewed pop
ular
fe
ature
e
x
trac
tion te
ch
nique
s in
AUSCS, incl
udin
g
fast
Fouri
e
r tra
n
sform (FF
T
) a
nd discrete
wavelet tra
n
sform (DWT),
identified cri
t
ical issue
s
i
n
feature
extra
c
tion, an
d
compa
r
ed th
e
rep
o
rted
ap
proa
ch
es to
dra
w
thei
r
stren
g
ths an
d
w
e
ak
ne
ss
es
[2
, 3
]
.
Yamani et al. developed a
databa
se of ultr
asoni
c A-scan si
gnal
s by using an
out-of-
servi
c
e
pressure
vessel
with lots
of hi
gh te
m
peratu
r
e hydro
gen
attach defe
c
ts.
Th
e
b
a
si
c
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 753
9 – 7547
7540
feature
extra
c
tion m
e
thod
co
uple
d
wit
h
pri
n
ci
pal
compon
ent a
n
a
lysis (P
CA) we
re
used
to
rep
r
e
s
ent th
ese
sets
of
A-scan
si
gna
ls. Exper
im
e
n
tal re
sult
s
showed th
at
a pri
o
ri
train
ed
cla
ssifie
r
b
a
sed o
n
ne
arest-neigh
bo
r cri
t
erion
ca
n di
stingui
sh
accurately the
h
y
droge
n atta
ck
from geom
etrically simila
r
defect
s
[4].
ISA et al. provided a conti
nuou
s
syste
m
for
oil
an
d gas pipelin
e con
d
ition
mo
nitoring.
The raw ultraso
n
ic
sig
nal
s were first
pro
c
e
s
sed
using DWT
an
d then
cla
s
si
fied usi
ng S
V
M.
Prelimina
r
y tests
sh
owed
that the SVM algorith
m
wa
s abl
e to cla
s
sify the sig
n
a
l
s a
s
abn
orm
a
l
in the pre
s
en
ce of wall thin
ning [5].
Matz et al. u
s
ed th
e DWT
based meth
od for filterin
g of ultra
s
oni
c si
gnal to
suppress
the ech
o
e
s
from grai
ns. S
V
M was u
s
e
d
to automat
i
c
ally cla
s
sify ultrasoni
c sig
nals, with fau
l
t
ech
o
, ech
o
from wel
d
and
back-wall
ech
o
, measur
e
d
on mate
rial u
s
ed fo
r con
s
tructing ai
rpl
a
n
e
engin
e
s [6].
Anasta
ssopo
ulos
et al. condu
cted
an
ext
ensive di
scrimin
a
tion
study on
ultraso
n
ic
sign
als very
simila
r to
e
a
ch
othe
r o
b
tained
from
artificial
in
serts i
n
a
CFRP pl
ate.
The
perfo
rman
ce
of fifteen cla
s
sificatio
n
sch
e
mes con
s
isti
ng of n
on-pa
rametri
c
patte
rn recognitio
n
and A
r
tificial
Neu
r
al
Net
w
ork (A
NN)
al
gorithm
s
wa
s asse
ssed,
a
nd a
n
u
ppe
r bou
nd fo
r t
h
e
cla
ssifi
cation
error expe
cte
d
with simila
r ultr
asoni
c sig
nals
wa
s defi
ned.
Moreove
r
, the Wilk’
s
Λ
crite
r
ion
wa
s proved effici
e
n
t for featur
e
sele
ction in th
eir expe
rimen
t
s [7].
Ca
cci
ola
et a
l
. pro
p
o
s
ed
a
n
he
uri
s
tic a
ppro
a
ch fo
r
cla
ssifying
th
e ultrasonic
ech
o
e
s
measured on
defective CF
RP. The prop
ose
d
me
tho
d
was
ba
sed o
n
the use of DWT and P
C
A
for featu
r
e
extraction
an
d
selectio
n [8]. Experim
e
n
tal result
s a
s
sure
d go
od p
e
rfo
r
mances of th
e
impleme
n
ted
SVM cla
ssifie
r
. They al
so
develop
ed
a softwa
r
e pa
ckag
e,
whi
c
h
allowes users
to
perfo
rm the cross wavelet tran
sform, the
wavele
t coh
e
r
en
ce an
d the fuzzy infere
n
c
e sy
stem for
impleme
n
ting
a data-ind
e
p
ende
nt cla
ssif
i
er [9].
Sambath
et al. improved t
he
sensibility of fl
aw detection in
ultrasonic testing by
usi
ng
an ANN
and
sig
nal p
r
o
c
e
ssi
ng te
chni
q
ue. Wavelet
transfo
rm
wa
s u
s
e
d
to d
e
r
ive a fe
ature
vector whi
c
h contai
ns
t
w
o-dimen
s
ion
a
l
i
n
formatio
n o
n
four type
s
of defect
s
, n
a
mely po
ro
sity,
lack of
fusi
on
, tungste
n i
n
clu
s
ion
an
d
non
defe
c
t. T
hese ve
ctors we
re
then
classified
by a
n
ANN traine
d
with the b
a
ck pro
pagatio
n
algorith
m
.
Using the
wavel
e
t feature
s
a
nd ANN, go
o
d
cla
ssifi
cation
at the rate of 94% wa
s obt
ained [10].
Schul
z et al
. focuse
d o
n
the autom
atic
evaluati
on of the b
a
ckscatte
red
signal
s
received
by the ultrasoni
c
sen
s
o
r
s.
T
he evaluation sy
stem wa
s
ba
sed on a
stati
s
tical cla
s
sifier
usin
g mo
st di
scrimin
a
tive feature
s
extra
c
ted
from the
backscattere
d echo
sign
al
s a
c
cording t
o
their
am
plitu
des, conto
u
r,
co
rrel
a
tion and regi
on.
By this mea
n
s th
ey impl
emented
reli
able
defect dete
c
ti
on for an aut
omatic
cha
r
a
c
teri
zation of
the CF
RP ma
terial [11].
Liu et
al. pro
posed
algo
rithms for
defe
c
t
dete
c
tion
based on discrete
wavelet
pa
cket
transfo
rm
an
d BP network. Furthe
rmo
r
e
,
the re
co
nf
ig
urabl
e a
r
chite
c
ture
of the
d
e
fect d
e
tectio
n
in embed
de
d system
was di
scusse
d
.
Accordi
ng
to the experiments of
ult
r
asoni
c sig
n
al
pro
c
e
ssi
ng, su
ch archite
c
ture
could
provide
a fle
x
ible and
efficient
sol
u
tion to em
bed
d
e
d
reconfigu
r
a
b
le sign
al pro
c
essing
syste
m
[12].
As mention
e
d
above, there
are two
strate
gi
es for featu
r
e extraction i
n
AUSCS.
(1)
Ch
oose feature
s
from
different do
m
a
ins
of
ultra
s
onic
sign
als.
The de
rived f
eature
s
mainly incl
ud
e statisti
cal
para
m
eters e
x
tracted
from
statistical m
o
ments of th
e time-d
omai
n
or/and
freq
u
ency-dom
ain
based
ultra
s
onic
sig
nal
s, su
ch
as
me
an, varia
n
ce, skewne
ss a
nd
kurto
s
i
s
. Usin
g su
ch
strate
gy, users n
e
ed effect
ive f
eature
sel
e
cti
on sch
e
me
s
to evluate th
e
discrimi
nation
of features, redu
ce the
re
d
unda
ncy an
d optimize the f
eature
set.
(2)
Use
dire
ct
ly the wh
ole
sign
al sectio
n
deriv
e
d
from
the ultra
s
o
n
i
c
sca
ns a
s
in
put to
the classifie
r
. The input feature
s
mainly
in
clude FF
T
coefficient
s and DWT co
efficients. Su
ch
strategy
dem
and
s mi
nimu
m prep
ro
ce
ssing, i.e
., no
t
much
featu
r
e sel
e
ctio
n scheme
s
are
employed. Howeve
r, coeff
i
cient feat
ure
s
are ofte
n high dimen
s
io
n
a
l.
More
over, there a
r
e two
convention
a
l cl
as
sifiers used
in AUSCS, namely ANN a
nd
SVM. Genera
lly, SVM has sup
e
rio
r
predi
ction an
d gen
erali
z
ation p
e
r
forma
n
ce in view of small
sampl
e
si
ze p
r
oble
m
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Ultra
s
oni
c Fla
w
Signal Cl
assificatio
n
usi
n
g Wa
velet Transfo
rm
and Suppo
rt… (Y
u Wan
g
)
7541
3. Methodol
og
y
3.1. Wav
e
let Trans
f
orm
Fouri
e
r t
r
an
sf
orm
ca
n b
e
use
d
to im
prove
the
perf
o
rma
n
ce of f
eature
extra
c
tion for
flaws by m
a
pping the ti
me domai
n
based si
gnal
s into fre
que
ncy dom
ain
based si
gnal
s.
Ho
wever, the
frequen
cy d
o
main c
haracteristics of trans
i
ent sign
a
l
s wo
uld not be refle
c
ted
by
Fouri
e
r tran
sform d
ue to it
s glo
bal p
r
op
erty. Yet wa
velet transfo
rm (WT) i
s
a
kind
of time
-
freque
ncy
do
main meth
od
with multi
-
resol
u
tion a
n
a
lysis,
whi
c
h
can
adju
s
t
the time an
d
freque
ncy p
r
o
perty as requi
red. The d
e
compo
s
ed p
a
rt
s of the sign
a
l
are re
solved
such that the
highe
r the fre
quen
cy, the f
i
ner th
e re
sol
u
tion
[13]. WT ha
s po
we
rful ability for
denotin
g lo
ca
l
sign
al ch
ara
c
teristics both
in time and freque
nc
y dom
ain. WT can
be co
nsi
dere
d
as a spe
c
ia
l
filtering ope
ration, and th
e frequ
en
cy segm
entatio
n
is obtaine
d by dilating the wavelet. It is a
wind
owin
g te
chni
que
with
variable
si
ze
d re
gion
s,
wh
ich
allows th
e
use of l
ong ti
me inte
rvals to
obtain mo
re pre
c
ise low f
r
equ
en
cy informatio
n and
shorte
r re
gi
ons
whe
r
e h
i
gh frequ
en
cy
informatio
n is neede
d.
Note that as
a fast algorith
m
to obtain the
wavelet transfo
rm of a discrete time
sign
al,
discrete
wavelet tran
sform (DWT
) h
a
s been
wid
e
ly
use
d
in the
u
l
traso
n
ic si
gn
al analy
s
is. T
h
e
DWT analyzes the sig
n
a
l
by decomp
o
sin
g
it
into its coarse
approximatio
n and detail
ed
informatio
n, whi
c
h is a
c
compli
sh
ed
by usi
ng su
ccessive hig
hpa
ss
an
d lowp
ass filtering
operation
s
in
the frequ
en
cy domain [1
4]. The o
r
iginal
sign
al
x
[
n
] i
s
first pa
ssed t
h
rou
gh a
half
-
band hig
hpa
ss filter
g
[
n
] and lowpa
s
s filter
h
[
n
], where
g
[
n
] an
d
h
[
n
] are quadrature mirro
r
filters of each other. After the
filtering
,
half of
the sample
s of the two outp
u
t signal
s are
discarded
by
downsamplin
g sin
c
e th
e si
gnal
s no
w h
a
v
e a ban
dwi
d
th of
π
/2 ra
di
ans i
n
ste
ad o
f
π
. This con
s
titutes one lev
e
l of decom
p
o
sition
a
nd it is expre
s
sed
mathemati
c
al
ly as:
[]
[
]
[
2
]
high
n
y
kx
n
g
k
n
(
1
)
[]
[
]
[
2
]
low
n
y
kx
n
h
k
n
(
2
)
Whe
r
e
y
high
[
k
] and
y
low
[
k
]
are the o
u
tp
uts of the hi
ghpa
ss and l
o
wp
ass filters after
downsamplin
g by
2. Th
e a
bove p
r
o
c
e
d
u
r
e i
s
re
peat
e
d
for furth
e
r d
e
com
p
o
s
ition
of the
lo
wpa
ss
filtered sig
nal
s.
3.2. Support
Vector Ma
ch
ine
Suppo
rt vect
or ma
chin
e (SVM) is a
struct
u
r
al
risk
based lea
r
ni
ng ma
chin
e, whi
c
h
c
o
ns
tr
uc
ts
N
-dime
n
si
onal
hyperpl
ane
to optimally sepa
rate th
e input data
into differe
nt
categ
o
rie
s
. A sigmoi
d ke
rn
el function
m
odel of SVM is equival
ent
to a two-l
a
yer, feed-forwa
r
d
neural net
work. Furth
e
rm
ore, SVM can
use
polynomi
a
l functio
n
or radial
ba
sis f
unctio
n
(RBF
)
in whi
c
h the
weig
hts of th
e netwo
rk are f
ound by solving a qua
dratic
pro
g
ra
mming p
r
obl
em
with lin
ear co
nstrai
nts [1
5]. Since
SVM i
s
robu
st in
hi
gh dim
e
n
s
ion
a
l spa
c
e
s
with a
sp
arse
set
of sample
s, it may be use
d
either to cl
assify
or pre
d
ict som
e
arbitrary patterns from a set of
labele
d
data
while avoidi
n
g
over-fitting t
he data at the
convergen
ce
of the training
[16, 17].
Let {
x
i
,
y
i
} be a dataset, where
x
i
is a
d
-dim
ensional sample (
i
=
1
,2,…
l
) and
y
i
is
the
corre
s
p
ondin
g
bipola
r
la
b
e
l (
y
i
{-1,1}
). Assu
me tha
t
we hav
e d
e
fined a lin
e
a
r sepa
ratin
g
hyperpl
ane b
y
w
x
+
b
for training
sampl
e
s, then it sh
ould meet:
(
)
1
,
{
1
,
2
,
.
.,
}
ii
y
wx
b
i
l
(
3
)
The optimal
separating hyp
e
rpla
ne (OSH) can
n
o
t only corre
c
tly sep
a
rate the
sam
p
les,
but also maxi
mize the ma
rgin between t
he clo
s
e
s
t po
sitive sampl
e
s and n
egativ
e sampl
e
s.
The se
pa
rabl
e margi
n
ca
n be cal
c
ul
ated
as follow:
{|
1
}
{|
1
}
11
2
(,
)
m
i
n
m
a
x
||
||
||
||
||
|
|
|
|
||
||
||
ii
ii
ii
xy
xy
wx
b
w
x
b
dw
b
ww
w
w
w
(
4
)
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 753
9 – 7547
7542
Obviou
sly the maximum
of
(,
)
dw
b
may be
achi
eved th
rough th
e mi
nimizatio
n
of
||
w
||
2
/2.
By u
s
ing a nu
mb
er of nonn
eg
ative slack variabl
es
i
, the trainin
g
o
f
SVM can be
formulate
d
as solving a qu
adrati
c
optim
al probl
em:
2
1
,,
1
mi
n
|
|
|
|
s
.
t
.
(
)
1
,
0
2
i
l
ii
i
i
i
i
wb
wC
y
w
x
b
(
5
)
Acco
rdi
ng to Lagrangi
an theory, it yields
ii
i
i
wy
x
with c
o
ns
traints
0
i
C
and
0
ii
i
y
, where
i
=1,
2
,
…
l
. Note that
i
can
be fou
nd aft
e
r the
follo
wing
proble
m
is
maximized:
,
1
2
D
ii
j
i
j
i
j
ii
j
Ly
y
x
x
(
6
)
The samples w
i
th value
i
>0 a
r
e
calle
d
sup
port ve
ctors
(SVs). Th
e de
cisi
on fu
nction
can b
e
derive
d
as follo
w:
1
(
)
sg
n(
)
s
gn
(
)
s
N
ii
i
i
f
x
w
xb
y
x
xb
(
7
)
Whe
r
e,
N
s
is t
he numb
e
r of
SVs.
Linea
r sepa
ration of data
s
ets can not b
e
achi
eved
succe
ssfully al
l the time. Therefo
r
e
the point
s in
origin
al spa
c
e sh
ould
be
expand
ed to
a feature spa
c
e
with hig
h
e
r
dime
nsi
onal
ity
and
hen
ce
li
near
sepa
rati
on
can
be
retried
[18].
This expa
nsi
on p
r
o
c
e
s
s i
s
reali
z
ed
wi
th
operator
()
and the OSH turns into th
e form
()
()
f
xwx
b
. We may consi
d
er an
augme
n
ted space by utilizi
ng ke
rnel fun
c
tion in the fo
rm of
(,
)
(
)
(
)
ii
Kx
x
x
x
[19].
4. Experimental proc
edu
r
e
4.1. Specimens
Carbon fibe
r reinfo
rced po
lymers (CF
R
Ps)
are man
u
factured by mixing carbo
n
fibers
and pla
s
tic re
sin und
er p
r
e
s
cribe
d
co
ndi
tions. The m
o
st com
m
on f
o
rm of CF
RP
s is the cro
ss-
ply laminate, su
ch a
s
layin
g
up a sequ
e
n
ce of uni
dire
ctional pli
e
s
[
21]. The mat
e
rial
s have hi
gh
elasti
c modul
us an
d tensil
e stren
g
th wit
h
low den
sity as well a
s
thermal exp
a
n
s
ion. They have
been
wi
dely
use
d
for vari
ous compo
n
e
n
ts a
nd
stru
ct
ure
s
,
such a
s
aircraft fusel
age
as well a
s
wing structures, helicopter roto
rs and windmill bl
ades,
due to
their
excell
ent
properties.
Ho
wev
e
r, the
CF
RPs a
r
e
r
e
lativ
e
ly
brittle co
mpa
r
ing
with metallic
materials
[22]
. Flaws
in form
of void, delamination an
d
debo
nding m
a
y occur in
CFRPs d
u
ri
ng
the manufa
c
t
u
ring p
r
o
c
e
ss or
unde
r co
mple
x environme
n
t
s and loa
d
in
g states.
In this stu
d
y, two CF
RP
specim
en
s we
re u
s
ed fo
r e
x
perime
n
t. An artificial
de
fective
CFRP
spe
c
i
m
en m
e
a
s
ures
300m
m
3
00mm
5mm,
with void (
3mm), top
del
amination,
mi
ddle
delamin
ation
and bottom
delamin
ation, which ar
e
depicte
d
in
Figure
1. Another
defe
c
tive
CFRP
spe
c
i
m
en is a pa
n
e
l with natura
l
debon
ding.
Figure 1. The
Specime
n
wi
th Void and
Delaminatio
n
5
300
T
op delam
i
nation
M
i
ddle delam
i
nation
Botto
m
dela
m
i
nati
on
300
Void
3
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Ultra
s
oni
c Fla
w
Signal Cl
assificatio
n
usi
n
g Wa
velet Transfo
rm
and Suppo
rt… (Y
u Wan
g
)
7543
4.2. Signal Acquisition
A PXU T227
digital flaw
detecto
r was use
d
to s
end ultras
onic
waves
into the CFRP
spe
c
ime
n
s u
nder test th
rough
a tran
sducer
ope
rati
ng at the
ce
ntral freque
n
c
y of 5M
Hz.
An
ec
ho was
reflec
ted back
eac
h
time when the ultr
ason
ic wave en
co
untere
d
a discontin
uity in the
prop
agatio
n
medium. T
he
A-scan
si
gnal
wa
s
digitise
d
at a
sam
p
lin
g freq
uen
cy o
f
100 M
H
z a
n
d
sampl
e
len
g
th of 4
k
u
s
in
g a Son
o
tek STR 8
100
A/D boa
rd, a
nd then
sto
r
ed in a
pe
rsona
l
comp
uter (P
C). The ult
r
a
s
onic te
sting system is sho
w
n in Figu
re
2.
Figure 2. The
Ultrasoni
c T
e
sting Syste
m
As the
data
s
et for furth
e
r
cla
ssifi
cat
i
on expe
rim
ents, the
co
llected
sig
n
a
l
s a
r
e
comp
osed by
:
(1) 30
ultra
s
onic pul
se
s
affected
by
delamin
ation-like fla
w
s at
the top, mi
d
d
le a
n
d
bottom of the in-stu
dy spe
c
i
m
en re
pe
ctively;
(2) 2
0
ultra
s
o
n
ic pul
se
s de
scribin
g
void of the in-stud
y
specim
en;
(3) 2
0
ultra
s
o
n
ic pul
se
s de
scribin
g
deb
o
nding of the i
n
-stu
dy spe
c
i
m
en;
(4) 3
0
ultra
s
o
n
ic pul
se
s sh
owin
g ab
sen
c
e of defect.
4.3. Featur
e Extrac
tion a
nd Selection
After pre
-
pro
c
e
ssi
ng, the
sign
als d
e
scribing
differen
t
flaws can b
e
cha
r
a
c
teri
zed by
wavelet coefficient
s whi
c
h
are the succe
ssive
conti
nuation of th
e app
roximat
i
on co
efficien
ts
and detail co
efficients by usin
g DWT. In this st
udy, each sign
al wa
s decomp
o
se
d into 3 levels
usin
g Daub
e
c
hie
s
wavele
t. The sig
nal
s for th
re
e types of fla
w
s (delami
nati
on, void an
d
debo
nding
) a
nd
the
re
pre
s
entation of
th
eir co
rre
sp
on
ding 512 sa
m
p
les
of DWT
coeffici
ents
a
r
e
s
h
ow
n in
F
i
gu
r
e
3
to F
i
gu
r
e
5
r
e
s
p
ective
ly. O
b
vio
u
s
ly, th
es
e
D
W
T
c
o
e
ffic
i
e
n
t
s
c
o
mp
le
te
ly
descri
be the
macro-t
r
en
d of each
sign
a
l
.
Figure 3. Ultraso
n
ic Sign
al
for Delami
na
tion and its DWT Coefficie
n
ts Re
pre
s
e
n
tation
Figure 4. Ultraso
n
ic Sign
al
for Void
and
its DWT Co
efficients
Rep
r
e
s
entatio
n
0
500
1
000
150
0
200
0
2
500
30
00
3500
4
000
0
50
100
150
200
250
300
De
l
a
m
i
n
a
t
i
o
n
0
50
10
0
150
20
0
25
0
30
0
35
0
40
0
45
0
500
-
300
-
200
-
100
0
100
200
300
cD
3
0
500
1000
1500
200
0
25
00
3000
350
0
400
0
0
50
10
0
15
0
20
0
25
0
30
0
vo
i
d
0
50
100
150
20
0
25
0
300
350
40
0
45
0
50
0
-
300
-
200
-
100
0
100
200
300
cD
3
Per
s
onal Co
m
puter
Digitizer c
a
rd
Ultrasonic f
l
aw det
ector
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 753
9 – 7547
7544
Figure 5. Ultraso
n
ic Sign
al
for Debo
ndin
g
and its DWT Coefficie
n
ts Representati
o
n
Eight inform
a
t
ive feature
s
were extra
c
te
d fr
om th
e DWT
coeffici
en
ts re
pre
s
e
n
ta
tion of
each sig
nal:
(1) Me
an valu
e:
1
1
N
i
i
A
VG
x
N
(2) Stand
ard deviation:
2
1
1
()
1
N
i
i
STD
x
AV
G
N
(3) Maxim
u
m amplitude
(4) Mini
mum
amplitude
(5) Maxim
u
m energy
(6) Average freque
ncy
(7) F
r
eq
uen
cy of minimum energy samp
les
(8)
Half point
(HaPo
)
: the frequen
cy that di
vides u
p
the spe
c
trum into two parts o
f
s
a
me
ar
ea
.
More
over, th
e conve
n
tion
al time-dom
ai
n based
stati
s
tical p
a
ra
me
ters of ea
ch
sign
al,
inclu
d
ing the
mean value,
root mean
square value,
standa
rd d
e
v
iation and a
b
sol
u
te value
,
were also cal
c
ulate
d
and t
a
ke
n as a
not
her fou
r
featu
r
es.
Due to the
still large
dimensionality of feat
ure space, t
he P
C
A method
was
exploi
ted to
redu
ce th
e n
u
mbe
r
of inp
u
ts into cl
assi
fier
by only consi
deri
ng th
e prin
cip
a
l co
mpone
nts (P
Cs)
who
s
e contri
bution
s
to total variation of the w
hole se
t of PCs are greate
r
than
2%. Finally, the
input numb
e
r
of elements f
o
r cla
s
sifier h
a
s be
en re
du
ced fro
m
12 to 6.
4.4. SVM Classifica
tion
The trai
ning
set an
d test
set for
cla
ssi
fication exp
e
r
iment
s were
comp
osed b
y
100
sign
als
colle
cted in sectio
n 4.2. Six PCs m
entio
n
e
d
in sectio
n 4.3 we
re ta
ken a
s
the in
put
vector for trai
ning SVM cl
assifier. We con
d
u
c
ted th
e one
-agai
nst-one meth
od
for multi-cl
a
s
s
cla
ssif
i
cat
i
on
(6 cla
s
s
e
s
in t
h
is st
udy
,
i.
e.
, top delamination, m
i
ddle dela
m
ination, bottom
delamin
ation,
void, debon
ding, no defe
c
t) and a
dopt
ed five fold cross validatio
n asse
ssm
en
t
for traini
ng.
First, the
sa
mples we
re
randomly divi
ded into five
grou
ps. In th
e trainin
g
sta
ge,
one
gro
u
p
wa
s left
out a
s
t
e
st
sam
p
le
s f
o
r ve
rifying th
e SVM
cla
ssif
i
er, a
nd th
e ot
her re
mainin
g
four g
r
ou
ps
were u
s
ed
a
s
trai
ning
sa
mples.
T
he p
r
ocess
did n
o
t terminate
until every g
r
ou
p
wa
s take
n as test sample
set. Finally, average of the
five recorde
d
results was taken a
s
the
result of the trained SVM cl
assifier.
5. Results a
nd Analy
s
is
In this
s
t
udy, three c
l
ass
i
c
a
l k
e
rnel func
ti
ons u
s
e
d
for SVM training
were as follo
ws:
(1) Li
nea
r ke
rnel:
(,
)
ii
K
xx
x
x
(2) Polyno
mi
al kernel:
(,
)
(
1
)
p
ii
Kx
x
x
x
(3) RBF
ke
rn
el:
2
(
,
)
e
xp(
||
||
)
ii
Kx
x
x
x
The re
cog
n
ition rate
s and
training time
s of SVMs with various kernel fun
c
tio
n
s are
resume
d in T
able 1. As is
sho
w
n in the
tabl
e, the mean recogniti
on rate
s of SVMs with
RBF
kernel
s
are
h
i
gher tha
n
th
ose
with
line
a
r
and
polyn
omial
ke
rnel
s. Ho
weve
r, S
V
Ms
with
RB
F
0
500
10
00
1
500
2
000
2500
3000
35
00
40
00
0
50
100
150
200
250
300
D
e
b
ondi
ng
0
50
100
150
200
250
30
0
35
0
40
0
45
0
500
-
300
-
200
-
100
0
100
200
300
cD
3
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Ultra
s
oni
c Fla
w
Signal Cl
assificatio
n
usi
n
g Wa
velet Transfo
rm
and Suppo
rt… (Y
u Wan
g
)
7545
kernel
s
had
the m
a
ximum
training
time
s due
to th
eir
expone
ntial
computation
a
l
compl
e
xity. Let
us fo
cu
s o
n
the SVMs
wi
th polynomi
a
l
ke
rnel
(
p
=3
)
and RBF kernel
(
=0.1),
denote
d
a
s
Poly3SVM and RBF0.1S
V
M resp
ectiv
e
ly. Poly3S
VM achieve
s
97.5% of training re
co
gniti
on
rate
within
81.4s
.
In this c
a
s
e
, t
w
o t
op de
lamin
a
tion
flaws of
the CFRP spe
c
ime
n
we
re
cla
ssifie
d
a
s
middle delam
ination
fla
w
s, whi
c
h wa
s n
o
t
affected
by false
po
sitive
or n
egative
at
all. Co
mpa
r
e
d
to
RBF0.1
SVM, 98.75
% of trai
ning
re
co
gnition
rate withi
n
2
7
0
.1s, Poly3S
VM
gain
s
230% improvem
ent
for training
efficiency where
a
s 1.2
5
% loss for reco
gnition ra
te.
Therefore, P
o
ly3SVM ca
n perfe
ctly achi
ev
e the
trade
-off b
e
twee
n the
comp
utatio
nal
compl
e
xity and cla
ssifi
catio
n
perfo
rman
ces.
Table 1. Re
cognition
Rate
s and T
r
aini
n
g
Times of S
V
Ms with Diff
erent Kernel
Functio
n
s
Kernel function
Recognition rate
of training data
(
%
)
Recognition rate
of test data (
%
)
Training time (s)
Linear (
C
=1
)
91.25
87.5
21
Poly
nomial,
p
=2 (
C
=0.1)
95
90
67.5
Pol
y
no
mial,
p
=3
(
C
=0.1)
97.5 92.5 81.4
Poly
nomial,
p
=4 (
C
=0.1)
97.5
93.75
101.3
RBF,
=10 (
C
=1
)
96.25
91.25
218.5
RBF,
=1 (
C
=1
)
97.5 92.5
230.3
RBF,
=0.1 (
C
=1
)
98.75
93.75
270.1
For fu
rthe
r compa
r
ation,
we al
so
impl
ement
ed
the
back p
r
o
pag
ation (BP) ne
twork b
y
usin
g MATLA
B
NN To
olbo
x for cla
ssifyi
ng the flaw
si
gnal
s from
CFRP spe
c
ime
n
. The outp
u
t of
BP network wa
s a 6 com
pone
nt vector. A compone
nt value in the 1±
δ
in
ter
v
a
l
w
a
s
co
ns
id
er
ed
as 1
and
a
compon
ent val
ue in the
0±
δ
in
te
r
v
a
l
w
a
s
c
o
ns
ide
r
ed
as
0
,
w
h
er
e
δ
>0. Th
e opti
m
al
BP
netwo
rk architectu
re wa
s sele
cted
ba
sed on
th
e average
of
the be
st cl
assificatio
n
results
for the 6 cla
s
ses of flaws. The value
s
of
all para
m
eters for training the B
P
network a
r
e
resume
d in
Table
2. Tab
l
e 3 di
spl
a
ys the cl
assification a
c
curacy results
obt
ained
by u
s
i
n
g
SVMs an
d B
P
netwo
rk. O
b
viously, the
SVM cla
ssifie
r
yield
s
b
e
tter
classificatio
n
pe
rform
a
n
c
e
than that of t
he BP ne
ural
network. In t
he
CFRP fl
a
w
ide
n
tificatio
n
case, we
can
con
c
lud
e
t
hat
the SVM o
u
tperfo
rms the BP network due
to
its high
er
gene
rali
zatio
n
ca
pability
for
cla
ssifi
cation
probl
em with
small sampl
e
size.
Table 2. The
Paramete
rs o
f
BP Network
Parameters
Val
ues
No. of input featu
r
es
6
Activation functio
n
at hidden la
y
e
r
tan-sigmoid trans
fer function
Activation functio
n
at output la
yer
tan-sigmoid trans
fer function
Training algorith
m
trainscg
No. of neu
rons at
hidden la
y
e
r
13
Performance g
o
a
l
0.001
Net
w
ork structu
r
e
6-13-6
Table 3.
Co
m
pari
s
on b
e
tween BP Network a
nd SVM
s
Classifier
Recognition rate
of training data
(
%
)
Recognition rate
of test data (
%
)
Training time (s)
BP net
w
o
rk
91.25
86.25
85
Standard SVM
98.75
93.75
170.5
6. Conclusio
n
In this pape
r, we u
s
ed the
digital flaw de
tector to a
c
qu
ire ultra
s
o
n
ic
sign
als from CF
RP
spe
c
ime
n
with void, d
e
lam
i
nation
and
d
ebon
di
ng, an
d
utilized adv
anced sig
nal pro
c
e
ssi
ng a
n
d
pattern
re
co
g
n
ition te
chni
q
ues to im
ple
m
ent a
u
toma
tic cl
assification fo
r the
s
e
flaw
sig
nal
s.
DWT an
d PCA we
re first u
s
ed
for fe
ature extra
c
ti
on a
nd sele
ct
ion.
I
n
cla
s
sif
i
cat
i
on
p
r
o
c
e
ss,
we
trained th
e SVM to identify different flaws. Mo
re
over, the sel
e
ction of kerne
l
function
wa
s
discussed detailly so as to train the SVM clas
sifier with the best co
mprehensiv
e performance.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 12, Dece
mb
er 201
3: 753
9 – 7547
7546
Experimental
result
s sh
owed that
the propo
sed SVM can effici
ently
classify different ultrason
i
c
flaw sig
nal
s with high recognition rate.
Ackn
o
w
l
e
dg
ements
This
wo
rk is
sup
porte
d by
the Natural
Scien
c
e F
o
u
ndation
of Ji
ang Xi Province
(No.
20122BAB201039), the F
ound
ation of
Key Laboratory of Nond
estructive Testi
ng (Nanchang
Han
g
kong University),
Mi
nistry
of Ed
ucatio
n (No.
ZD2
0122
90
03), the S
c
i
entific Resea
r
ch
Found
ation
o
f
Educatio
n
Dep
a
rtme
nt of Jian
gXi
Province
(No.
GJJ13
515
) a
nd the S
c
ient
ific
Re
sea
r
ch Fo
undatio
n of Chong
qing Mu
nicip
a
l Educa
t
ion Commi
ssion (No. KJ11
1218
).
Referen
ces
[1]
Lee
K, Estivil
l
-
Castro V. F
e
ature
e
x
tractio
n
a
nd
gati
n
g
techn
i
qu
es f
o
r ultr
ason
ic
shaft sig
n
a
l
classification.
Appl
ied S
o
ft Computi
n
g
. 20
0
7
; 7(1): 156-1
6
5
.
[2] K
y
u
ngm
i
L
ee.
F
eature
extrac
tion sch
e
m
es f
o
r ultras
onic
si
gna
l pr
ocessi
n
g
. Proce
edi
ngs
of the
5th
Internatio
na
l C
onfere
n
ce
on
Comp
uter Sci
e
nces
and
C
o
n
v
erge
nce I
n
for
m
ation
T
e
chno
log
y
. S
e
o
u
l.
201
0: 366-
372.
[3]
K
y
u
ngm
i Le
e. F
eature E
x
traction Sc
hem
es for
Ultrasonic Non-destructive T
e
sting Inspections.
Advanc
es in In
formati
on Sci
e
nces an
d Servi
c
e Scienc
es
. 2011; 3(3): 1
25-
135.
[4] Ahmed
Yamani,
Mo
hame
d
Derich
e.
Auto
matic
Det
e
ctio
n of
Hig
h T
e
mp
eratur
e Hy
drog
en Atta
c
k
defects from U
l
trason
ic A-sca
n Sign
als
. Pro
c
eed
ings
of the 4th Midd
le E
a
st NDT
Conferenc
e an
d
Exh
i
b
i
tion. Ne
w
York. 2
007:
58-6
1
.
[5]
Isa D, Rajk
um
ar R.
Pip
e
li
ne
Defect Detecti
on Us
ing
Sup
p
o
rt Vector Mac
h
in
es
. Proce
e
d
i
ngs
of th
e
6th Internati
o
n
a
l Co
nfere
n
ce
on Circ
u
its, S
y
stems, Electronics, Contro
l a
nd Sig
n
a
l
Proc
essin
g
. Cair
o.
200
7: 162-
168.
[6]
Vaclav
Matz,
Marcel
Krei
dl,
Radis
l
av
Sm
id.
Cl
assificati
on
of ultras
on
ic si
gna
ls.
Intern
ati
ona
l Jo
urn
a
l
of Materials a
n
d
Product T
e
ch
nol
ogy
. 20
06; 27(3-
4): 145-1
55.
[7]
Anastasso
po
ul
os A, N
i
kol
a
i
d
is V, P
h
il
ip
pi
di
s T
.
A Co
mparativ
e Stu
d
y
of P
a
ttern
Rec
ogn
itio
n
Algorit
hms
for
Classific
a
tio
n
o
f
Ultraso
nic Si
gna
ls.
Ne
ura
l
Co
mp
uting
& Appl
icatio
ns
.
1
999; 8(1): 53-
66.
[8]
Caccio
l
a M, C
a
lca
gno
S, Mo
rabito F
.
Com
p
utatio
n
a
l
intel
l
i
genc
e as
pects
for defect cl
as
sificatio
n
in
aero
nautic
co
mposites
b
y
u
s
ing
ultras
onic
puls
e
s.
IEEE
Transactions on Ultras
onics,
Ferroelectric
s
and F
r
eq
ue
ncy
Control
. 20
08;
55(4): 870-
87
8.
[9]
Caccio
l
a M,
C
a
lca
gno
S, Me
gali
G.
W
a
vele
t Coher
enc
e a
nd F
u
zz
y
S
ubtr
a
ctive C
l
uster
i
ng for
Defect
Classific
a
tio
n
i
n
Aer
o
n
autic
C
F
RP
. Procee
di
ngs
of Intern
ati
ona
l C
onfer
en
ce o
n
Comp
le
x, Inte
lli
ge
nt
and Soft
w
a
r
e
Intensiv
e S
y
ste
m
s.
Krako
w
. 2
010: 10
1-1
07.
[10]
Sambath S, Nagar
aj P, Selv
akumar N. Au
tomatic Defect Classi
ficati
on
in Ultrason
i
c NDT
Using
Artificial Intelligence.
Jour
na
l of nond
estructi
ve eval
uati
o
n
. 201
1; 30(1): 20
-28.
[11]
Ra
ymo
nd Sc
h
u
lz, H
e
rman
n
Rohl
in
g.
Auto
matic
Eva
l
uati
o
n of A
e
ros
pace
CF
RP Structu
r
es Bas
ed
o
n
Ultraso
nic Ech
o
Sign
als
. Proc
eed
ings
of the 9th Europ
e
a
n
C
onfer
ence
on
NDT
. Hambur
g. 2006: 1-
9.
[12]
Liu Sh
oush
an,
Chen Ya
n, Bi Liju
n.
Reco
nfig
urab
le arch
itec
ture of
ultraso
n
ic defect dete
c
tion bas
ed
on w
a
ve
let
pa
cket an
d
back
pro
p
a
gatio
n
artificial
n
eur
al
netw
o
rk
. Pro
c
eed
ings
of I
n
ternati
o
n
a
l
Confer
ence
on
Comput
er an
d Commu
nic
a
tion T
e
chno
lo
gi
es in Agr
i
cultu
r
e Engi
ne
erin
g
.
Chen
gdu
.
201
0: 498-
501.
[13]
Gang Qi, Alan
Barhorst, Java
d Hash
emi, Girish
Kama
la. Di
screte
w
a
ve
let decom
positi
on
of acousti
c
emission s
i
gnals from
carb
on-fiber-reinfor
c
ed c
o
mpos
ites.
Co
mp
osites
Scie
nce an
d T
e
chno
logy
.
199
7; 57(4): 38
9-40
3.
[14]
K
y
u
ngm
i Le
e. F
eature E
x
traction Sc
hem
es for
Ultrasonic Non-destructive T
e
sting Inspections.
Advanc
es in In
formati
on Sci
e
nces an
d Servi
c
e Scienc
es
. 2011; 3(3): 1
25-
135.
[15]
Elan
gova
n
M,
Sugum
aran
V,
Ram
a
cha
ndr
a
n
K,
R
a
vik
u
mar S. Effect of
SVM ker
nel
f
unctio
n
s
on
classificati
on o
f
vibration si
g
nals of a sin
g
l
e
poi
nt cutting
tool.
Expert Systems w
i
th Appl
icatio
ns
.
201
1; 38(1
2
): 1520
2-15
20
7.
[16]
A
y
at N, C
heri
e
t M, Suen C. Automatic mo
de
l
selecti
on for t
he o
p
timizati
on
of SVM kernel
s.
Pattern
Reco
gniti
on
. 2
005; 38(
10): 17
33-1
745.
[17]
Rush
di Sal
eh,
Martin-Val
divi
a. Exp
e
rime
nts
w
i
t
h
SVM to classif
y
o
p
i
n
io
ns in d
i
fferent doma
i
ns
.
Expert Systems w
i
th Applicati
ons
. 201
1; 38(
12): 147
99-
148
04.
[18]
Ergun Gum
u
s
,
Ni
yazi Ki
lic,
Ahmet Sertb
a
s,
Osman N
Ucan. Eva
l
u
a
tion
of face
recog
n
itio
n
techni
qu
es usi
ng PCA,
w
a
ve
lets an
d SVM.
Expert Syste
m
s w
i
th Ap
plic
ations
. 20
10
;
3
7
(
9
)
: 64
04
-
640
8.
[19]
Ron
an Co
llo
be
rt, Samy
Be
ng
io. SVMT
o
rch: Support Vect
or Machin
es for Larg
e
-Scal
e
Regressi
on
Probl
ems.
Jour
nal of Mach
in
e Lear
nin
g
Res
e
arch
. 200
1; 1(9
)
: 143-16
0.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Ultra
s
oni
c Fla
w
Signal Cl
assificatio
n
usi
n
g Wa
velet Transfo
rm
and Suppo
rt… (Y
u Wan
g
)
7547
[20]
Yiqi
ang Z
h
an,
Ding
g
a
ng S
h
e
n
. An ad
aptive
error
pe
nal
iza
t
ion meth
od fo
r trainin
g
a
n
efficient a
n
d
gen
eral
ize
d
SVM.
Pattern Recogn
ition
. 2
006;
39(3): 342-
35
0.
[21]
Qing-Qua
n
L
i
u
.
Coord
i
n
a
ted
Motion
Co
ntrol
of Auto
nomo
u
s
an
d Sem
i
aut
onom
ous M
obi
le Ag
ents
.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
ng.
2012; 1
0
(8): 1
929-
193
5.
[22]
Junfan
g Li, B
u
-ha
n
Z
han
g. Limitati
on of Sma
ll-
w
o
rld T
o
p
o
lo
g
y
for Ap
pli
c
ation i
n
Non-
domi
nate
d
Sorting D
i
ffere
ntial Ev
oluti
on.
T
E
LKOMNIKA Indon
esia
n
Journ
a
l of El
e
c
trical Eng
i
n
e
e
r
ing
. 2
012;
10(2): 40
0-4
0
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.