TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.5, May 2014, pp
. 3544 ~ 35
5
4
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i5.3587
3544
Re
cei
v
ed
Jun
e
19, 2013; Revi
sed
De
ce
m
ber
12, 201
3; Acce
pted Janua
ry 3, 20
1
4
Acoustic Emission Source Iden
tification Based on
Pattern Recognition Method
Zhigang Fen
g
*, Jiu Yao
Schoo
l of Auto
mation, She
n
y
ang Aer
o
sp
ace
Un
iversit
y
, Li
a
oni
ng, 11
013
6,
P.R.China
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: fzg1023
@
y
e
ah.net
A
b
st
r
a
ct
A new
pattern
recog
n
itio
n
method
bas
ed o
n
har
mo
nic w
a
vel
e
t packet
(HW
P
T
)
and h
i
erarch
y
supp
ort vector
machi
ne (
H
-
SVM) is pr
op
osed t
o
so
lve
the fatig
u
e
d
a
mag
e
i
dentifi
c
ation
pro
b
le
m
of
helic
opter c
o
mpon
ent. In this
appr
oach, HW
PT
is used to
extract the en
e
r
gy feat
ure of
acoustic e
m
iss
i
on
(AE) sign
als
o
n
differe
nt freq
uency
ba
nds
a
nd to r
educ
e t
he d
i
mens
io
nal
ity of orig
in
al d
a
ta featur
es. T
h
e
H-SVM classifi
er is use
d
to i
dentify the AE
source
typ
e
. A subset of the
exper
imenta
l
d
a
ta for know
n
A
E
source typ
e
is
used to trai
n the H-SVM cla
ssifier,
the re
ma
ini
ng s
e
t of data is
us
ed t
o
test the H-S
V
M
classifier.
Also
, the pr
essur
e
off exp
e
ri
me
n
t
on s
peci
m
en
of car
bon
fib
e
r
materi
als
is
inv
e
stigate
d
.
T
h
e
results in
dicate
that the propo
sed
ap
proac
h
can i
m
pl
e
m
e
n
t AE source ty
pe id
entificati
o
n effectively, a
n
d
has better p
e
rformanc
e on co
mp
utatio
nal
efficiency a
nd i
d
e
n
tificatio
n
accu
racy than w
a
ve
let packet (W
PT
)
feature extracti
on an
d RBF
ne
ural n
e
tw
ork classificati
on.
Ke
y
w
ords
:
harm
onic wavelet packet, support
vector
mac
h
in
e, RBF
ne
u
r
al n
e
tw
ork, acoustic e
m
issi
on,
pattern recognition
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
Due to th
e fa
ct that heli
c
o
p
ter movin
g
compon
ents
a
r
e ea
sy to p
r
odu
ce fatigu
e
damag
e
su
ch a
s
cracks,
whi
c
h
are
se
riou
sly en
dang
er th
e o
peratin
g
stabi
lity and safety of helicopte
r
, it
is n
e
cessa
r
y
to monito
r th
e initiation
of
cra
c
ks
a
nd to
ma
ster th
e d
e
veloping
tre
nd of th
e
cracks.
Acou
stic emi
ssi
on (AE) i
s
a noticea
ble
choi
ce of
no
nde
stru
ctive testing
meth
o
d
becau
se of
its
extremely hi
gh sen
s
itivity. AE has b
e
en p
r
oved
t
o
be a
very
sen
s
itive m
e
thod fo
r de
fect
recognitio
n
of compo
s
ite
material
s whi
c
h have b
e
e
n
use
d
in typical appli
c
atio
n area
s such
as
aero
s
p
a
ce, vehicl
e ind
u
st
ry and i
n
fra
s
tructure. Intere
st toward
s au
tomatic
re
cog
n
ition of
defe
c
t
types ba
sed
on their AE signal
s has in
cre
a
sed
an
d many re
cent studie
s
have
been pu
blish
ed
[1-3]. In the
AE techniq
u
e
, AE sou
r
ce
type identification i
s
u
s
e
d
to dete
r
mi
ne the m
ode
l of
fatigue dama
ge.
AE sou
r
ce type ide
n
tificati
on is a typica
l pro
b
lem
of pattern
re
cog
n
ition, whi
c
h
inclu
d
e
s
two
step
s, i.e. feature
e
x
traction
and
pattern
c
l
a
s
s
i
fic
a
tion
. AE s
i
gn
a
l
s
ar
e n
o
n
-
s
ta
tion
ar
y
sign
als,
so
the tra
d
itional
tech
niqu
es i
n
the
time
a
nd fre
que
ncy
domai
ns a
r
e
not
suitabl
e
for
analyzi
ng the
m
. The wave
let transfo
rm
(WT
) h
a
s b
e
en dem
on
strated a
s
an
al
terative tool for
feature extra
c
tion. The scaling ope
rati
on in wavele
t transform produ
ce
s a serie
s
of wav
e
let
function
s
with different
windo
w si
ze
s,
enabli
ng
m
u
lti-re
sol
u
tion
analysi
s
th
at is suited
for
rep
r
e
s
entin
g the non
-stati
onary
sign
al
s. A major d
r
awba
ck
of wavelet tra
n
s
form i
s
its l
o
w-
freque
ncy
re
solutio
n
in
th
e hig
h
frequ
ency
ra
nge.
The
wavelet
pa
cket tra
n
s
form
(WPT
), in
comp
ari
s
o
n
, further
de
co
mposes the
detailed
i
n
formatio
n of
the sig
nal,
whi
c
h h
a
s
b
een
su
ccessfully applie
d in the
featur
e extra
c
tion of sen
s
or fault and
machi
ne h
eal
th diagno
si
s [4-6].
Of the different types of wavelets d
e
vel
ope
d, the harm
oni
c wavelet po
sse
s
ses
com
pact
freque
ncy ex
pre
ssi
on an
d
has overco
me the limitat
ions of tradit
i
onal wavelet includi
ng en
ergy
leakage, i
n
fle
x
ible freq
uen
cy ba
nd
sel
e
ction
and
different
freq
ue
ncy re
solutio
n
s on differe
nt
levels
[7-8].
So in this
res
e
ar
ch, ha
rm
onic wavelet
pa
cket tran
sform
is u
s
e
d
to extract
the
feature of AE sou
r
ces.
Nume
ro
us p
a
ttern re
co
g
n
ition metho
d
s have b
e
e
n
develope
d
within the intelligent
system
s. Am
ong th
e met
hod
s, stati
s
tical l
ear
ning
method and
ANN are
mo
stly
used
i
n
AE
sign
als a
naly
s
is
of com
p
o
s
ite material
s.
ANN h
a
s
bee
n wid
e
ly appli
ed in AE sig
n
a
l cla
s
sificati
on
probl
em
s ba
sed on le
arni
n
g
pattern fro
m
example
s
or em
piri
cal
data mod
e
lin
g in the la
st two
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Em
issi
on Sou
r
ce Identification
Based o
n
Pattern Re
co
gnit
i
on Method
(Zhigan
g Fen
g
)
3545
decade
s [9
-1
2]. However,
as a
typical
machi
ne le
arning
cla
ssifie
r
, the ANN me
thod is ba
sed
on
the empi
rical
risk mi
nimization p
r
in
cipl
e, whi
c
h
ha
s been
re
co
gn
ized
as a m
e
thod that
can
not
alway
s
mini
m
i
ze th
e
actual
ri
sk.
The
tra
d
iti
onal
neu
ra
l network
app
roa
c
he
s have
limitation
s
o
n
gene
rali
zatio
n
giving rise
to modals t
hat can ove
r
-fit the data. This defi
c
ie
ncy is du
e to the
optimizatio
n
algorith
m
s u
s
ed
in A
N
N for the
se
l
e
ction
of pa
ramete
rs an
d the
statisti
cal
measures u
s
ed to
sel
e
ct t
he m
odel. M
e
anwhile, th
e
effectivene
ss
of the A
N
N m
e
thod
s i
s
clo
s
ely
related to th
e
numbe
r of training
sam
p
l
e
s. In mo
st
case
s, it is dif
f
icult to obtai
n larg
e samp
le
sets
of AE signal
s in com
posite m
a
teri
al and the
effectivene
ss of
the ANN m
e
thods
ca
n ha
rdly
be improved.
In orde
r to overcome the di
sadva
n
tage
s
of
ANN, sup
p
o
rt vector ma
chin
e (SVM) i
s
used
for cla
s
sifica
tion of AE
sou
r
ces. SV
M, based
o
n
statistical learni
ng the
o
ry, is gaini
ng
appli
c
ation
s
i
n
the area
s o
f
machin
e lea
r
ning,
co
m
put
er visio
n
an
d pattern
re
cog
n
ition be
cau
s
e
of the high
accuracy
an
d goo
d ge
ne
ralizati
on
ca
pability. The
SVM trainin
g
se
eks a
gl
obal
optimal soluti
on and
avoi
d over fitting
, so it has t
he ability to deal with
a large
numb
e
r of
features
. It is
very s
u
itable
for patte
rn re
cog
n
ition
with
small sa
mple
s.
In this p
ape
r, we
di
scuss the
appli
c
ati
on of h
a
rmonic wavel
e
t pa
cket in
feature
extraction
an
d hie
r
archy
suppo
rt vecto
r
machine
cla
ssifi
cation i
n
AE sou
r
ce type ide
n
tificati
on,
and verify the algorithm u
s
i
ng pre
s
su
re o
ff ex
periment on sp
ecim
en
of carb
on fibe
r material
s.
2. AE Sourc
e
Feature Ex
trac
tion Bas
e
d on Harmo
n
ic Wav
e
let Packe
t
2.1.
Harm
oni
c Wav
e
let
In esse
nce, the
wavelet transfo
rm
ch
aracteri
ze
s th
e
correl
ation o
r
simila
rity bet
wee
n
the
sign
al to be analyzed and
the mother wavelet func
ti
on. Such a correlation is
expre
s
sed by
the
wavelet
co
efficient
s a
s
so
ci
ated
with the
wavel
e
t tra
n
s
form,
whi
c
h
can
be
calcul
ated th
roug
h
a
correl
ation o
peratio
n bet
wee
n
the
sig
nal
x
(
t
) and the
co
njug
ate
)
(
t
w
of
the cho
s
en moth
e
r
wav
e
let
)
(
t
w
:
d
t
w
x
t
W
)
(
)
(
)
(
(1)
If the signal
x
(
t
) i
s
closely correlate
d
with the
mother
wave
let
)
(
t
w
; the wavelet
coeffici
ent
)
(
t
w
will be la
rge, i
ndicating a
g
ood m
a
tch
b
e
twee
n the
mother
wave
let and th
e
sign
al bei
ng
analyzed. As
a re
sult, the i
n
format
io
n e
m
bedd
ed i
n
the
signal
can
be extracte
d
by
analyzi
ng the
wavelet
coef
ficients
with l
o
cal m
a
xima.
At 1993, pro
f
essor
D.E.Newla
nd [13
-
1
5
]
from Camb
ri
dge
Universi
ty propo
sed
the ha
rm
o
n
ic
wavelet
whi
c
h h
a
s ideal ‘Box-like’
cha
r
a
c
teri
stic in freque
ncy
domain. In thi
s
study,
the h
a
rmo
n
ic
wav
e
let is chose
n
as the mot
her
wavelet, due
to the simplici
t
y of its expre
ssi
on in the freque
ncy dom
ain, and is d
e
f
ined by:
elsewhere
n
m
m
n
H
n
m
2
2
0
)
(
2
/
1
)
(
,
(2)
W
h
er
e
m
and
n
a
r
e th
e scal
e p
a
rameters. T
h
ese
pa
ram
e
ters a
r
e
real
but n
o
t
necessa
rily the integ
e
rs.
By taking
th
e
inverse F
o
u
r
ier tran
sform
of
)
(
,
n
m
H
, the time
domain
expre
ssi
on of
the harmo
nic wavelet is ob
tained a
s
:
t
m
n
i
t
im
t
in
t
h
n
m
)
(
2
/
)]
2
exp(
)
2
[exp(
)
(
,
(3)
If the harmo
nic wavelet is tran
slate
d
by a step
Z
k
m
n
k
)
/(
, i
n
whic
h
k
is the
transl
a
tion p
a
ram
e
ter, a
gene
rali
zed
expre
ssi
on
that is ce
ntered
at
)
/(
m
n
k
t
wi
th a
band
width of
)
(
2
m
n
can b
e
obtain
ed as:
)
)(
(
2
)]
(
2
exp[
)]
(
2
exp[
)
(
,
m
n
k
t
m
n
i
m
n
k
t
im
m
n
k
t
in
m
n
k
t
h
n
m
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3544 – 35
54
3546
Based
on the
gene
rali
zed
expre
ssi
on, the ha
rmoni
c
wavelet tra
n
sform of a si
g
nal
x
(t)
can b
e
perfo
rmed as:
1
0
,
)
(
)
(
)
(
)
,
,
(
N
r
n
m
m
n
k
r
h
r
x
N
m
n
k
n
m
hwt
(5)
Whe
r
e
)
,
,
(
k
n
m
hwt
is the harm
oni
c wavelet coefficien
t. By taking the Fouri
e
r transfo
rm
of Equation (5), an eq
uiva
lent expre
ssi
on of t
he harmonic
wavel
e
t transfo
rm i
n
the frequ
e
n
cy
domain
can b
e
expre
s
sed
as:
]
)
[(
)
(
)
,
,
(
,
m
n
H
X
n
m
HWT
n
m
(6)
Whe
r
e
)
(
X
is the Fourie
r transfo
rm of the sign
al
x
(
t
), and
]
)
[(
,
m
n
H
n
m
is the
conj
ugate of
]
)
[(
,
m
n
H
n
m
, whi
c
h i
s
the
Fouri
e
r tran
sform of th
e ha
rmoni
c
wavel
e
t at the
scal
e
(
m
,
n
). Since
the h
a
rm
oni
c
wavelet h
a
s
comp
act
f
r
eque
ncy
expression,
as sh
own
in Eq
uat
ion
(2), th
e
harm
onic wavelet
transfo
rm
ca
n be
read
ily
obtaine
d th
ro
ugh
a p
a
ir of
Fouri
e
r tran
sf
orm
and inverse F
ourie
r tran
sfo
r
m ope
ration
s.
H
Figure 1. Algorithm for Im
plementin
g
the Harmoni
c
Wavelet Tran
sform [8]
As sho
w
n i
n
Figu
re 1,
after ta
king
the
Fou
r
ier tran
sform
of a
si
gnal
x
(
t
) to obtain its
freque
ncy d
o
m
ain exp
r
e
s
sion
)
(
X
, the inner pro
duct
)
,
,
(
n
m
HWT
of
)
(
X
and the
co
nju
gate
of the ha
rmo
n
ic
wavelet
]
)
[(
,
m
n
H
n
m
at the scal
e (
m
,
n
)
is
cal
c
ulated. Fin
a
ll
y, the harm
o
nic
wavelet tra
n
sform of the
signal
x
(
t
), de
noted a
s
)
,
,
(
k
n
m
hwt
is obtaine
d by takin
g
the inv
e
rse
Fouri
e
r tra
n
sf
orm of the inn
e
r produ
ct
)
,
,
(
n
m
HWT
.
2.2.
Harmoni
c Wav
e
let Packet
Algorith
m
The scale p
a
ram
e
ter
m
and
n
dete
r
mine the
ba
ndwi
d
th that
the h
a
rm
oni
c
wavelet
c
o
vers
. Similar to the Wavelet Pack
et Trans
f
or
m
(WPT), the n
u
m
ber
of freq
ue
ncy
sub
-
ba
nd
s for
the Ha
rmo
n
i
c
Wavelet P
a
cket T
r
an
sf
orm
(HWPT
)
has to be
s
po
we
rs of
2, in which
s
corre
s
p
ond
s
to the de
com
positio
n level
for WPT.
A
c
cordingly, the
sign
al can b
e
de
comp
ose
d
into 2
s
freque
ncy su
b-b
and
s with the ba
ndwi
d
th in He
rtz for ea
ch
sub-b
and d
e
fin
ed by:
s
h
band
f
f
2
/
(7)
Whe
r
e
h
f
is the
highe
st freq
uen
cy com
p
o
nent of the si
gnal to be
an
alyzed. Sin
c
e
the
band
width
of
the h
a
rm
oni
c
wavelet i
s
)
(
2
m
n
, sele
ction
of
the valu
es f
o
r
m
an
d
n
of
the
HWPT has
to s
a
tis
f
y the following c
o
nditions
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Em
issi
on Sou
r
ce Identification
Based o
n
Pattern Re
co
gnit
i
on Method
(Zhigan
g Fen
g
)
3547
band
f
m
n
2
)
(
2
(8)
Thus th
e harmonic
wavele
t packet coeff
i
cient
s
hwpt
(
s
,
i
,
k
)
can be
obtaine
d as:
)
,
,
(
)
,
,
(
k
n
m
hwt
k
i
s
hwpt
(9)
Whe
r
e
s
is the decomposition level,
i
is the index
of the sub-b
and,
k
is index of the
coeffici
ent. The paramete
r
s
m
and
n
need to satisfy the followi
ng condition:
1
2
,
,
1
,
0
2
)
1
(
)
1
(
2
s
s
h
band
s
h
band
i
f
i
f
i
n
f
i
f
i
m
(10)
2.3.
AE Signal Featur
e Extra
c
tion
With the AE
sign
al bein
g
decompo
se
d
into a
num
be
r of sub-ban
d
s
, the featu
r
e
s
can b
e
extracted fro
m
the harmo
nic wavelet packet co
effi
ci
ents in ea
ch
sub
-
ba
nd to p
r
ovide inform
ation
on the
type
of AE source. The
fact
of differe
nt
energy di
stri
bution
of si
g
nals in
different
freque
ncy
ba
nds
mu
st be
cau
s
e
d
by th
e differe
nce i
n
formatio
n co
ntained i
n
the
sign
als. F
o
r t
he
AE sign
al, b
e
ca
use of th
e differe
nt A
E
so
urce f
e
a
t
ures,
the
ch
ara
c
teri
stic e
nergy
dist
rib
u
tion
coeffici
ent of
harmoni
c
wavelet pa
cket
is
sel
e
ct
e
d
as th
e featu
r
es. T
he e
n
e
r
gy co
ntent of
a
sign
al can
be
cal
c
ulate
d
, b
a
se
d on
the
coeffici
ent
s o
f
the sig
nal’
s
tran
sform. In
the case of
a
HWPT, the
coefficient
s
hwpt
(
s
,
i
,
k
) q
u
a
n
tify the energy associ
ated
with ea
ch sp
ecific
sub
-
b
a
nd.
The detail
s
of feature extra
c
tion
procedu
re are sh
own as follo
ws.
Step 1
: No
rm
alizin
g the AE signal u
s
in
g:
)]
(
[
~
1
X
X
X
E
D
(11)
Whe
r
e
X
is the AE signal,
)
(
X
E
an
d
D
is the mean
and stan
dard
deviation of
X
.
Step 2
: Decompos
ing
X
~
with
four levels
of harmoni
c
wavelet
pa
cket tran
sform, an
d
getting the
coefficient
s ve
ctors
of the
sixteen
nod
e
s
,
15
,
4
,
4
,
4
H
H
H
,
,
,
1
0
, where
i
H
,
4
r
e
pr
es
e
n
t
s
1
,
1
,
0
)
,
,
4
(
N
k
k
i
hwpt
, in which
N
is the length o
f
AE signal.
Step 3
: Cal
c
u
l
ating the ene
rgy of each n
ode an
d normalizin
g the
m
.
N
j
j
i
i
i
H
dt
H
EH
1
2
,
,
4
2
,
4
,
4
(12)
15
0
2
,
4
,
4
,
4
|
|
i
i
i
i
EH
EH
EH
(13)
Step 4
: the feature ve
ctor
]
,
,
,
,
[
15
,
4
2
,
4
1
,
4
0
,
4
EH
EH
EH
EH
T
is used to i
dentify the AE
s
o
urce types
.
3. AE Sourc
e
Identifica
ti
on Using Hie
r
arch
y
Support Vec
t
or M
achine (H-S
VM) Classi
fier
AE source identification is a
typical p
r
o
b
lem of
patte
rn
re
cognition with sm
all
sampl
e
,
bec
au
se in most
ca
se
s,
it
is dif
f
i
cult
to obt
ai
n larg
e sampl
e
set
s
of AE sign
als in compo
s
ite
material
to train the
cl
assi
fiers. In
this
pape
r,
supp
o
r
t vecto
r
m
a
chine
(SVM) i
s
sele
cted
a
s
the
basi
c
cla
s
sifier, becau
se it provide
s
a novel appr
oa
ch to the two-category cl
assi
fication probl
em
with goo
d sm
all sampl
e
ge
neratio
n [16-17].
The co
ncept of
co
mpo
s
ite damag
e wa
s prop
osed
by Professo
r
K. L.
Reif
snid
er [18]
at
1977 du
rin
g
his re
se
arch
on com
p
o
s
ite
fatigue dam
age. The
r
e a
r
e four da
ma
ge types of fiber
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3544 – 35
54
3548
comp
osite, i.e. matrix cra
cki
ng, interfa
c
ial
de
bon
din
g
, delaminati
on and fibe
r fractu
re. Th
e task
of AE sou
r
ce
identificatio
n i
s
to di
stingui
sh
the
dam
a
ge type u
s
ing
the AE sign
a
l
s. Becau
s
e t
h
e
feature
s
of
AE sign
al fo
r inte
rfaci
a
l
debo
nding
a
nd del
amin
ation a
r
e
simil
a
r a
nd th
ese
two
damag
e type
s are always
occurre
d
at the sam
e
time
for the
carbon
fiber mate
rial
s, in this p
a
p
e
r,
three
dama
g
e
types
are
studie
d
, i.e. matrix cra
cki
ng, interfa
c
ial
debo
ndin
g
a
nd fibe
r fra
c
t
u
re.
Obviou
sly, AE source id
en
tification is a
multi-cl
assification pro
b
lem
.
There a
r
e two stan
da
rd a
ppro
a
che
s
to
con
s
truct a
n
d co
mbin
e th
e re
sult
s fro
m
bina
ry
cla
ssif
i
e
r
s f
o
r
a
C
-cl
a
ss
probl
em. Th
e first o
ne i
s
the o
n
e
-
vs-re
st meth
od
, in whi
c
h e
a
ch
cla
ssifie
r
di
stingui
she
s
o
n
e
cla
ss f
r
om th
e other
C
-1
classe
s, and t
he cl
ass la
be
l of the input
is
deci
ded
by wi
nner-ta
k
e-all method
[
19].
Each
cl
assifie
r
n
eed
s to
be
trained
on
the
wh
ole t
r
aini
n
g
set, and
there is n
o
g
uarantee that g
ood di
sc
rimi
n
a
tion exist
s
betwe
en o
n
e
cla
s
s and
the
remai
n
ing
cl
a
s
ses.
The
se
con
d
stand
ard ap
proa
ch t
o
combi
ne
bi
nary
cla
s
sifie
r
s is the
one
-vs
-
one m
e
thod,
in whi
c
h
the
deci
s
io
n is m
ade by m
a
jo
rity voting stra
tegies. T
h
is
requires traini
ng
and testin
g of
C
(
C
-1
)/2 bin
a
ry cla
ssifie
r
s. This app
roa
c
h is p
r
ohi
bitive when
C
is
large [20].
Thus, we ch
o
s
e a bina
ry hi
era
r
chical cla
ssifi
cation
structure in Figu
re 2. Each no
de is a
binary cla
ssifi
er.
Co
arse separation
a
m
ong cla
s
se
s
occurs in th
e
begin
n
ing
(a
t upper l
e
vels) in
the hierarchy
and a finer
cla
ssifi
cation
result is
obta
i
ned in late
r (at lower lev
e
ls). At the top
node,
we
divi
de the
o
r
igin
al 4
cla
s
se
s i
n
to two
small
e
r
gro
u
p
s
of
cla
s
ses (m
acro-cla
sse
s
). T
h
is
clu
s
terin
g
pro
c
ed
ure i
s
re
p
eated in sub
s
eq
uent
level
s
, until there
is only one
cl
ass in the fin
a
l
sub
-
g
r
ou
p. T
h
is
hierarchi
c
al st
ru
cture
d
e
com
p
o
s
e
s
t
he p
r
obl
em i
n
to 3 bi
na
ry su
b-p
r
obl
ems.
For
testing, only about
3
log
2
classifi
ers a
r
e
requi
red to traver
se
a path from top to bottom.
Figure 2. Hierarchical Multi
-
cla
s
sificati
o
n
Structure for AE Source Id
entification
In this
pape
r,
the
standa
rd
k
-mea
ns
clu
s
terin
g
i
s
u
s
e
d
to de
sig
n
t
he bin
a
ry
hie
r
archi
c
al
s
t
r
u
c
t
u
r
e
,
as
s
h
ow
n in
F
i
gu
r
e
2
.
SVM1
is
us
ed
to
cla
ssify
Normal
vs othe
r th
re
e patterns,
SVM2
is use
d
to cl
assify Matrix
cra
c
king vs
Fibers
fra
c
ture, Interfacial
debo
nding, S
V
M3 is use
d
to
cla
ssify Fibe
rs fra
c
ture vs I
n
terfaci
a
l deb
ondin
g
.
In the trainin
g
ph
ase, the
traini
ng
sa
m
p
les a
r
e
gro
uped
a
c
cording to
Figu
re
2. The
n
SVM1 to SVM3 a
r
e train
ed u
s
ing
the
co
rrespondi
ng g
r
ou
p of t
r
ainin
g
sampl
e
s. After th
at, by
inputting the feature ve
ctor
into the traine
d mult
i-cl
assif
i
er, the AE so
urce type can
be identified.
4. Experiment and Resul
t
s
4.1.
Experimental Setup
In orde
r to v
e
rify the pro
posed meth
o
d
, a
se
rie
s
of pre
s
sure
off experime
n
ts were
carrie
d o
u
t o
n
the
spe
c
im
en of
carb
on
fiber mate
ri
als,
whi
c
h i
s
one
of th
e
comm
only u
s
ed
material
s
of
helicopter m
o
ving
com
p
o
nent. Th
e A
E
sig
nal m
e
asu
r
em
ent
system i
s
sh
own
schemati
c
ally
in Fi
gu
re
3.
Figure 4
sho
w
s the
pr
essure
off expe
ri
ment p
r
o
c
e
s
s on
carb
on
fiber
spe
c
ime
n
. Th
e dimen
s
ion
s
of all sampl
e
s a
r
e all 41
8mm×120m
m
×
2mm. T
w
o
AE sensors
are
distrib
u
ted
on
the
ca
rbo
n
fi
ber spe
c
imen
, one
is
80m
m di
stan
ce
a
w
ay from th
e
central lin
e
of the
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Em
issi
on Sou
r
ce Identification
Based o
n
Pattern Re
co
gnit
i
on Method
(Zhigan
g Fen
g
)
3549
spe
c
ime
n
in up dire
ction, and the othe
r is 80mm
di
stan
ce away from t
he ce
ntral line of the
spe
c
ime
n
in
down di
re
ctio
n. The
ce
ntra
l point
of the
sp
ecim
en i
s
the force
poi
nt. The lo
adi
ng
spe
ed of the
pre
s
sure
off experime
n
t is 500
N/
s. Fi
gure
5 is the
AE signal a
c
qui
sition
system
employed
on
-site. Th
e si
g
nal conditioni
ng is
per
fo
rm
ed by the p
r
e-am
plifiers. The conditio
n
ed
sign
al (with a
gain of 40dB
) is fed to the
main data
-
a
c
quisitio
n
boa
rd in whi
c
h th
e AE waveforms
and
paramet
ers a
r
e
sto
r
e
d
. The
in
stru
ments an
d e
q
u
ipment
s u
s
e
d
in th
e
expe
riment
s a
r
e
li
sted
belo
w
:
(1) MTS el
ect
r
o-hydra
u
lic l
oadin
g
syst
e
m
(MTS 810
material te
st system
).
(2) Vallen AM
SY-5 AE si
gn
al acqui
sition
system
with
1
6
chann
el a
n
d 16
-bit, 10
-MHz AD
conve
r
ter o
n
each ch
ann
el
.
(3) T
w
o Valle
n VS150-M A
E
sensors.
(4) T
w
o Vallen AEP4 pre-amplifiers (20-2000K
Hz).
(5) Valle
n AE application software Vall
e
n
Visual AE.
(6) Noteb
o
o
k
comp
uter.
Figure 3. Sch
e
matic of the
AE Measu
r
e
m
ent System
Figure 4. Pre
s
sure
Off Exp
e
rime
nt Process on
Carbon Fi
ber
Specim
en
Figure 5. AE
Signal Acq
u
isition System
Employed On
-site
The sampli
ng
rate of the a
c
qui
sition
system is 1M
Hz. In orde
r to a
c
qui
re all AE
sign
al
s
durin
g the p
r
essu
re
off pro
c
e
ss, AM
SY-5 wo
rks in co
ntinuo
us a
c
q
u
isitio
n mode. T
h
ree
spe
c
ime
n
s of
carbon fibe
r material
s wi
th the same
dimen
s
ion
s
are u
nde
r the pre
s
sure o
f
f
experim
ent. For ea
ch AE source type, 50 grou
ps of d
a
ta are gath
e
red.
Figure 6, Figure 7 a
nd Figure 8 sho
w
s t
he AE signal and its spe
c
trum of
matrix
cra
c
king, the
AE signal and its spe
c
trum of
interfacial deb
ondi
n
g
and the AE signal an
d
its
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3544 – 35
54
3550
spe
c
tru
m
of f
i
ber fra
c
ture,
respe
c
tively. The th
ree
AE sig
nal
are al
l no
rmali
z
ed
usin
g Equ
a
tion
(11
)
. Th
e
spe
c
trum
of AE
sign
al fo
r the
thre
e type
s i
ndicates that
the en
er
gy di
stributio
n
of th
e
three type
s i
s
different. T
h
e en
ergy
of
matrix cra
cki
ng i
s
mai
n
ly i
n
lo
w fre
que
ncy b
and
an
d the
freque
ncy
ba
nd i
s
ve
ry na
rrow.
The
en
e
r
gy of i
n
te
rfa
c
ial de
bon
ding
dist
ribute
s
i
n
wid
e
frequ
en
cy
band. The freque
ncy ba
n
d
of fiber fra
c
ture i
s
wid
e
r
than matrix
cra
cki
ng but
narrower th
an
interfaci
a
l de
bondi
ng.
Figure 6. AE
Signal and Its Spectru
m
of Matrix Cra
c
ki
ng
Figure 7. AE
Signal and Its Spectru
m
of Interfacial
De
bondi
ng
Figure 8. AE
Signal and Its Spectru
m
of Fiber F
r
a
c
ture
0
100
200
300
400
500
600
700
800
900
1
000
-10
0
10
Ti
me
(
u
s
)
S
i
gnal
(
m
V
)
AE s
i
gn
al (M
at
ri
x
C
r
a
c
k
i
ng
)
0
50
100
150
200
250
300
350
400
450
500
0
100
200
300
F
r
equ
enc
y
(k
H
z
)
A
m
pl
i
t
ude (
m
V
)
Spec
t
r
um
of
AE s
i
gna
l (M
at
rix
C
r
ac
k
i
ng)
0
100
200
30
0
400
500
600
700
800
900
1000
-1
0
0
10
Ti
me
(
u
s
)
S
i
gnal
(
m
V
)
A
E
s
i
gnal
(I
nt
erf
a
c
i
al
D
ebondi
ng)
0
50
100
15
0
200
250
300
350
400
450
50
0
0
100
200
300
F
r
e
quenc
y
(k
H
z
)
A
m
pl
i
t
ude (
m
V
)
S
pec
t
r
um
of
AE s
i
gnal (
I
nt
er
f
a
c
i
a
l
D
ebo
nding
)
0
10
0
20
0
30
0
40
0
50
0
60
0
70
0
80
0
90
0
10
00
-10
0
10
Ti
m
e
(
u
s
)
Si
g
n
a
l
(
m
V)
AE s
i
gn
al
(F
i
b
e
r
F
r
ac
t
u
re
)
0
50
10
0
15
0
20
0
25
0
30
0
35
0
40
0
45
0
500
0
10
0
20
0
30
0
F
r
eq
uen
c
y
(k
H
z
)
A
m
p
lit
u
d
e
(
m
V
)
Sp
ec
t
r
um
of
AE
s
i
gn
al
(F
ib
er F
r
a
c
t
u
re
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Em
issi
on Sou
r
ce Identification
Based o
n
Pattern Re
co
gnit
i
on Method
(Zhigan
g Fen
g
)
3551
AE source identification is a
typical p
r
o
b
lem of
patte
rn
re
cognition with sm
all
sampl
e
,
bec
au
se in most
ca
se
s,
it
is dif
f
i
cult
to obt
ai
n larg
e sampl
e
set
s
of AE sign
als in compo
s
ite
material
to train the
cl
assi
fiers. In
this
pape
r,
supp
o
r
t vecto
r
m
a
chine
(SVM) i
s
sele
cted
a
s
the
basi
c
cla
s
sifier, becau
se it provide
s
a novel appr
oa
ch to the two-category cl
assi
fication probl
em
with goo
d sm
all sampl
e
ge
neratio
n [16-17].
4.2.
Feature Extrac
tion
Firstly the experime
n
t of feature extraction
is p
e
rformed a
c
cording to the algorith
m
given in the se
ction 1.3. Table 1 sho
w
s t
he featu
r
e node
s and
their freque
ncy ran
g
e
s
. The
freque
ncy
ba
nd for ea
ch
feature
nod
e
is 3
1
.25K
Hz. Figure 9
shows the
no
rmalize
d
e
nergy
distrib
u
tion fo
r different AE sou
r
ces at 15
frequen
cy su
b-ba
nd
s.
Table 1. Feat
ure Node
s an
d Their F
r
eq
u
ency Band
Range
Feature
nodes
Freque
nc
y
ban
d
No
Freque
nc
y
ban
d
range (K
Hz)
H4,0
0
0~31.25
H4,1
1
31.25~62.5
H4,2
2
62.5~93.75
H4,13
13
40.625~43.75
H4,14
14
43.75~468.75
H4,15
15
468.75~500.00
0
0.1
0.2
0.3
0.4
0.5
012
34
567
89
1
0
1
1
1
2
1
3
1
4
1
5
M
atrix c
racking
Fi
ber frac
ture
Inte
rfacial
debondi
ng
normal
Figure 9. Normalize
d
Energy Distrib
u
tio
n
for
Differe
nt AE Source a
t
15 Frequ
en
cy Sub-b
and
s
As sh
own in
Figure 9, the
energy
distri
b
u
tion of no
rm
al state
is
ap
proximately u
n
iform in
every freq
ue
ncy ba
nd, b
e
c
au
se th
e AE
sign
al of
n
o
rmal state i
s
a
pproxim
ately white n
o
ise.
The
energy distri
b
u
tion of Matri
x
cra
cki
ng is
mainly
co
nce
n
trated in fre
quen
cy ban
d 3, 4 and 5. T
he
energy distri
b
u
tion of Fibe
rs brea
king i
s
mainly co
nce
n
trated in
fre
quen
cy ban
d
4, 5, 6, and 7
.
The e
n
e
r
gy d
i
stributio
n of i
n
terface
sep
a
r
ation i
s
bro
a
d
, app
roximat
e
ly from freq
uen
cy ban
d 3
to
8. The
r
efo
r
e
combi
n
ing
ab
ove an
alysi
s
, the AE
sou
r
ce type
s
ca
n
be
distin
gui
she
d
u
s
in
g t
he
harm
oni
c wa
velet packet energy features.
4.3.
AE Sour
ce Identific
a
tion Using H-SVM Classifi
er
After the experime
n
t of feature extra
c
ti
on,
two grou
ps of data a
r
e a
c
qui
red,
i.e. the
training
sam
p
les
and th
e
testing
data.
20 g
r
ou
ps
o
f
data for
ea
ch type
are
use
d
a
s
trai
n
i
n
g
sampl
e
s, a
n
d
the other 30
grou
ps of da
ta fo
r each type are used
as testin
g dat
a. The H-SVM
cla
ssifie
r
is trained u
s
in
g the traini
ng sa
mples
acco
rd
ing to se
ction
2. The ke
rnel
function
s of the
three
SVM
s
i
n
the H-SVM
cla
s
sifier are
all sele
cted
as RBF ke
rn
els, sho
w
n
a
s
Equ
a
tion (1
4).
The ke
rn
el wi
dth para
m
ete
r
,
, for each S
V
M is sele
cte
d
as 1.0.
)
/
exp(
)
,
(
2
2
j
i
j
i
X
X
X
X
K
(14)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 5, May 2014: 3544 – 35
54
3552
Table 2
sho
w
s the AE sou
r
ce ide
n
tificati
on re
sult u
s
in
g HWPT
and
H-SVM. The
results
indicate that the pro
p
o
s
ed
approa
ch can
implement
AE source type
identification
effectively
Table 2. AE Source Identi
f
ication Resul
t
Using
HWP
T
and H-SVM
AE source t
y
pe
Test sample No (
C
orrect N
o
)
Identification rate (%)
Matr
ix
cr
acking
30 (28)
93.33
Fibers breaking
30 (27)
90.00
Interface separ
at
ion
30 (28)
93.33
Normal
30 (30)
100.00
In orde
r to verify the advantage
s of th
e HW
PT feature extra
c
tio
n
, the com
p
a
r
iso
n
of
WPT feature extraction
an
d H-SVM cl
a
ssifie
r
with
HWPT and
H-SVM is studi
ed. For the
WPT
feature extra
c
tion, the wa
velet function
is se
le
cted a
s
Db1
0
, and the decomp
o
sing level is also
4. Similar to
the HWPT fe
ature
extra
c
tion, the
featu
r
e ve
ctor
of
WPT i
s
al
so
the no
rmali
z
ed
energy in
ea
ch frequ
en
cy ban
d. Tabl
e
3
sho
w
s th
e
com
p
a
r
ison
of feature
extractio
n
time f
o
r
HWPT
an
d
WPT. Th
ese
algo
rithm
s
are
all im
ple
m
ented
by
Matlab 7.1
o
n
Intel
Dual
Core
2.4GHz a
nd
1G RAM. Th
e re
sults i
ndi
cate that
the feature
extra
c
tion spee
d of
HWPT
i
s
over
nine time
s a
s
qui
ck
as th
e WPT. Such an advant
a
ge of the HWPT over
WPT is even
more
appreci
able
whe
n
the de
compo
s
ition le
vel is lar
ger t
han 4, be
ca
u
s
e of the a
d
d
i
tional re
cu
rsi
v
e
operation
s
ne
eded for
WP
T.
Table 3. Feat
ure Extra
c
tion Time Com
p
arison of HWPT and WPT
Feature
ext
r
action method
Feature
ex
t
r
action time for 50 sa
mples (s)
HWPT 0.65
WPT 5.88
Table 4 sho
w
s the com
p
a
r
i
s
on of AE so
urce
identification re
sult for HWP
T
and
H-SVM
with
WPT an
d H-SVM. Th
e re
sult
s indi
cate that th
e i
dentificatio
n rate of HWPT
and
H-SVM i
s
a
little higher than WPT an
d H-SVM. HWPT over
co
mes the en
e
r
gy leakage
sho
r
tco
m
ing
of
traditional
wa
velet, and ca
n extract t
he
energy feature more a
c
curacy.
Table 4. AE Source Identi
f
ication Comp
ar
ison of HWPT and H-SVM with WPT
and H-SVM
AE source t
y
pe
Identification rate (%)
HWPT and H
-
SV
M
WPT and H-SVM
Matrix cracking
93.33
86.67
Fibers breaking
90.00
83.33
Interface separ
at
ion
93.33
90.00
Normal
100.00
100.00
In ord
e
r to ve
rify the adva
n
t
ages of the
H-SVM
cla
s
si
fication, the
compa
r
ison
of HWPT
feature extra
c
tion a
nd H-SVM classifi
er with
HWP
T
and
RBF
neural net
wo
rk i
s
stu
d
ied.
The
RBF neu
ral n
e
twork i
s
a three laye
r ne
twork. The
first layer is th
e input layer,
and the se
cond
layer has
RADBAS neurons
as well
as
the
output layer has
PURELI
N neurons
.
For t
he
cla
ssifi
cation
of AE so
urce
type, the in
put ne
uro
n
s
are
16, e
a
ch
for o
ne fe
ature. T
he
out
put
neuron
s a
r
e
4. Figu
re
10
sho
w
s the
stru
ctur
e
of RBF
ne
ural
network for AE
so
urce
cla
ssifi
cation.
Table 5 sh
o
w
s the
relatio
n
shi
p
betwee
n
RBF output
and AE source type.
Table 5. Rel
a
tionshi
p between RB
F O
u
tput and AE Source Type
AE source t
y
pe
Output of
RB ne
ural net
w
o
rk [
Y
1
Y
2
Y
3
Y
4
]
Matr
ix
cr
acking
[1 0 0 0]
Fibers breaking
[0 1 0 0]
Interface separ
at
ion
[0 0 1 0]
Normal
[0 0 0 1]
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Acou
stic Em
issi
on Sou
r
ce Identification
Based o
n
Pattern Re
co
gnit
i
on Method
(Zhigan
g Fen
g
)
3553
Figure 10. Structure of RB
F Neu
r
al Ne
twork for AE Source
Class
i
fic
a
tion
The trai
ning
pro
c
e
s
s of
the RBF n
e
u
ral n
e
two
r
k is sho
w
n a
s
follo
ws. Ini
t
ially the
RADBAS layer has no neuron
s, whil
e the input layer has 16 n
eurons and the output layer has
4
neuron
s. The
followin
g
ste
p
s a
r
e
rep
eat
ed until
the
n
e
twork'
s m
e
a
n
sq
uared e
r
ror fall
s b
e
lo
w
GOAL or the
maximum nu
mber of ne
urons a
r
e rea
c
h
ed:
(1) T
he network i
s
sim
u
lat
ed usi
ng the trainin
g
sa
mpl
e
s.
(2) T
he input
vector
with the greate
s
t error is foun
d.
(3) A RA
DBAS neuron is added
with wei
ghts equal to that vector.
(4) T
he PURELIN layer weights a
r
e red
e
sig
ned to mi
nimize e
r
ror.
Table 6 sho
w
s the
comp
arison of trai
ning time for H-SVM and
RBF neu
ral netwo
rk.
These alg
o
rit
h
ms a
r
e all i
m
pleme
n
ted
by Matlab
7.1 on Intel Du
al Co
re 2.4G
Hz a
nd 1
G
RAM.
The
re
sults in
dicate
that th
e traini
ng
sp
e
ed of
H-SVM
is
about th
re
e times a
s
q
u
ick a
s
the
RB
F
neural netwo
rk, which verify that the conver
g
e
n
c
e
perfo
rman
ce
of SVM is better than RBF
neural network.
Table 6. Training Time
Co
mpari
s
o
n
of H-SVM an
d RBF Ne
ural
Network
Classification method
Training time for
20 samples (s)
H-SVM
0.120
RBF neu
ral net
work
0.358
Table 7 sho
w
s the com
p
a
r
i
s
on of AE so
urce
identification re
sult for HWP
T
and
H-SVM
with HWPT
a
nd RBF
ne
ural network. T
he results in
dicate th
at th
e identificatio
n rate
of HWPT
and
H-SVM i
s
hig
her tha
n
HWPT an
d RBF n
e
u
r
a
l
netwo
rk, which ve
rify that SVM is very
suitabl
e for cl
assificatio
n
with small traini
ng sam
p
le
s.
Table 7. AE Source Identi
f
ication Comp
ar
ison of HWPT and H-SVM with HWPT
and RBF
N
e
ur
a
l
Ne
tw
or
k
AE source t
y
pe
Identification rate (%)
HWPT and H
-
SV
M
WPT and RBF
n
eural net
w
o
rk
Matr
ix
cr
acking
93.33
83.33
Fibers breaking
90.00
80.00
Interface separ
at
ion
93.33
66.67
Normal
100.00
90.00
5. Conclusio
n
In this pap
er,
the HWPT feature
extra
c
tion
and
H-S
V
M cla
ssifie
r
are firstly ap
plied to
the AE source i
dentificat
ion. The
ex
perim
ental
system is bui
lt up an
d th
e pressu
re
off
experim
ents
on spe
c
imen
of carbon fi
ber m
a
terial
s is carried
o
u
t. The com
pari
s
on
re
sul
t
s of
HWPT
and
H-SVM with
WPT and
H-SV
M indi
cate th
at the propo
sed ap
pro
a
ch
can i
m
plem
e
n
t
AE source t
y
pe identification effectively, and
it has bette
r pe
rforma
nce o
n
com
putatio
nal
efficien
cy and identificatio
n accu
racy than WPT fea
t
ure extra
c
tio
n
. The comp
arison re
sult
s of
HWPT
an
d
H-SVM with
HWPT a
nd
RB
F ne
ural
net
work indi
cate
that the
pro
posed
app
ro
ach
has
better p
e
r
forma
n
ce on
comp
utation
a
l effici
en
cy and ide
n
tifica
tion accu
ra
cy than the RB
F
neural n
e
two
r
k cla
s
sificati
on. Th
e p
r
o
posed
app
ro
ach
is very
suitabl
e for
small
sa
mpl
e
s
∩
∩
∩
…
…
…
X
1
X
2
X
1
Y
1
Y
2
Y
4
Evaluation Warning : The document was created with Spire.PDF for Python.