Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8
, p
p.
4221
~
4229
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp4221
-
42
29
4221
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
A Detail
Stu
dy of
Wavelet
Families
for EM
G Patt
ern
Recognit
ion
Jing
w
ei
T
oo
1
,
A.
R.
A
bdulla
h
2
,
N
orhas
himah M
oh
d
S
aad
3
, N
M
oh
d
Ali
4
, H
Musa
5
1
,2,4
Fakult
i
Kejur
ute
ra
an
Ele
k
tri
k
,
Univer
sit
i
T
ekn
ika
l
Malay
si
a
M
el
ak
a, Ma
lay
si
a
2
,3
Cent
re
of Exc
el
l
enc
e
in
Robot
ic
and
Industr
ia
l
Autom
at
ion,
Uni
ver
siti
Te
kn
ika
l
Malay
s
ia Mel
ak
a,
Ma
lay
si
a
3
Fa
kult
i
K
ej
urut
era
an
E
le
k
tron
i
k
dan
K
ej
urut
eraa
n
Kom
pute
r
,
Uni
ver
siti
Te
kn
ika
l
Malay
s
ia Mel
ak
a,
Ma
lay
si
a
5
Fakult
i
P
engur
u
san
Te
kno
logi d
an
T
eknousa
haw
ana
n
Technol
og
y
,
Univer
si
ti T
ek
nika
l
Malay
si
a M
el
aka,
Ma
lay
si
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Feb
23
, 201
8
Re
vised
Ju
l
19
,
201
8
Accepte
d
Aug
7
, 2
01
8
W
ave
le
t
tra
nsfor
m
(W
T)
has
recent
l
y
dr
awn
the
at
t
ent
ion
of
the
rese
arc
h
ers
due
to
it
s
pot
ent
i
al
in
e
lectr
om
y
ogr
aph
y
(
E
MG
)
rec
ognitio
n
s
y
stem.
How
eve
r,
the
o
pti
m
al
m
othe
r
wave
let
sele
c
ti
o
n
remai
ns
a
cha
ll
eng
e
to
the
appl
i
ca
t
ion
of
W
T
in
EMG
signal
proc
essing.
Thi
s
pape
r
pr
e
sents
a
de
t
ai
l
stud
y
for
diffe
r
ent
m
othe
r
wav
el
e
t
func
ti
on
in
discre
te
wav
elet
tra
nsform
(DW
T)
and
co
nti
nuous
wave
l
et
tr
ansform
(CW
T).
Additionally
,
the
per
form
anc
e
of
diffe
ren
t
m
othe
r
wave
l
et
in
DW
T
and
CWT
at
dif
fer
e
n
t
dec
om
positi
on
l
eve
l
and
sc
al
e
ar
e
al
so
inve
s
ti
g
ated.
The
m
ea
n
ab
solute
val
u
e
(MA
V)
and
wave
le
ng
th
(W
L)
fea
tur
es
are
ext
r
ac
t
ed
from
each
CW
T
and
rec
onstruc
te
d
DW
T
wave
le
t
coeffic
i
ent
.
A
popul
ar
m
ac
hine
learn
ing
m
et
hod,
support
vec
tor
m
ac
hine
(SV
M)
is
emplo
y
ed
to
cl
assif
y
the
diff
e
ren
t
t
y
pes
of
hand
m
ovements.
The
result
s
show
ed
tha
t
t
h
e
m
ost
suita
ble
m
ot
her
wave
l
et
in
CW
T
are
Me
xic
an
h
at
and
Sy
m
le
t
6
at
sca
l
e
16
and
32,
resp
ec
t
ive
l
y
.
On
the
othe
r
h
and,
S
y
m
le
t
4
and
Dau
bec
hi
es
4
at
th
e
sec
ond
dec
om
po
siti
on
le
v
el
are
found
to
be
the
opti
m
al
wave
le
t
in
D
W
T.
From
the
ana
l
y
sis,
we
deduc
ed
tha
t
S
y
m
le
t
4
a
t
the
sec
ond
de
co
m
positi
on
le
ve
l
i
n
DW
T
is
the
m
ost
suita
bl
e
m
othe
r
wave
l
et
for
accura
t
e
c
la
s
sific
a
ti
on
of
EMG
signal
s
of
diff
ere
nt
h
and
m
ovements.
Ke
yw
or
d:
Con
ti
nu
ou
s
w
a
velet
t
ran
s
f
or
m
Discrete
w
avel
et
t
ran
s
form
Ele
ct
ro
m
yogr
a
ph
y
Mother
w
a
vele
t
Patt
ern
r
eco
gni
ti
on
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Jingwei
T
oo,
Faculty
of Elec
tric
al
Engineer
ing
,
Un
i
ver
sit
i
Te
knikal
Ma
la
ysi
a Me
la
ka,
Hang T
ua
h
Jay
a
, 76100
D
ur
ia
n
T
unggal
, Me
la
ka,
Ma
la
ysi
a
.
Em
a
il
: ja
m
esjam
es868@gm
ail.co
m
1.
INTROD
U
CTION
Ele
ct
ro
m
yogr
a
ph
y
(
EMG)
si
gn
al
c
on
ta
ins
r
ic
h
m
us
cl
e
info
rm
at
ion
that
can
be
us
e
d
in
cl
inica
l
and
reh
a
bili
ta
ti
on
app
li
cat
io
n.
T
he
po
te
ntial
of
E
MG
sig
nal
in
m
yoel
ect
ric
con
tr
ol
has
been
wides
pr
ea
d
si
nc
e
la
st
two
decad
e
s
[
1]
.
EMG
si
gn
a
l
reco
r
de
d
f
rom
a
con
tract
in
g
m
us
cl
e
no
t
on
ly
m
easur
es
the
tim
e
detect
ion
of
m
us
cl
e
act
ivatio
n
but
al
so
pro
vid
es
el
ect
rical
sign
s
of
m
us
cular
beh
a
vior
[
2]
.
Re
centl
y,
t
he
a
naly
sis
of
EMG
sign
al
us
in
g
a
powe
rful sig
na
l processi
ng techn
i
qu
e
h
a
s
be
com
e the att
ention o
f
the
r
esea
rch
e
rs.
In
bio
m
edical
sign
a
l
pr
ocessi
ng,
short
ti
m
e
Four
ie
r
T
ra
ns
f
or
m
(S
TFT)
,
w
avelet
trans
for
m
(W
T)
a
nd
e
m
pirical
decom
po
sit
ion
m
ode
(EMD)
a
re
frequ
e
ntly
us
ed
[
3]
-
[
5]
.
I
n
the
previ
ou
s r
esearc
h,
it
has
bee
n
f
ound
that
W
T
ou
t
pe
rfor
m
ed
oth
e
r
tim
e
-
fr
eq
uen
cy
m
et
ho
ds
in
di
scrim
inati
ng
E
MG
patte
r
ns
[
3],[6]
.
W
T
e
xhibit
s
good
ti
m
e
reso
l
utio
n
at
high
fr
e
quency
a
nd
go
od
f
re
quency
re
so
l
utio
n
at
l
ow
f
requ
ency
c
om
po
ne
nts
[
7]
.
In
ge
ne
ral,
W
T
can
be
cat
egorized
into
di
screte
and
co
ntinuo
us
f
or
m
.
In
co
ntinuo
us
wav
el
et
tran
sfor
m
(C
W
T
),
t
he
w
avelet
trans
f
orm
at
ion
cha
nge
s
co
ntin
uous
ly
.
O
n
one
side
,
discrete
wa
ve
le
t
transfor
m
(
D
WT)
deco
m
po
ses
th
e sig
nal into
m
ulti
reso
luti
on c
oeffici
ents
us
in
g hig
h pass a
nd lo
w pass
f
il
te
rs.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4221
-
4229
4222
Most
stu
dies
t
o
date
in
dicat
ed
the
perform
a
nces
of
C
WT
a
nd
D
WT
we
re
dep
e
ndin
g
on
t
he
sel
ect
io
n
of
a
m
oth
er
w
avelet
f
un
ct
i
on
[
3],[8]
-
[
10
]
.
I
n
th
e
past
stu
di
es,
Ka
koty
et
al
.
[
8]
in
vestig
at
ed
the
be
st
m
oth
er
wav
el
et
in
D
WT
a
nd
C
WT
at
dif
fer
e
nt
scal
e
an
d
de
com
po
sit
ion
l
evel.
T
he
a
uth
ors
reco
m
m
e
nd
e
d
t
he
Gau
s
sia
n
an
d
Sy
m
le
t
8
to
be
the
op
ti
m
al
m
o
ther
wa
velet
s
in
C
W
T
an
d
D
WT,
res
pecti
ve
ly
.
Ph
inyom
ark
et
al
.
[11]
sugg
est
e
d
that
the
us
e
of
D
WT
with
t
he
Da
ube
chies
7
an
d
8
to
e
nsure
higher
cl
a
ssific
at
ion
acc
ur
acy
.
Om
ari
et
al
.
[6]
stud
ie
d
four
m
ot
her
wa
velet
fu
nctio
ns
at
fo
ur
dif
fere
nt
deco
m
po
sit
io
n
le
vels.
The
au
thors
repor
te
d
Sym
l
et
4
offe
re
d
th
e
low
cl
assi
ficat
ion
er
r
or
rate.
Pr
e
vious
st
udie
s
sh
owe
d
tha
t
the
analy
sis
of
bes
t
m
oth
er
wa
velet
in
WT
is
c
riti
cal
ly
i
m
po
rtant,
le
a
ding
to
the
op
ti
m
u
m
cl
assifi
cat
ion
pe
rfor
m
ance.
Howeve
r
,
the sele
ct
ion o
f
m
oth
er w
a
vel
et
is rem
ai
ns
ch
al
le
ng
i
ng in
m
any areas.
The
best
m
oth
er
wa
velet
is
m
os
tl
y
su
bj
e
ct
ind
e
pende
nt
,
w
hich
m
eans
dif
fer
e
nt
m
oth
er
wa
velet
offer
s
dif
fer
e
nt
kind
of
pe
rform
ance
on
diff
e
ren
t
s
ubj
ect
.
I
n
ad
diti
on
,
previ
ou
s
w
orks
m
os
tl
y
fo
cus
on
four
t
o
ei
gh
t
m
oth
er
w
avelet
s
in
the
c
la
ssific
at
ion
of
EMG
sig
nals,
wh
ic
h
is
ins
uff
ic
ie
nt.
More
over,
the
perf
or
m
ance
of
m
oth
er
wa
ve
le
t
at
diff
e
rent
scal
e
an
d
dec
om
po
sit
ion
le
ve
l
pro
vid
e
si
gnific
ant
dif
fe
rence
in
cl
assifi
ca
ti
on
perform
ance.
I
t
is
ob
vious
th
at
the
analy
sis
of
the
m
oth
er
wav
el
et
in
CWT
an
d
D
WT
is
re
m
ai
n
insuffici
ent
and
uncl
ear
in
EMG
patte
r
n
recog
niti
on
.
T
her
e
fore,
this
s
tud
y
ai
m
s
to
e
valuate
t
he
be
st
m
oth
er
wa
ve
le
t
in
C
W
T
a
nd
D
WT
by
em
plo
yi
ng
a
la
rg
e
nu
m
ber
of
m
oth
e
r
wav
el
et
f
un
ct
io
ns
wit
h
diff
e
re
nt
sc
al
e
an
d
deco
m
po
sit
io
n l
evel, r
es
pecti
ve
ly
.
This
pap
e
r
pr
e
sents
a
detai
l
stud
y
of
t
he
sel
ect
ion
of
m
othe
r
wa
velet
in
D
WT
an
d
C
W
T.
14
m
oth
e
r
wav
el
et
s
of
D
WT
an
d
12
m
oth
e
r
wa
velet
s
of
C
WT
at
three
dif
fe
re
nt
de
com
po
sit
ion
l
evels
an
d
scal
es
ar
e
inv
est
igate
d,
r
especti
vely
.
T
wo
popula
r
fe
at
ur
es
m
ean
abso
l
ute
value
(MA
V)
an
d
w
avelen
gth
(
W
L)
are
extracte
d
f
r
om
each
wav
el
et
coeffic
ie
nt
for
perform
ance
evaluati
on.
T
he
m
ulti
c
la
ss
su
pport
vecto
r
m
a
chine
(S
VM
)
is
u
se
d
to
cl
assify
EMG
sign
al
si
nc
e
it
of
fer
s
bet
te
r
perform
ance
in
pr
e
vious
work
[8
]
,
[
12
]
.
Finall
y,
the b
e
st m
oth
er wavelet
of
D
WT
a
nd C
WT t
hat offe
r
the
best
cl
assifi
cat
ion
perform
ance ar
e
po
i
nted.
2.
MA
TE
RIA
L
AND RESE
A
RCH
METH
OD
2.1.
EMG
data col
le
ction
This
stu
dy
wa
s
perform
ed
on
te
n
healt
hy
su
bject
s
(
8
m
ales
and
2
fem
al
es)
with
m
ean
age
of
28.
6
(
=
9.7
)
ye
ars.
Each
sub
j
ect
pro
vid
e
d
inf
orm
ed
con
se
nt
to
pa
rtic
ipate
in
the
e
xp
e
rim
ent.
A
ddit
ion
a
ll
y,
al
l
su
bject
s
we
re
fr
ee
from
neur
ologica
l
an
d
m
us
cular
dis
order.
T
wo
w
ear
able
EMG
devi
ces
nam
ed
Shim
m
er
(S
him
m
er3
Co
ns
e
ns
ys
EMG
De
velo
pm
ent
Kits)
with
st
and
a
r
d
set
ti
ng
wer
e
use
d
i
n
data
c
ollec
ti
on
.
T
he
reso
l
ution
was
set
at
24
bits
with
a
gain
of
12.
The
EM
G
sign
al
was
gathe
red
f
ro
m
fo
ur
us
ef
ul
ha
nd
m
us
cl
es
nam
ely
extens
or
di
gitorum
(ch
1),
flex
or
ca
rp
i
rad
ia
li
s
(c
h2)
,
e
xtens
or
ca
rp
i
rad
ia
li
s
lo
ngus
(c
h3)
an
d
fl
ex
or
carp
i
ul
nar
is
(c
h4)
with
t
wo
r
efere
nce
el
ect
r
od
e
s
at
the
el
bow
.
The
si
gn
al
was
sam
pled
at
1024
Hz
an
d
ba
nd
-
pass
filt
ere
d
be
tween
20
a
nd
500
Hz.
T
he
s
kin
was
sh
a
ve
d
an
d
cl
eane
d
with
al
co
ho
l
pa
d
be
fore
the
e
le
ct
ro
de
placem
ent.
The
su
r
face
el
ect
r
od
e
s
wit
h
30
m
m
dia
m
e
te
r
wer
e
us
e
d
an
d
the
inter
-
el
ect
ro
de
distance
w
as
set
at
20 m
m
to r
e
duce the c
ro
s
sta
lk. The
b
i
po
la
r e
le
ct
ro
de
con
figurati
on
was
s
how
n
in
Fig
ure
1
.
Figure
1. Ele
ct
rodes
c
onfi
gur
at
ion
Subj
ect
w
as
s
eat
ed
com
fo
rta
bly
on
a
chai
r
with
the
ha
nd
in
ne
utral
posit
ion.
The
s
urface
EMG
sign
al
s
we
re
re
corde
d
as
the
su
bject
pe
rfor
m
ed
te
n
dif
fer
e
nt
han
d
m
ov
em
ents
inclu
ding
thu
m
b
flexio
n
(M1),
thu
m
b
exten
sion
(M
2)
,
w
rist
flexi
on
(M3
),
wr
ist
e
xtensi
on
(M
4)
,
m
aki
ng
a
fist
(M5
)
,
pi
nch
in
dex
to
th
e
thu
m
b
(M6)
,
pi
nch
m
idd
le
to
the
thu
m
b
(M7),
pin
c
h
rin
g
to
the
thu
m
b
(M8)
,
pin
c
h
li
ttle
to
the
thu
m
b
(M9
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A D
et
ail Stu
dy of W
avelet
Famil
ie
s for
EM
G
P
attern
Rec
ogniti
on (J
ingw
ei
Too)
4223
and
rest
(M1
0)
.
The
e
xp
e
rim
ents
co
ns
ist
ed
of
te
n
tria
ls.
W
it
hin
eac
h
tria
l,
the
subj
ect
was
aske
d
to
pe
rfor
m
te
n
dif
fer
e
nt
ha
nd
m
ov
em
ents
for
5
s
eac
h,
fo
ll
owe
d
by
a
resti
ng
sta
te
of
4
s.
Mo
reove
r
,
a
resti
ng
pe
riod
of
1
m
in
was
intr
oduce
d
at
the
e
nd
of
t
rial
to
a
vo
i
d
m
ental
and
m
us
cl
e
fati
gu
e
.
T
he
resti
ng
sta
te
was
r
e
m
ov
ed
befor
e
d
at
a
seg
m
entat
ion
.
A
rece
nt
re
port
of
real
ti
m
e
E
MG
ap
plica
ti
on
in
dicat
ed
tha
t
the
op
ti
m
al
w
indow
le
ng
t
h
was
ra
ngin
g
from
15
0
to
250
m
s
to
balance
the
co
ntr
oller
delay
an
d
cl
assifi
cat
ion
erro
r
rate
[
13
]
.
A
ddit
ion
a
ll
y,
an
ov
e
rlap
pe
d
wi
ndowin
g
t
ech
ni
qu
e
was
intr
oduce
d
to
pro
duce
bette
r
cl
as
sific
at
ion
accu
racy
in
EMG
patte
rn
recog
niti
on
[14]
. I
n
this w
ork, t
he
EMG d
at
a
wer
e
div
ide
d
i
nto
250
m
s w
i
ndow (2
56 sa
m
ples)
with 5
0%
(
128
sam
ples)
ov
erl
app
e
d.
I
n
total
,
a
data
m
a
trix
of
39
segm
ents
256
sam
ples
4
cha
nnel
s
we
r
e
ob
ta
ine
d
from
each m
ov
em
ent f
r
om
each
subj
ect
.
Figure
2
s
how
s
the
flo
w
diagr
am
of
t
he
pro
posed
rec
og
niti
on
syst
e
m
.
In
the
first
st
age,
t
he
raw
EMG
sig
nals
are
colle
ct
ed
and
se
gm
ented.
Ne
xt,
MA
V
an
d
W
L
fe
at
ur
es
are
e
xtr
act
ed
from
C
WT
an
d
reconstr
ucted
D
WT
co
ef
fici
ents
at
di
ff
e
ren
t
scal
e
an
d
dec
om
po
sit
ion
le
ve
l
us
in
g
diff
e
r
ent
m
oth
er
w
a
velet
,
re
sp
ect
ively
.
I
n
the
final
st
age,
t
he
S
VM
is
us
e
d
to
re
cognize
the
E
MG
sig
nals
of
te
n
dif
fer
e
nt
ha
nd
m
ov
e
m
ents.
Figure
2. The
fl
ow
diag
ram
o
f
the
pro
po
se
d r
ecognit
ion sy
stem
2.2.
Wavel
et Tr
ansfo
r
m
Wav
el
et
tra
nsf
or
m
(
W
T
)
is
a
powe
rful
m
ath
em
atical
too
l
that
is
su
cces
sfu
l
i
n
the
a
na
ly
sis
of
bi
o
-
sign
al
incl
ud
i
ng
EMG
sig
nal.
W
T
offer
s
hi
gh
fr
e
quency
r
esolutio
n
f
or
low
fr
e
quency
com
po
ne
nt
an
d
go
od
tim
e
reso
luti
on
f
or
t
he
high
f
reque
ncy
com
pone
nt
[
13]
.
G
ener
al
ly
,
WT
can
be
cat
e
gor
iz
ed
int
o
c
on
ti
nuous
and
disc
rete
f
orm
s.
Con
ti
nuous
wa
ve
le
t
tra
ns
f
or
m
(C
W
T
)
dec
om
po
ses
t
he
si
gn
al
base
d
on
the
dilat
ion
s
a
nd
translat
ions
of
a
sing
le
m
oth
er
wav
el
et
f
unct
ion
.
C
W
T
is
m
or
e
con
sist
e
nt
and
ef
fici
en
t
becau
se
it
prov
i
des
local
iz
at
ion
ti
m
e
-
fr
e
qu
e
ncy
inf
or
m
at
ion
wi
thout
do
wn
-
sa
m
pl
ing
[
11]
.
Additi
on
al
ly
,
C
W
T
is
c
on
ti
nuous
in
te
rm
o
f
sh
ifti
ng a
nd it
g
iv
es
use
fu
l t
im
e
-
fr
eq
uen
cy
i
nfor
m
at
ion
[
15
]
. C
WT
can be
de
fine
d as:
s,
(
s,
)
(
)
(
)
xb
CW
T
b
x
t
t
d
t
(1)
wh
e
re
x(
t)
is
the
i
nput
si
gn
al
an
d
ψ
s,b
(
t
)
is
the
transfor
m
at
ion
of
the
m
oth
er
wa
velet
f
unct
ion.
The
tra
nsfo
rm
a
ti
on
ca
n be e
xp
resse
d
as:
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4221
-
4229
4224
s,
1
()
b
tb
t
s
s
(2)
wh
e
re
s
is
the
scal
ing
pa
ram
e
te
r,
b
is
re
ferre
d
to
the
tra
ns
la
ti
on
pa
ram
et
er
and
(
t)
is
ca
lled
m
oth
er
wa
ve
le
t.
The
va
riables
s
and
b
prov
i
de
the
tim
e
scaling
a
nd
s
hiftin
g
operati
on,
re
sp
ect
ively
[16]
.
By
us
in
g
e
q
ua
ti
on
1
and 2, C
WT
ca
n be c
om
pu
te
d as:
1
(
s,
)
(
)
tb
CW
T
b
x
t
dt
s
s
(3)
Figure
3
dem
on
strat
es
the
sc
al
ogram
of
CWT
at
scal
e
32
us
i
ng
Me
xican
hat
wa
velet
.
The
ye
ll
ow
areas
represe
nt h
ig
he
r
am
plit
ud
e at ea
c
h
scal
e. In t
urn,
da
rk b
lue
areas
r
e
fe
r
to
lo
w
am
plitu
de
.
Figure
3. Scal
ogram
o
f
c
onti
nuous
wa
velet
tr
ansfo
rm
at scale 32 usin
g
Me
xican hat
Discrete
wa
vel
et
transfor
m
(DWT)
is
der
iv
ed
f
ro
m
CWT
[17]
.
D
WT
is
m
or
e
widely
use
d
becau
se
it
offer
s
l
ow
c
om
pu
ta
ti
on
cost
[11]
.
I
n
D
W
T,
the
sig
nal
is
deco
m
po
se
d
int
o
the
ap
pr
ox
im
at
ion
and
detai
l
coeffic
ie
nt
w
hich
involve
s
th
e
change
of
sa
m
pl
ing
rate
[
18]
.
The
dec
om
po
sit
ion
of
D
WT
com
pr
ise
s
of
tw
o
dig
it
al
filt
ers,
wh
ic
h
are
high
-
pass
an
d
lo
w
-
pass
filt
ers.
T
he
low
-
pas
s
an
d
high
-
pass
filt
er
do
wn
-
sam
ple
the
input
sig
na
l
a
nd
pro
vid
e
t
he
appr
ox
im
at
ion
,
A
a
nd
detai
l,
D
,
res
pecti
vely
[
11
]
,
[19]
.
F
or
each
dec
om
posit
ion
le
vel, the
f
il
te
r
s down
-
sam
ple the si
gn
al
by t
he fact
or of
2.
The first l
ev
el
of d
ec
om
po
sit
ion i
s
def
ine
d
a
s:
D
[
]
[
k
]
[
2
]
n
n
x
h
n
k
(4)
A
[
n
]
[
k
]
[
2
]
n
x
g
n
k
(5)
wh
e
re
x[
k
]
is
the
input
sig
nal,
D[
n
]
is
r
efer
red
t
o
the
detai
l,
D
1
an
d
A[
n
]
is
t
he
approxim
at
io
n,
A
1
.
The
dec
om
po
s
it
ion
process
is
rep
eat
ed
un
ti
l
the
desire
d
fin
al
le
vel
is
achi
eved.
I
n
the
pr
evio
us
resea
rc
h,
eac
h
coeffic
ie
nt
sub
set
was
reconst
ru
ct
e
d
to
ob
ta
i
n
m
or
e
reli
able
EMG
sign
al
pa
rt,
resu
lt
in
g
in
bette
r
cl
assifi
cat
ion
accuracy
[3
]
,
[
13
]
.
The
refor
e
,
the
in
ve
rse
wav
el
et
tra
nsf
or
m
is
us
ed
t
o
reconstr
uct
ea
ch
wav
el
et
c
oe
ff
ic
ie
nt
into
m
or
e
eff
e
ct
ive
subset,
nam
el
y,
est
i
m
ated
ap
pro
xim
ati
on,
rA
a
nd
e
sti
m
at
ed
detai
l,
rD
.
F
or
exam
ple,
the
est
i
m
at
ed
su
bse
t
rD
3
is
obta
ined
by
perfor
m
ing
the
i
nve
rse
wav
el
et
tr
ansfo
rm
on
th
ird
-
le
vel
detai
l,
D
3
.
The
wa
velet
reco
nst
ru
ct
io
n
of
est
i
m
a
te
d
detai
l
(
rD
1
-
rD
6
)
an
d
est
i
m
at
ed
app
r
oxim
a
ti
on
(
rA
1
-
rA
6
)
we
re
s
how
n
in Figu
re
4.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A D
et
ail Stu
dy of W
avelet
Famil
ie
s for
EM
G
P
attern
Rec
ogniti
on (J
ingw
ei
Too)
4225
Figure
4.
W
a
ve
le
t reco
ns
tr
uction o
f D
W
T
at
sixth
dec
om
po
sit
ion
level
us
i
ng Sym
let 4
2.3.
Mother
W
av
e
le
t
Sele
cti
on
and Ev
alu
at
i
on
Re
cent
stud
ie
s
ind
ic
at
ed
WT
has
been
rec
ognized
a
s
on
e
of
t
he
best
t
i
m
e
-
fr
eq
ue
ncy
m
et
ho
d
i
n
bio
m
edical
sign
al
proces
sin
g
[3
]
,[1
8],[20
]
.
Howe
ver,
the
perform
ance
of
W
T
is
m
os
tl
y
dep
e
nd
i
ng
on
the
m
oth
er
wav
el
e
t
fu
nctio
n.
T
he
sel
ect
ion
of
m
oth
e
r
wav
el
et
is
rem
a
ined
cha
ll
eng
in
g
in
m
a
ny
areas.
The
r
efore,
this
w
ork
ai
m
s
to
eval
uate
th
e
best
m
oth
er
wav
el
et
in
D
WT
a
nd
C
WT
for
E
MG
sig
na
l
processin
g.
In
t
his
stud
y,
14
m
oth
er
wav
el
et
s
in
D
WT
an
d
12
m
oth
er
wa
ve
le
ts
in
CW
T
are
inv
e
sti
gated.
Ta
ble
1
is
a
loo
ku
p
ta
ble
of
the
m
oth
e
r
wa
velet
us
e
d
in
C
W
T
and
D
WT.
It
is
worth
noti
ng
diff
e
ren
t
scal
e
and
dec
om
po
sit
ion
le
vel
in
C
WT
and
D
WT
pr
ov
ide
di
ff
e
ren
t
prop
e
rty
.
F
or
thi
s
reas
on,
t
he
pe
rfor
m
ance
of
the
m
oth
er
wa
velet
at
the scal
e
8,
16,
32 a
nd d
ec
ompo
sit
io
n
le
vel
of 2,
4
a
nd 6 ar
e exam
ined.
Table
1.
M
oth
e
r
w
avelet
of
C
WT
a
nd
D
W
T
us
e
d
in
this st
udy
CWT
DW
T
1
Haar
Haar
2
Dau
b
echi
es 2
Dau
b
echi
es 2
3
Dau
b
echi
es 4
Dau
b
echi
es 4
4
Dau
b
echi
es 6
Dau
b
echi
es 6
5
Sy
m
let 2
Dau
b
echi
es 8
6
Sy
m
let 4
Dau
b
echi
es 1
0
7
Sy
m
let 6
Sy
m
let 2
8
Morlet
Sy
m
let 4
9
Mayer
Sy
m
let 6
10
Mexican
h
at
Sy
m
let 8
11
Gau
ss
ian
2
Co
if
let 2
12
Gau
ss
ian
4
Co
if
let 3
13
-
Co
if
let 4
14
-
Co
if
let 5
2.4.
Feature E
xt
r
ac
tion usin
g
W
av
el
e
t
Tr
ansf
or
m
Feat
ur
e e
xtract
ion
is a
n
esse
nt
ia
l st
ep
to r
e
duce the d
im
ension
al
it
y an
d
ext
ract t
he usef
ul
inf
or
m
at
ion
from
the
sign
a
l.
In
this
w
ork
,
wa
velen
gth
(
WL)
a
nd
m
ea
n
abs
ol
ute
val
ue
(M
AV)
are
extracte
d
fro
m
each
wav
el
et
c
oeffic
ie
nt.
MA
V
a
nd
W
L
can
b
e
express
ed
as
[6]
:
1
1
L
n
n
M
A
V
x
L
(6)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4221
-
4229
4226
1
1
1
L
nn
n
W
L
x
x
(7)
wh
e
re
X
n
is t
he
input si
gn
al
a
nd
L
is t
he
len
gth
of sig
nal.
2.5.
Support
V
ec
t
or Mac
hine
Suppor
t
vect
or
m
achine
(SV
M)
ha
s
been
pro
ve
d
to
be
an
outst
an
ding
s
up
e
r
vised
m
achine
le
a
rn
i
ng
m
et
ho
d
i
n
EM
G
patte
rn
r
ec
ogniti
on
[14]
. Mo
re
ov
e
r,
SVM
has
s
how
n
i
ts
superi
or
it
y, especial
ly
for
non
-
li
near
and
high
dim
e
ns
io
nal
patte
r
n
recogn
it
io
n
[
21]
.
SV
M
m
aps
the
pr
e
dictors
on
to
a
high
di
m
ension
al
spa
ce
by
us
in
g
th
e
co
nc
ept
of
hype
rp
l
ane
par
ti
ti
on
f
or
t
he
data
[22]
.
So
m
e
dr
a
wbacks
of
S
VM
are
the
com
plexity
of
the
sel
ect
ion
of
kernel
f
unct
ion
an
d
t
he
l
onger
com
pu
ta
ti
on
ti
m
e
[14]
.
A
previ
ou
s
stu
dy
repor
te
d
t
hat
ra
dial
basis
functi
on
(RBF)
was
the
best
kernel
functi
on
because
it
gav
e
a
hig
he
r
cl
assifi
cat
ion
perf
or
m
anc
e
[6]
.
I
n
this re
gard,
S
V
M wit
h
RB
F
kernel
functi
on i
s appli
ed
a
nd it
can b
e
de
fine
d as:
2
2
||
||
(
,
)
e
x
p
2
i
i
xx
K
x
x
(8)
wh
e
re
x
-
x
i
is t
he
Eu
cl
idea
n dis
ta
nce b
et
wee
n feat
ur
e
v
e
ct
ors
and
is t
he
k
e
rn
el
par
am
et
er.
3.
RESU
LT
S
A
ND AN
ALYSIS
In
this
w
ork,
10
-
f
old
cr
os
s
validat
io
n
is
a
pp
li
ed
in
the
c
la
ssific
at
ion
of
EMG
sign
al
s.
The
data
is
separ
at
e
d
int
o
10
eq
ual
pa
rts.
Eve
ry
pa
rt
ta
ke
s
tur
n
to
te
st
and
the
rem
ai
n
ing
par
ts
a
re
use
d
in
t
rainin
g
ph
a
se.
In
the
first
pa
rt
of
the
experim
ents,
14
m
oth
er
wav
el
et
f
un
ct
ion
s
in
D
WT
a
t
the
three
different
dec
om
posit
ion
le
vel
are
e
valu
at
ed.
Ta
bl
e
2
ou
tl
ines
t
he
m
ean
cl
assifi
cat
ion
accu
racy
of
14
m
oth
er
wa
velet
s
of
D
WT
at
a
deco
m
po
sit
io
n
le
vel
of
2,
4
and
6
acr
os
s
t
en
di
ff
e
ren
t
s
ubj
ect
s
.
From
t
he
res
ults,
the
m
ean
cl
assifi
cat
ion
accuracy
is
found
to
be
a
bove
97%
f
or
al
l
14
m
oth
er
wa
velet
functi
on
s
in
bo
th
WL
and
MA
V
feat
ure
set
s.
Additi
on
al
ly
,
MAV
has
sho
wn
to
be
a
n
ef
fecti
ve
an
d
reli
able
featu
re
be
cause
it
offer
s
bette
r
pe
rfor
m
ance
in
discrim
inati
ng
EMG
patte
r
ns.
By
em
plo
yin
g
MA
V
featur
e
,
it
is
ob
vi
ou
s
that
t
he
hi
gh
est
cl
assifi
cat
ion
accuracy
is o
bt
ai
ned
b
y
Sym
l
et
4
(
98.74%
),
f
ollo
we
d
by Daub
ec
hies 4
(
98.
72%)
at
the
s
econd d
ecom
posit
ion
le
ve
l
.
On
the
ot
her
ha
nd,
Coif
le
t
3
ou
tpe
rform
s
oth
er
m
oth
er
wav
el
et
s
wit
h
a
m
ean
cl
assifi
cat
ion
accu
r
acy
of
98.49%
at
the fourth d
ec
om
po
sit
ion
le
vel
w
he
n
WL
is
us
e
d.
Fr
om
the
anal
ysi
s,
Sy
m
le
t
4
and
D
a
ub
ec
hie
s
4
at
the sec
ond dec
om
po
sit
ion
le
ve
l are fo
und
t
o be the
m
os
t suita
ble m
oth
er wavelet
in D
WT
.
Table
2.
Cl
assi
ficat
ion
Acc
uracy
(m
ean ±
S
TD) of
14
Mot
her
W
a
velet
s
of DWT
at
T
hr
e
e D
if
fer
e
nt
Deco
m
po
s
it
io
n Level
Across
Ten Su
bject
s
Moth
er
wav
elet
Clas
sif
icatio
n
perfor
m
an
ce
(%)
Moth
er
wav
elet
Clas
sif
icatio
n
perfor
m
an
ce
(%)
WL
MAV
WL
MAV
Haar
Level 2
9
7
.90
±
1
.02
9
8
.43
±
0
.88
Sy
m
4
Level 2
9
8
.09
±
0
.92
9
8
.74
±
0
.66
Level 4
9
8
.00
±
0
.90
9
8
.28
±
0
.80
Level
4
9
8
.36
±
0
.78
9
8
.53
±
0
.67
Level 6
9
7
.28
±
0
.94
9
7
.64
±
0
.85
Level 6
9
7
.50
±
0
.79
9
7
.67
±
0
.81
Db
2
Level 2
9
7
.97
±
1
.01
9
8
.63
±
0
.68
Sy
m
6
Level 2
9
8
.18
±
0
.87
9
8
.67
±
0
.76
Level 4
9
8
.31
±
0
.72
9
8
.44
±
0
.70
Level 4
9
8
.39
±
0
.67
9
8
.55
±
0
.68
Level 6
9
7
.32
±
0
.94
9
7
.45
±
0
.78
Level 6
9
7
.58
±
0
.84
9
7
.65
±
0
.87
Db
4
Level 2
9
8
.08
±
0
.88
9
8
.72
±
0
.67
Sy
m
8
Level 2
9
8
.19
±
0
.87
9
8
.70
±
0
.71
Level 4
9
8
.36
±
0
.78
9
8
.56
±
0
.68
Level 4
9
8
.45
±
0
.69
9
8
.57
±
0
.72
Level 6
9
7
.36
±
0
.91
9
7
.55
±
0
.82
Level 6
9
7
.67
±
0
.87
9
7
.74
±
0
.85
Db
6
Level 2
9
8
.23
±
0
.90
9
8
.65
±
0
.73
Co
if
2
Level 2
9
8
.10
±
0
.90
9
8
.69
±
0
.70
Level 4
9
8
.48
±
0
.63
9
8
.53
±
0
.69
Level 4
9
8
.34
±
0
.79
9
8
.52
±
0
.66
Level 6
9
7
.48
±
0
.88
9
7
.52
±
0
.85
Level 6
9
7
.59
±
0
.91
9
7
.70
±
0
.77
Db
8
Level 2
9
8
.20
±
0
.90
9
8
.69
±
0
.71
Co
if
3
Level 2
9
8
.18
±
0
.90
9
8
.69
±
0
.71
Level 4
9
8
.44
±
0
.67
9
8
.59
±
0
.67
Level 4
9
8
.49
±
0
.70
9
8
.62
±
0
.62
Level 6
9
7
.57
±
0
.74
9
7
.60
±
0
.82
Level 6
9
7
.56
±
0
.85
9
7
.71
±
0
.73
Db
10
Level 2
9
8
.17
±
0
.94
9
8
.70
±
0
.68
Co
if
4
Level 2
9
8
.22
±
0
.93
9
8
.70
±
0
.70
Level 4
9
8
.48
±
0
.66
9
8
.62
±
0
.60
Level 4
9
8
.42
±
0
.74
9
8
.56
±
0
.68
Level 6
9
7
.43
±
0
.91
9
7
.49
±
0
.88
Level 6
9
7
.61
±
0
.83
9
7
.71
±
0
.72
Sy
m
2
Level 2
9
7
.97
±
1
.01
9
8
.63
±
0
.68
Co
if
5
Level 2
9
8
.23
±
0
.88
9
8
.70
±
0
.71
Level 4
9
8
.31
±
0
.72
9
8
.44
±
0
.70
Level 4
9
8
.45
±
0
.72
9
8
.56
±
0
.59
Level 6
9
7
.32
±
0
.94
9
7
.45
±
0
.78
Level 6
9
7
.50
±
0
.86
9
7
.59
±
0
.85
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A D
et
ail Stu
dy of W
avelet
Famil
ie
s for
EM
G
P
attern
Rec
ogniti
on (J
ingw
ei
Too)
4227
In
the
sec
ond
pa
rt
of
the
exp
e
rim
ents,
12
m
oth
er
w
avelet
s
of
C
WT
a
re
st
ud
i
ed.
Table
3
dem
on
strat
es
the
m
ean
cl
assifi
cat
ion
acc
ur
a
cy
of
12
m
oth
er
wa
velet
s
of
C
W
T
at
scal
e
8,
16
an
d
32
f
or
te
n
diff
e
re
nt
sub
j
e
ct
s.
At
scal
e
8,
Gaussi
an
2
a
nd
4
ex
hib
it
the
high
est
cl
assif
ic
at
ion
accu
rac
y
of
98.42%
usi
ng
WL
a
nd
M
AV
featu
re
set
s,
r
especti
vely
.
H
ow
e
ve
r,
t
heir
perform
ance
di
d
no
t
sho
w
m
uch
im
pr
ov
em
ent
at
a
higher
scal
e.
At
scal
e
16,
it
has
bee
n
f
ou
nd
t
hat
the
Sym
le
t
6
achieve
s
the
best
cl
assifi
cat
ion
acc
uracy
of
98.
56%
,
f
ollo
wed
by
Sym
let
4,
98.
53%
w
hen
MA
V
is
use
d.
F
or
i
ns
ta
nc
e,
the
Me
xica
n
hat
ha
s
s
hown
it
s
su
pe
rio
rity
at scal
e 3
2 wit
h
th
e b
est
m
ean cl
assifi
cat
ion
acc
ur
acy
of
98.64
% in
W
L
featu
re s
et
. Unfo
rtu
natel
y,
MAV
s
hows
t
he
dec
rem
ent
i
n
cl
assifi
cat
ion
perform
ance
at
scal
e
32
.
T
his
sh
ows
that
M
AV
featu
re
set
is
no
t
su
it
able
f
or
hi
gh
scal
e
wavel
et
functi
on
i
n
C
W
T.
As
a
re
su
lt
,
the
m
os
t
su
it
able
m
oth
er
wa
velet
in
CWT
are
Me
xican hat
at
scale
32 a
nd S
ym
let 6
at scal
e 16.
Table
3.
Cl
assi
ficat
ion
Acc
uracy
(m
ean ±
S
TD) of
12
Mot
her
W
a
velet
s
of C
WT
at
T
hr
e
e D
if
f
ere
nt Sca
le
Across
Ten
S
ubj
ect
s
Moth
er
wav
elet
Clas
sif
icatio
n
perfor
m
an
ce
(%)
Moth
er
wav
elet
Clas
sif
icatio
n
perfor
m
an
ce
(%)
WL
MAV
WL
MAV
Haar
Scale 8
9
7
.70
±
0
.96
9
8
.00
±
1
.08
Sy
m
6
Scale 8
9
8
.05
±
0
.86
9
8
.17
±
0
.97
Scale 16
9
8
.38
±
0
.92
9
8
.31
±
0
.96
Scale 16
9
8
.48
±
0
.81
9
8
.56
±
0
.74
Scale 32
9
8
.51
±
0
.76
9
8
.19
±
0
.79
Scale 32
9
8
.49
±
0
.72
9
8
.35
±
0
.73
Db
2
Scale 8
9
7
.88
±
1
.01
9
8
.06
±
1
.13
Morl
Scale 8
9
8
.00
±
0
.83
9
8
.07
±
0
.83
Scale 16
9
8
.44
±
0
.88
9
8
.42
±
0
.86
Scale 16
9
8
.40
±
0
.86
9
8
.40
±
0
.82
Scale 32
9
8
.50
±
0
.72
9
8
.29
±
0
.73
Scale 32
9
8
.34
±
0
.74
9
8
.26
±
0
.78
Db
4
Scale 8
9
7
.99
±
0
.92
9
8
.13
±
1
.03
Meyr
Scale 8
9
8
.06
±
0
.95
9
8
.13
±
0
.95
Scale 16
9
8
.46
±
0
.90
9
8
.47
±
0
.78
Scale 16
9
8
.45
±
0
.84
9
8
.49
±
0
.75
Scale 32
9
8
.45
±
0
.76
9
8
.27
±
0
.76
Scale 32
9
8
.36
±
0
.79
9
8
.29
±
0
.81
Db
6
Scale 8
9
7
.94
±
1
.01
9
8
.08
±
1
.08
Mexh
Scale 8
9
8
.36
±
0
.82
9
8
.15
±
0
.79
Scale 16
9
8
.41
±
0
.93
9
8
.46
±
0
.78
Scale 16
9
8
.08
±
0
.76
9
7
.49
± 0.8
1
Scale 32
9
8
.36
±
0
.75
9
8
.27
±
0
.76
Scale 32
9
8
.64
±
0
.66
9
6
.26
±
1
.00
Sy
m
2
Scale 8
9
7
.88
±
1
.01
9
8
.06
±
1
.13
Gau
s 2
Scale 8
9
8
.42
±
0
.83
9
8
.35
±
0
.84
Scale 16
9
8
.44
±
0
.88
9
8
.42
±
0
.86
Scale 16
9
8
.28
±
0
.76
9
8
.00
±
0
.77
Scale 32
9
8
.50
± 0.7
2
9
8
.29
±
0
.73
Scale 32
9
8
.50
±
0
.67
9
7
.01
±
0
.87
Sy
m
4
Scale 8
9
8
.03
±
0
.87
9
8
.18
±
0
.99
Gau
s 4
Scale 8
9
8
.39
±
0
.93
9
8
.42
±
0
.83
Scale 16
9
8
.48
±
0
.83
9
8
.53
±
0
.74
Scale 16
9
8
.48
±
0
.70
9
8
.42
±
0
.71
Scale 32
9
8
.52
±
0
.69
9
8
.34
±
0
.74
Scale 32
9
8
.31
±
0
.69
9
7
.80
±
0
.77
In
t
he
final
pa
rt
of
the
ex
pe
rim
ents,
the
paire
d
tw
o
-
ta
il
t
-
te
st
is
us
e
d
to
m
easur
e
t
he
sta
ti
sti
cal
diff
e
re
nce
bet
ween
t
he
cl
assi
ficat
ion
perfor
m
ances
of
WL
and
MA
V
feat
ur
es
w
hen
dif
f
eren
t
m
oth
er
w
avelet
functi
on
is
us
e
d.
Ta
ble
4
an
d
5
ou
tl
ine
the
resu
lt
of
t
-
te
st
of
t
he
cl
assifi
cat
ion
perform
ance
ob
ta
ine
d
from
D
WT
a
nd
C
WT
acr
os
s
te
n
s
ubj
ect
s.
I
n
t
-
te
st,
the
null
hy
po
t
hesis
is
reje
ct
ed
if
t
he
p
-
va
lue
is
le
ss
t
ha
n
0.05.
This s
h
ow
s
tha
t t
her
e is a
stat
ist
ic
al
d
iffe
ren
c
e b
et
wee
n WL
and MA
V feat
ur
e
sets.
Fr
om
Table
4,
the
res
ults
of
the
WL
a
nd
M
AV
are
sta
ti
stical
dif
fer
e
nce
for
al
l
wav
el
et
f
un
ct
io
ns
at
the
sec
ond
dec
om
po
sit
ion
le
ve
l.
At
four
t
h
de
com
po
sit
ion
l
evel,
the
p
-
value
il
lus
trat
es
t
hat
the
Da
ub
e
c
hies
6
and
Coi
flet
5
sh
ow
no
si
gnific
ant
diff
e
re
nc
e
w
he
n
WL
ver
s
us
MA
V.
At
sixt
h
decom
po
sit
ion
le
ve
l,
onl
y
Haar,
Da
ub
ec
hies
4
a
nd
Sy
m
le
t
4
exh
ibit
the
sign
ific
a
nt
diff
ere
nce
.
F
ro
m
Table
5,
Haar,
Sym
le
t
4
an
d
Me
xican
hat
s
how
sig
nifica
nt
diff
ere
nce
in
scal
e
8.
A
dd
it
ion
al
ly
,
at
scal
e
16
,
on
ly
Me
xi
can
hat,
Ga
us
s
ia
n
2
and
Ga
us
sia
n
4
obta
in
p
-
val
ue
lo
wer
t
han
0.05.
Mo
reove
r,
ot
her
t
han
Daubec
hies
6
and
Sym
le
t
6
exh
i
bit
sign
ific
a
nt
differences
b
et
wee
n
the
classi
fica
ti
on
perform
ance of
WL
a
nd
MAV
at
scal
e
32.
Table
4.
Res
ult o
f
t
-
te
st
of
t
he
Cl
assifi
cat
ion
Perfo
rm
an
ce
be
tween M
A
V
a
nd
WL usin
g D
WT
Moth
er
wav
elet
p
–
v
alu
e
Level 2
Level 4
Level 6
Haar
0
.00
0
7
0
.00
0
7
3E
–
05
Db
2
0
.00
1
2
0
.01
9
5
0
.05
2
1
Db
4
0
.00
0
6
0
.00
8
7
0
.00
3
1
Db
6
0
.00
7
0
0
.37
5
4
0
.23
4
0
Db
8
0
.00
3
7
0
.01
8
5
0
.60
8
5
Db
10
0
.00
2
0
0
.00
3
6
0
.31
6
3
Sy
m
2
0
.00
1
2
0
.01
9
5
0
.05
2
1
Sy
m
4
0
.00
0
9
0
.00
5
7
0
.01
3
8
Sy
m
6
0
.00
0
8
0
.00
4
6
0
.03
8
0
Sy
m
8
0
.00
0
7
0
.02
8
9
0
.10
8
1
Co
if
2
0
.00
1
2
0
.01
7
8
0
.08
5
4
Co
if
3
0
.00
3
1
0
.01
0
9
0
.06
2
5
Co
if
4
0
.00
3
1
0
.01
5
7
0
.05
0
4
Co
if
5
0
.00
1
0
0
.08
6
0
0
.08
0
7
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4221
-
4229
4228
Table
5.
Res
ult o
f
t
-
te
st
of
t
he
Cl
assifi
cat
ion
Perfo
rm
ance
be
tween M
A
V
a
nd
WL usin
g
C
WT.
Moth
er
wav
elet
p
–
v
alu
e
Scale 8
Scale 16
Scale 32
Haar
0
.03
7
7
0
.26
7
6
0
.00
0
3
Db
2
0
.05
3
9
0
.81
4
1
0
.00
0
3
Db
4
0
.11
0
4
0
.84
7
8
0
.01
0
4
Db
6
0
.05
2
5
0
.38
5
5
0
.06
7
0
Sy
m
2
0
.05
3
9
0
.81
4
1
0
.00
0
3
Sy
m
4
0
.04
0
9
0
.52
5
6
0
.00
3
7
Sy
m
6
0
.06
2
5
0
.21
7
2
0
.05
2
6
Morl
0
.06
3
5
0
.91
6
2
0
.00
5
0
Meyr
0
.12
6
6
0
.38
6
5
0
.02
0
7
Mexh
3E
–
05
7E
–
07
2E
–
06
Gau
s 2
0
.28
6
4
7E
–
05
7E
–
07
Gau
s 4
0
.56
8
3
0
.03
3
4
5E
–
05
4.
CONCL
US
I
O
N
In
this
st
ud
y,
the
us
ef
uln
e
ss
of
the
m
oth
er
wa
velet
functi
on
in
D
WT
an
d
C
WT
has
be
e
n
inv
est
igate
d.
T
wo
popula
r
fea
tures,
WL
a
nd
MAV
a
re
ext
ra
ct
ed
from
the
wav
el
et
co
ef
fici
ents
as
the
in
pu
t
to
the
SV
M
.
I
n
C
W
T,
t
he
Me
xican
hat
at
sc
al
e
32
a
nd
Sy
m
le
t
6
at
sca
le
16
are
s
ugge
s
te
d
to
be
the
optim
a
l
m
oth
er
wav
el
e
t
sel
ect
ion
f
or
the
cl
assifi
cat
ion
of
EMG
sign
al
s
.
O
n
the
oth
e
r
ha
nd,
t
he
reconstr
ucte
d
D
WT
coeffic
ie
nt
with
Da
ub
e
chies
4
an
d
Sym
let
4
at
second
de
com
po
sit
ion
le
vel
are
rec
omm
end
ed
t
o
be
us
e
d
in
EMG
patte
r
n
r
ecognit
ion.
T
he
ex
per
im
ental
res
ults
ind
ic
at
ed
D
W
T
not
only
offer
e
d
l
ow
c
om
pu
ta
ti
on
cost,
bu
t
al
so
yi
el
de
d
a
high
cl
assif
ic
at
ion
accurac
y.
As
com
par
ed
to
C
W
T
,
D
WT
is
m
or
e
ap
proar
ia
te
to
be
us
e
d
in
reh
a
bili
ta
ti
on
a
nd cli
nical
app
li
cat
ion
.
ACKN
OWLE
DGE
MENTS
The
aut
hors
w
ou
l
d
li
ke
to
thank
the
Un
i
versi
ti
Tekn
ikal
Ma
la
ysi
a
Mela
ka
(UTeM
),
S
kim
Za
m
al
ah
UTeM
a
nd
Mi
nister
of
Higher
Ed
uca
ti
on
Ma
la
ysi
a
(MO
HE)
f
or
f
undi
ng
re
search
un
der
gr
a
nt
PJP/1/2
01
7/FKEKK/
H19/S
01
526.
REFERE
NCE
S
[1]
A.
Phin
y
om
ark
,
et
al.
,
“
Feat
ur
e
red
uction
and
sele
ction
for
EMG
signal
cl
a
ss
ifi
ca
ti
on
,
”
E
x
pert
Syst
em
wit
h
Appl
ic
a
ti
on
,
vol
/
issue:
39
(
8
)
,
pp.
7420
–
7431,
201
2.
[2]
G.
Vannoz
z
i,
e
t
al.
,
“
Autom
at
i
c
detec
t
ion
of
surfac
e
EMG
ac
t
iva
ti
on
ti
m
ing
u
sing
a
wave
l
et
tra
nsform
base
d
m
et
hod,
”
Journa
l
of
Elec
tromyography
and
Kin
esiol
ogy
,
vol
/i
ss
ue
:
20
(
4
)
,
pp.
767
–
772,
2010
.
[3]
A.
Phin
y
om
ark
,
et
al.
,
“
Applica
ti
on
of
W
ave
le
t
Anal
y
sis
in
E
MG
Feat
ure
Extrac
t
ion
for
Patt
e
rn
Cla
ss
ifi
cation
,
”
Me
as
urement
S
c
i
ence
R
ev
iew
,
vo
l
/i
ss
ue:
11
(
2
)
,
pp
.
45
–
52
,
2011
.
[4]
A.
C.
Tsa
i,
et
a
l
.
,
“
A
novel
STF
T
-
ran
king
featur
e
of
m
ult
i
-
cha
nn
el
EMG
for
m
ot
ion
patter
n
r
ec
o
gnit
ion,”
Ex
pert
Syst
em
wi
th
App
l
ic
at
ion
,
vol
/
issue:
42
(
7
)
,
pp
.
332
7
–
3341,
2015
.
[5]
R.
H.
Chowdhury
,
et
al
.,
“
Surfa
ce
Elec
trom
y
ogr
aph
y
Sign
al
Pro
ce
ss
ing
and
Clas
sifi
ca
ti
on
Tech
nique
s
,
”
S
ensors
,
vol
/i
ss
ue:
13
(
9
)
,
pp.
12431
-
1246
6,
2013
.
[6]
F.
A.
Om
ari
,
et
al.
,
“
Patt
e
rn
R
ec
ogni
ti
on
of
E
ight
Hand
Mot
i
ons
Us
ing
Feature
Ex
tra
c
ti
on
of
Forea
rm
EM
G
Signal
,
”
Proce
e
dings
of
the
Nati
onal
Ac
ademy
of
Sci
ences,
India
Sec
t
ion
A:
Phy
si
cal
Sci
en
ce
s
,
vol
/i
ss
ue:
84
(
3
)
,
pp
.
473
–
480,
2014
.
[7]
M.
R.
Cana
l
,
“
Com
par
ison
of
W
ave
le
t
and
Short
Ti
m
e
Fourie
r
Tra
nsform
Methods
in
the
Anal
y
s
is
of
EMG
Signal
s,”
Journa
l
of
medi
cal
syst
ems,
vol
/
issue:
34
(
1
)
,
pp
.
91
–
94
,
2010.
[8]
N.
M.
Kakot
y
,
e
t
al.
,
“
Expl
or
ing
a
famil
y
of
wav
el
e
t
tra
nsform
s
for
EMG
-
base
d
gra
sp
rec
ognition,
”
Signal
,
Image
and
Vi
d
eo Proces
sing
,
vol
/
issue:
9
(
3
)
,
pp
.
553
–
55
9,
2015
.
[9]
J.
Rafi
e
e,
et
a
l.
,
“
W
ave
le
t
b
asi
s
func
ti
ons
in
biomedic
a
l
sign
al
proc
essing,
”
Ex
pert
S
yst
em
wit
h
Appl
ic
at
io
n
,
vol
/i
ss
ue:
38
(
5
)
,
pp.
6190
–
6201
,
2011.
[10]
M.
Saini
,
e
t
al.
,
“
Algorit
hm
for
Fault
Locat
ion
and
Cla
ss
ifi
c
at
i
on
on
Para
ll
e
l
Tra
nsm
ission
Li
ne
using
W
ave
l
e
t
base
d
on
Cla
rke
’s
Tra
nsform
at
ion,
”
Int
ernati
o
nal
J
ournal
of
El
e
ct
r
ic
a
l
and
Comput
ute
r
Eng
ine
ering
.
IJE
C
E
,
vol
/i
ss
ue:
8
(
2
)
,
p
p.
699
–
710
,
201
8.
[11]
A.
Phin
y
om
ark
,
et
al
.
,
“
Feat
ur
e
Ext
ra
ct
ion
and
Reduc
ti
on
of
W
ave
le
t
Tra
nsf
orm
Coeff
ic
ie
n
t
s
for
EMG
Patt
ern
Cla
ss
ifi
c
at
ion
,
”
El
e
kt
ronika
ir
Elek
trot
ec
hni
ka
,
v
ol
/i
ss
ue:
1
22
(
6
)
,
pp.
27
–
32
,
2012
.
[12]
J.
Yous
efi
and
A.
H
.
W
right
,
“
Chara
c
te
ri
zi
ng
E
MG
dat
a
using
m
ac
hine
-
l
ea
rn
in
g
tool
s,”
Comput
er
in
Bi
ol
og
y
a
nd
Me
d
icine
,
vol
.
5
1
,
pp
.
1
–
13
,
201
4.
[13]
L.
H.
Sm
it
h,
et
a
l.
,
“
Dete
rm
ini
ng
the
Opt
imal
W
i
ndow
Le
ngth
for
Patt
ern
Re
cognition
-
Based
M
y
o
el
e
ct
ri
c
Contro
l:
Bal
an
ci
ng
the
C
om
pet
ing
Eff
ect
s
of
Cla
ss
ifi
ca
tion
Err
or
and
Control
ler
Delay
,
”
IEE
E
Tr
ansacti
ons
on
Neural
Syste
ms
and
Re
h
abil
it
a
ti
on
Engi
n
ee
ring
,
vol
/
issue:
19
(
2
)
,
pp.
186
–
192,
2011
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A D
et
ail Stu
dy of W
avelet
Famil
ie
s for
EM
G
P
attern
Rec
ogniti
on (J
ingw
ei
Too)
4229
[14]
M.
Hakone
n,
et
al.
,
“
Curre
n
t
sta
te
of
digital
sign
al
proc
essing
in
m
y
o
elec
tr
ic
in
terfac
es
and
r
elated
appl
icat
ions,
”
Bi
omedi
cal
S
ign
al
Proc
essing
an
d
Control
,
vol
.
1
8,
pp
.
334
–
359
,
2015.
[15]
J.
Rafiee
,
et
a
l.
,
“
Feat
ure
ext
r
action
of
fore
arm
EMG
signal
s
fo
r
prosthetics,
”
E
xpe
rt
S
yst
em
w
i
th
App
li
ca
ti
on
,
vol
/i
ss
ue:
38
(
4
)
,
pp.
4058
–
4067
,
2011.
[16]
L.
Fraiwa
n
,
e
t
al
.
,
“
Autom
at
ed
sl
ee
p
stag
e
ide
n
ti
f
ic
a
ti
on
s
y
st
em
base
d
on
ti
m
e
–
fre
quency
an
aly
sis
of
a
single
EE
G
cha
nne
l
and
ran
dom
fore
st
cl
assifie
r
,
”
Computer
methods
and
pr
ogram
s
in
biome
d
ic
in
e
,
vol
/
issue:
108
(
1
)
,
pp.
10
–
19,
2012
.
[17]
S.
H.
Cho,
et
al.
,
“
Ti
m
e
-
Freque
nc
y
Anal
y
sis
of
Pow
er
-
Quali
t
y
Di
sturbanc
es
vi
a
the
Gabor
W
igne
r
Tra
nsform
,
”
IEE
E
transacti
o
ns on
power
d
el
i
ve
ry
,
vol
/i
ss
ue:
25
(
1
)
,
pp
.
494
–
49
9,
2010
.
[18]
A.
Subasi,
“
Cla
s
sific
a
ti
on
of
EMG
signal
s
using
PS
O
opti
m
iz
ed
SV
M
for
dia
gnosis
of
neur
om
uscula
r
d
isorder
s,
”
Comput.
B
iol.
M
ed.
,
vol
/i
ss
ue:
43
(
5
)
,
pp
.
576
–
586
,
2013
.
[19]
M.
H.
D.
Moham
m
adi
,
“
Im
prov
ed
Denoising
Method
for
Ultra
s
onic
Ec
ho
wi
th
Mother
W
ave
let
Opt
imiza
ti
on
a
nd
Best
-
Basis
Selec
ti
on,
”
Int
ernat
io
nal
J
ournal
of
E
le
c
tr
ic
al
and
Co
mput
ute
r
Eng
in
e
ering.
IJ
ECE
,
vo
l
/i
ss
ue:
6
(
6
)
,
pp.
2742
–
2754,
201
6.
[20]
A.
Subasi,
“
Cla
ss
ifi
cation
of
EM
G
signal
s
using
combined
featur
es
and
soft
computing
techniqu
e
s,”
Appl
ied
Sof
t
Comput
ing
,
vo
l
/is
sue:
12
(
8
)
,
pp.
2188
–
2198,
201
2.
[21]
S.
V.
S.
Prasad,
et
al
.
,
“
Com
par
ison
of
Acc
ura
c
y
Me
asure
s
for
RS
Im
age
Cla
ss
ifi
cation
using
SV
M
and
ANN
Cla
ss
ifi
ers,”
Int
ernati
onal
J
ournal
of
E
le
c
tr
ic
al
and
Comput
uter
Eng
ine
ering
.
I
JE
CE
,
vol
/
issue:
7
(
3
)
,
pp.
1180
–
1187,
2017
.
[22]
A.
Subasi
and
M.
I
.
Gurs
o
y
,
“
EE
G
signal
cl
as
sific
a
ti
on
using
PC
A,
ICA,
LDA
and
support
vec
tor
m
ac
h
ine
s
,
”
Ex
pert
Syst
em
w
it
h
Appl
i
cat
ion
,
vol
/i
ss
ue:
37
(
12
)
,
pp
.
8659
–
8666
,
2010.
BIOGR
AP
H
I
ES
OF
A
UTH
ORS
Too
Jing
W
ei
has
rec
ei
ved
his
B.
Eng.
from
Univer
siti
T
ekni
ka
l
Malay
s
ia
in
201
7.
He
is
cur
ren
t
l
y
pursuing
his
M
aste
r
Eng
.
in
Univer
siti
T
ekn
ika
l
Mal
a
y
s
ia.
His
rese
arc
h
ar
ea
s
are
in
sign
al
proc
essing,
class
ifi
c
at
ion
and
fe
ature
sel
ection
for
EMG pa
ttern
r
ecogniti
on.
A
ss
oci
at
e
Prof. D
r.
Abdul
Rahim
Bin
Abdulla
h
has
recei
ved
h
is
B.
Eng.,
Mast
er Eng.
,
PhD
Degre
e
from
Univer
siti
Te
knologi
Ma
lay
sia
in
2001
,
2004
and
2011
in
El
e
ct
ri
ca
l
Eng
in
ee
ring
and
Digi
t
al
Signal
Proce
ss
in
g
respe
c
ti
ve
l
y
.
He
is
cur
r
ent
l
y
an
As
socia
t
e
Pr
ofe
ss
or
with
the
Depa
rtment
of
El
e
ct
ri
ca
l
Eng
in
ee
ring
for
Univ
e
rsiti
Te
knik
al Mal
a
y
si
a
Mel
aka (
UTe
M).
Dr.
Norhashim
ah
Bint
i
Mohd
Saad
is
cur
ren
tly
working
as
a
senior
lectur
er
in
Depa
rtment
Com
pute
r,
FK
EKK,
UTe
M.
She
fini
shed
h
er
stud
y
in
Bac
h
e
lor
of
Engi
n
ee
r
ing,
Master
of
Engi
ne
eri
ng
and
PhD
in
Medi
cal
Im
age
Proc
essing
from
UTM, M
al
a
y
s
ia.
Evaluation Warning : The document was created with Spire.PDF for Python.