TELKOM
NIKA
, Vol. 11, No. 12, Decem
ber 20
13, pp.
7339
~73
4
3
e-ISSN: 2087
-278X
7339
Re
cei
v
ed
Jun
e
25, 2013; Revi
sed
Jul
y
2
8
, 2013; Acce
pted Augu
st 15, 2013
A New Method for Intrusion Detection using Manifold
Learning Algorithm
Guoping Ho
u
1
, Xuan Ma
1
, Yuelei Zhang
2
1
Cho
n
g
q
in
g El
ectric Po
w
e
r C
o
lle
ge, No.
9
W
u
lo
ngmi
ao, Jiu
l
ong
po District, Cho
ngq
in
g 40
0
053, Ch
in
a
T
e
l: 86-23-68
5
014
90
2
Unit 94
27
0 of PLA, No.19 H
a
ngk
o
ng R
oad,
Jina
n 25
011
7, Chin
a
T
e
l: 86-521-8
6
710
09
1
Corresp
on
din
g
author, e-mai
l
: houg
uo
pin
g
1
9
82@
163.com
1
,
z
y
l3
83
6@1
63.
com
2
A
b
st
r
a
ct
Co
mp
uter an
d
netw
o
rk security has rec
e
ive
d
an
d w
ill still receiv
e
much
attenti
on. Any
unex
pecte
d i
n
trusio
n w
ill
da
mage t
he
netw
o
rk. It is theref
or
e i
m
p
e
rativ
e
to
detect th
e n
e
tw
ork intrusio
n t
o
ensur
e th
e n
o
r
ma
l o
per
atio
n
of the
i
n
tern
et. T
here
are
ma
ny stu
d
ies
in th
e i
n
trusi
o
n d
e
tectio
n a
n
d
intrusi
on p
a
tter recog
n
itio
n. T
he artifici
al n
e
u
r
al netw
o
rk (A
NN) has
prove
n
to be p
o
w
e
rful for the i
n
trus
io
n
detectio
n
. H
o
w
e
ver, very
little
w
o
rk has
disc
ussed
the
opti
m
i
z
at
io
n of th
e
inp
u
t i
n
trusio
n
features
for th
e
ANN. Gen
e
ral
l
y
, the i
n
trusi
o
n
features
co
nta
i
n
a cert
ain
n
u
m
b
e
r
of us
eles
s features,
w
h
i
c
h is
use
l
ess
for
the i
n
trusio
n d
e
tection.
Larg
e
di
m
ens
ions
of the fe
ature
data w
i
l
l
a
l
so
affect the i
n
trusio
n d
e
tectio
n
perfor
m
a
n
ce
of the AN
N. In
or
der to
i
m
pr
ove
the
ANN
p
e
rfor
ma
nce,
a n
e
w
appr
oach
for
n
e
tw
ork intrusi
o
n
detectio
n
base
d
o
n
non
lin
ea
r feature
d
i
me
nsio
n re
duc
tion
an
d AN
N i
s
p
r
op
o
s
ed
i
n
th
i
s
wo
rk. Th
e
ma
nifo
ld le
arni
ng al
gorith
m
w
a
s used to red
u
ce the
intrus
i
on feature vect
or. T
hen an ANN classifi
er wa
s
empl
oyed
to
id
entify the
i
n
tru
s
ion. T
h
e
effici
ency
of the
pr
opos
ed
metho
d
w
a
s ev
al
uat
ed w
i
th th
e r
e
al
intrusi
on d
a
ta. T
he test result show
s tha
t
t
he prop
ose
d
ap
proac
h h
a
s go
od i
n
tru
s
ion
detectio
n
perfor
m
a
n
ce.
Ke
y
w
ords
: intr
usio
n detecti
on
, nonli
n
e
a
r feature red
u
ctio
n,
artificial
neur
al
netw
o
rk, man
i
fold l
earn
i
n
g
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Great adva
n
c
e
s
have be
en made in
the field of commu
nication and co
mpute
r
techn
o
logy in
re
cent ye
ars.
The i
n
ternet
now is
indi
sp
ensi
b
le fo
r p
e
ople’
s life. Ho
wever,
due
to
compl
e
x operation enviro
n
m
ent, the network suffers fr
om vario
u
s o
ffensives an
d
violations. It is
therefo
r
e im
p
e
rative to
de
tect the
network intr
usi
o
n
to en
su
re t
he n
o
rm
al o
peratio
n of t
he
intnet. The intrusi
on dete
c
ti
on re
sea
r
ch h
a
s
he
nce re
ceived extensi
v
e attentions
[1].
Intrusio
n dete
c
tion is
cruci
a
l for com
puter
security and
defen
se. Terrible intru
s
ion
may
damag
e the internet for
wee
k
s [2]. To reali
z
e
effe
ctive intru
s
io
n detectio
n
, many advan
ced
techn
o
logie
s
have b
een
propo
sed, i
n
cl
u
d
ing th
e
a
r
tificial neu
ral ne
twork (ANN)
[3],
roug
h set
s
[1], and sup
port vector
machi
ne (SV
M
) [4] et
c. Among them, ANN is the
most pro
m
isi
ng
method [5]. ANN h
a
s the a
b
ility to find the nonlin
ea
r con
n
e
c
tion b
e
twee
n the Intrusio
n features
and the Intru
s
ion
pattern
s,
and h
a
s
bee
n wid
e
ly us
e
d
in the Intru
s
i
on dete
c
tion.
Ho
wever, A
N
N
detectio
n
pe
rforman
c
e i
s
mainly dete
r
mined
by it
s stru
ctu
r
al p
a
ram
e
ters, e
s
pe
cially by
the
input featu
r
e
vector of the
i
n
trusi
o
n
data.
Alt
houg
h the
pri
n
ci
pal
co
mpone
nt a
nal
ysis
(PCA
)
an
d
its derivative
algorithm
s
have bee
n proved to be
a useful tool
for feature
redu
ction a
n
d
extraction
to
improve
the
netwo
rk attack d
e
tect
ion accuracy, thei
r main li
mitation lies in their
ability is to capture the nonlinea
r properties of the original dat
a [6-8]. The sam
e
problem
s
are
also fou
nd in
other line
a
r
method
s [7], inclu
d
ing m
u
lti-dimen
s
io
na
l scali
ng (MDS) and lin
ea
r
discrimi
nate
analysi
s
(L
DA). Fortunat
ely, the
manifold learni
ng
algorithm
s
provide a n
e
w
mean
s to dea
l with the nonl
inear dim
e
n
s
i
onality redu
ct
ion pro
b
lem
s
. The Isoma
p
[6] and locally
linear em
bed
ding (L
LE) [7] etc., are able to deal
with the underlyin
g nonline
a
r b
ehavior of th
e
data. Co
mpa
r
ed
with the
linear metho
d
s, the
purp
o
se
of manif
o
ld lea
r
nin
g
i
s
to p
r
oje
c
t t
he
origin
al hig
h
-dimen
sion
al
data into
a lo
wer di
me
nsi
o
nal featu
r
e
space by p
r
e
s
erving th
e lo
cal
topology of t
he o
r
iginal
d
a
ta [9-1
0]. Th
us, the i
n
trin
sic st
ru
cture
o
f
the data of i
n
tere
st can b
e
extracted effe
ctively.
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: 733
9 – 7343
7340
The a
d
vantag
e of the
Isom
ap a
nd
LLE i
n
the i
n
tru
s
io
n dete
c
tion
is the id
entifica
t
ion o
f
unde
rlying n
o
n
linea
r manif
o
ld. Ho
weve
r, in the ident
i
f
ication of n
o
nlinea
r manif
o
ld the Isom
ap
and L
L
E ma
inly use
the
neigh
bo
rhoo
d graph
s, which
so
meti
mes m
a
y fai
l
to en
sure t
he
con
n
e
c
tedn
e
s
s of th
e
con
s
tru
c
ted
nei
g
hborhoo
d g
r
a
phs. In
o
r
de
r
to improve th
e robu
stne
ss
of
the neig
hbo
rhood
gra
p
h
s
of the manifo
ld learning
al
grithm to e
n
h
ance the int
r
usio
n dete
c
ti
on,
this wo
rk p
r
ese
n
ts a ne
w method b
a
se
d on the
improved L
L
E. An adap
tive schem
e
is
prop
osed to
build
suitabl
e
neigh
bo
rhoo
d graph
s. By
doing
so, it is rea
s
on
able
to red
u
ce th
e
origin
al featu
r
e spa
c
e an
d
find the disti
n
ct nonl
i
nea
r cha
r
a
c
teri
stics a
bout the
intrusi
on data
.
Then,
an A
N
N
cla
ssifie
r
i
s
empl
oyed to
re
co
gni
ze
th
e intrusi
on
p
a
tterns.
By i
m
pleme
n
ting
the
intrusi
on det
ection expe
ri
ments, the a
nalysi
s
re
su
lt
s sh
ow that the feature re
ductio
n
is very
essential in the intru
s
ion
detectio
n
because t
he orig
inal feature
spac
e have m
any redu
nda
nt
feature
s
to inf
l
uen
ce the int
r
usi
on id
entificati
on. Elimin
ate these re
d
unda
nci
e
s
ca
n enh
an
ce th
e
intrusi
on dete
c
tion. In addit
i
on, the com
p
arisi
on of the
improved L
L
E
and the ori
g
inal LLE ha
s
been don
e. The comp
ari
s
ion re
sult shows
that
th
e improved
LLE with
ad
aptive sche
me
outperfo
rm
s the origi
nal LL
E in the intrusion dete
c
tion.
2. Rese
arch
Metho
d
2.1. The Improv
ed LLE
Here in we p
r
opo
se a
n
ad
aptive schem
e to build sui
t
able neig
hbo
rhoo
d graph
s. The
details a
r
e ex
pre
s
sed bel
o
w
.
Given a nonli
near
high
-di
m
ensi
onal d
a
t
aset
12
[]
p
l
SR
ss
s
, where
l
is
the total
sampl
e
num
b
e
r an
d
p
the
dimen
s
ion
a
lity of each
sa
mple, the
obj
ective of LLE
is to re
co
nstruct
a nonli
nea
r
mappin
g
to
proje
c
t
S
int
o
a redu
ced
manifold
sp
ace
12
[]
q
rr
r
r
l
SR
ss
s
(
q
<<
p
). The i
m
prove
d
LLE
algorithm i
s
descri
bed a
s
followin
g
.
Step 1: Comp
ute
k
neig
hbo
urs of eve
r
y sample.
Step 2: Identify the neighborh
ood g
r
a
ph and fi
nd
s the points o
u
t of the con
necte
d
grap
h.
Step 3: Incre
a
se
k
if exist unconn
ecte
d points. Othe
rwise, go to step 5
Step 4: Comp
ute new
k
ne
are
s
t neigh
bo
rs.
Step 5: Compute the local
re
co
nst
r
u
c
tion weight matrix
W
by minimizing the fol
l
owi
n
g
co
st function:
2
11
mi
n
(
)
(
)
lk
i
ji
i
j
ij
Ww
s
s
-
(1)
Whe
r
e
i
j
w
is the weig
ht values. If
i
s
and
j
s
are not
neighb
ours,
0
i
j
w
and
1
1
i
j
k
w
j
.
Step 6: Map
the origin
al
dataset int
o
the embe
dded
coo
r
di
nates. Comp
ute the
r
e
co
ns
tr
uc
te
d
q
-dime
n
si
on
al manifold space
r
S
by minimizing the foll
owin
g co
nstraint:
2
11
mi
n
(
)
lk
i
r
i
j
rij
ij
s
Sw
s
r
(2)
Whe
r
e
ri
s
is the proje
c
tion ve
ctor of
i
s
in the embed
ded
co
ordin
a
tes, an
d
rij
s
are the n
e
ig
hbou
rs
of
ij
s
. Equation (2) ca
n be rewritten as:
11
mi
n
(
)
(
)
lk
iT
T
r
j
ri
rj
r
r
ij
Sm
s
s
t
r
S
M
S
,
(3)
Whe
r
e the
co
st matrix
M
can be expressed a
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A New Meth
o
d
for Intrusi
o
n
Detectio
n usi
ng Manifold L
earni
ng Algo
rithm
(Guopin
g
Hou
)
7341
()
()
T
ll
ll
M
IW
IW
.
(4)
Hen
c
e, the
m
i
nimizatio
n
of
(4) can b
e
re
duced to
an
eigenvalu
e
p
r
oblem, a
n
d
r
S
co
uld
be determine
d by the
q
sm
allest no
nzero eigenve
c
tors of
M
.
2.2. The Bac
k
propa
gatio
n
Neur
al Netw
o
r
k (BPNN)
A neu
ral
net
work ha
s
a
natural
p
r
op
ensity fo
r
storing
expe
rie
n
tial kno
w
led
ge a
n
d
makin
g
it available fo
r u
s
e. Then the
Input
-O
utput Mappin
g
p
r
o
perty
and ca
pability
ca
n be
provide
d
by the ANNs [8
-13]. One of
the mo
st co
mmonly use
d
sup
e
rvi
s
ed
ANN mo
del
is
BPNN. The B
P
NN u
s
e
s
o
n
e
of the well-kno
w
n
algo
rithms, ba
ckp
r
o
pagatio
n lea
r
ning alg
o
rith
m
[13]. The structure of BPNN i
s
arra
ng
ed into di
fferent layers: input layer, mi
ddle laye
r an
d
output layer. The wo
rkflow of the BPNN
can b
e
expre
s
sed a
s
follo
ws [8
-13]:
(1) Th
e in
put
layer
pro
pag
ates
a
parti
cu
lar i
nput ve
ct
or’s
comp
one
nts to
ea
ch
n
ode i
n
the middle la
yer.
(2)
The mi
ddl
e layer n
ode
s comp
ute out
put value
s
, which
be
come
inputs to th
e
node
s
of the output layer.
(3) T
he outpu
t layer node
s comp
ute the netwo
rk
o
u
tp
ut for the part
i
cula
r input vector.
The forwa
r
d
pass p
r
od
uces a
n
outp
u
t vector fo
r a
given inp
u
t vector b
a
sed
on th
e
curre
n
t state
of the network
wei
ghts.
Since
the
ne
twork weight
s are initiali
zed to ra
ndo
m
values, it is
unlikely that rea
s
on
able
o
u
tputs
will re
sult befo
r
e training. Altho
ugh the
wei
g
hts
can
be
adju
s
ted to redu
ce the e
r
ror
b
y
prop
agat
in
g the o
u
tput
error
ba
ckwa
rd th
roug
h th
e
netwo
rk,
the
traini
ng
ma
y suffer fro
m
the l
o
cal mi
nimum. T
h
u
s
, the imp
r
ov
ed L
L
E featu
r
e
redu
ction i
s
a
n
efficient co
mpen
sation t
o
the training
pro
c
e
ss of th
e ANN.
2.3. The Proposed Intr
us
ion Detectio
n Method
In this pape
r the improved
LLE-ANN are use
d
for the netwo
rk intrusio
n detecti
on. Th
e
prop
osed net
work intrusi
o
n detectio
n
proce
s
se
s are
given as follo
ws:
Step 1: Pre-treat the origin
al netwo
rk in
t
r
usi
on data to
standa
rdi
z
ed
data format.
Step 2: Extra
c
t distin
ct features from the
input network int
r
u
s
ion
data in the form of
manifold by improve
d
LLE
.
Step 3: T
r
ain
the BP
NN u
s
ing
the
ne
w features,
an
d dete
r
min
e
t
he n
e
two
r
k in
trusio
n
detectio
n
re
sult accordi
ng
to each ANN
model outp
u
t.
Step 4: Test
the perfo
rma
n
ce of the
propo
sed n
e
twork i
n
tru
s
ion
detectio
n
mo
del, and
provide the t
e
st re
sult as the base fo
r a va
lid network intru
s
io
n manag
eme
n
t deci
s
ion. A
diagram block of the proposed network i
n
trusi
on detection method is illustrated i
n
Figure 1.
Figure 1. The
Netwo
r
k Intrusio
n Dete
ct
i
on ba
sed o
n
the Improve
d
LLE and ANN
2.4. Experiments
In orde
r to e
v
aluate the
perfo
rman
ce
of
the prop
o
s
ed
com
pute
r
intru
s
ion m
e
thod,
experim
ent tests
have b
een impl
eme
n
ted in th
is work. Figu
re 2 sh
ows t
he expe
rime
nt
prin
ciple. A mini com
put
er network h
a
s be
en
esta
blish
ed to co
ndu
ct the experim
ents. T
he
comp
uter net
work is comp
ose
d
of
a lin
u
x
se
rver
, a
wi
ndo
ws serve
r
, the web
link,
two li
nux h
o
st
and fo
ur
win
dows
host
s
.
Some ma
nua
l attacks hav
e be
en
simul
a
ted a
nd te
sted u
s
ing thi
s
experim
ental system.
In this wo
rk, the De
nial of
Service
(DoS),
Rem
o
te to Local (R2L
), User to
Root
(U2
R
)
and Pro
be o
r
Sca
n
(PoS
) are intro
d
u
c
ed into
th
e
experime
n
t system to va
lidate the ne
w
detectio
n
method. Forty-on
e feat
ure
s
are monitored
and re
co
rd
ed
for every intrusi
on. The
s
e
feature
s
in
cl
ude the byte
s issue
d
fro
m
sou
r
ce
to destin
a
tion, the bytes fro
m
destin
a
tio
n
to
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: 733
9 – 7343
7342
sou
r
ce, duration, teard
r
op,
nept
un
e, etc. There a
r
e 5,
000 sample
s
for each intru
s
ion type an
d
the total sam
p
les a
r
e 20,0
00.
Figure 2. The
Principl
e of the Experime
n
t Tests
3. Results a
nd Analy
s
is
In experi
m
ent
s, L
L
E was a
dopted
to red
u
ce
the
41 di
mensi
on
of th
e ori
g
inal
dat
a to 2
and 3
dime
n
s
ion
s
, respe
c
tively. Figure
3 sho
w
s th
e
feature
re
du
ction
re
su
lt o
f
the improve
d
LLE with 2 di
mensi
o
n
s
, an
d Figure 4 shows the
feat
ure redu
ction
result
of the improve
d
LL
E
with 3
dimen
s
ion
s
. It ca
n
be seen
from
Figure 3
and
Figure 4 th
at after featu
r
e
redu
ction
the
r
e
are o
b
viou
s
boun
deri
e
s
b
e
twee
n different intru
s
ion
type besi
d
e
s
som
e
overl
a
ps. Hen
c
e, the
feature
red
u
ction ca
n effici
ently eliminte
the usel
e
ss f
eature
s
in
th
e origi
nal feat
ure ve
ctor
an
d
extract the m
o
st distin
gui
shed informati
on for t
he intrusio
n dete
c
tion. Mo
re
over, it can provi
de
a virtual
3D
pre
s
entatio
n
of the intrusi
on feat
u
r
e
space u
s
ing
3
dimen
s
io
ns
in the featu
r
e
redu
ction.
Figure 3. Fea
t
ure re
du
ction
result of the
Improved L
L
E
with 2 Dim
ensi
o
n
s
Figure 4. Fea
t
ure Re
du
ctio
n Re
sult of the
Improved L
L
E
with 3 Dim
ensi
o
n
s
In the intru
s
i
on re
co
gnitio
n
, 3, 000 sa
mples
of ea
ch intru
s
ion ty
pe was u
s
e
d
for train
the ANN, an
d
the remind
ers we
re u
s
ed f
o
r testi
ng. Ta
ble 1 lists the
intrusio
n det
ection results
usin
g the
pro
posed m
e
tho
d
. He
re in
th
e dete
c
tion
rate and
false
positive
rate
we
re u
s
e
d
t
o
evaluate the
ntrusio
n
det
ection pe
rformance. The detectio
n
rat
e
mean
s the
hits of correct
sampl
e
s to th
e total sampl
e
s an
d the false po
sitive ra
te is defined
as the mi
scla
ssifi
cation
s.
-2
0
2
4
6
8
10
12
-2
0
2
4
6
8
10
12
F
eat
u
r
e 1
F
eat
u
r
e 2
U2
R
Do
S
R2
L
Po
S
-2
0
2
4
6
8
10
1
2
-5
0
5
10
15
-4
-2
0
2
4
6
8
F
e
a
t
u
r
e
2
F
e
a
t
u
r
e
1
Fe
a
t
ur
e
3
Do
S
R2
L
U2
R
Po
S
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A New Meth
o
d
for Intrusi
o
n
Detectio
n usi
ng Manifold L
earni
ng Algo
rithm
(Guopin
g
Hou
)
7343
Table 1. The
Perform
a
n
c
e
of the ANN Detection
Feature
reductio
n
method
KPCA-PSO
-SVM
Detection rate (
%
)
False positive rate (%
)
None
81.6
13.8
LLE
w
i
th 2 dimen
s
ions
90.2
6.4
LLE
w
i
th 2 dimen
s
ions
90.6
6.8
Improved LLE
w
i
th 2 dimensions
93.8
6.6
Improved LLE
w
i
th 3 dimensions
94.4
6.4
The intru
s
io
n
detection
pe
rforma
nce of
the
use of L
L
E feature
selectio
n and
without
the LLE sele
ction is com
pare
d
in Tab
l
e 1. It
can be see
n
fro
m
Table 1 that by the LLE
pro
c
e
ssi
ng, t
he di
stinct fe
ature
s
a
r
e o
b
t
ained a
nd th
us the i
n
tru
s
i
on dete
c
tion
rate is e
nha
nced
by 8.6% or better and the
false po
sitive rate is
d
e
creased at lea
s
t by 7.4%. H
ence, it can be
see
n
that th
e
LLE fe
ature
sele
ction
can
improve
the
intrusi
on dete
c
tion rate
efficiently.
It
also
can
be n
o
tice
d that the im
proved
LLE i
n
crea
se
s th
e
detectio
n
rate by 3.6% or better tha
n
th
e
LLE.
4. Conclusio
n
Intelligent me
thod h
a
s be
e
n
wi
dely u
s
e
d
in i
n
tru
s
ion
detectio
n
, e
s
peci
a
lly for th
e ANN
based meth
o
d
s. However,
rea
s
on
able i
nput featur
e
vector of the
ANN mo
del
plays a
criti
c
l
e
role in de
sired dete
c
tion
perform
an
ce. Therefo
r
e,
this pape
r prop
osed a
new intrusi
o
n
detectio
n
method ba
sed o
n
the improv
ed LLE and
BP
NN. The i
nnovation of this wo
rk i
s
that
the ne
w met
hod u
s
e
s
th
e improved
LLE algo
rith
m to redu
ce
the dimen
s
i
ons
of the input
feature
s
of th
e BPNN to
el
iminate u
s
ele
ss i
n
form
atio
n. The real p
r
acti
ce
data
wa
s ap
plied t
o
the validation
of the p
r
op
o
s
ed
app
ro
ach. The
anal
y
s
is re
sult
s ve
rify the effect
iveness of thi
s
method.
Referen
ces
[1]
Z
hao
X, Jin
g
R
,
Gu M. Adaptive intrusi
on
det
ection
alg
o
rith
m base
d
on r
o
ugh s
e
ts.
J T
singh
ua U
n
iv
(Sci & Tech)
. 2008; 48: 1
165-
116
8.
[2]
Sin Y,
Lo
w
W
,
W
ong
P.
Lear
ning finger
print
s
for a da
tabase intrusion
detection system
.
Proc. 7t
h
Eur. S
y
mp. Re
search i
n
Com
puter Secur
i
t
y
.
Z
u
rich. 200
2; 7
:
264-28
0.
[3]
Li Z
,
Yan X, Yuan C, Z
hao
J, Peng Z
.
F
ault detecti
on
and d
i
ag
nos
is
of the gearb
o
x
in mar
i
n
e
prop
ulsi
on s
y
stem base
d
on
bisp
ectrum an
al
ysis an
d artificial n
eur
al net
w
o
rks.
Journ
a
l
of Marine
Scienc
e an
d Applic
atio
n
. 201
1; 10: 17-2
4
.
[4]
Li Z, Yan
X, Yuan C, Pe
ng
Z. Intellige
n
t faul
t di
ag
nosis
method for ma
rine d
i
es
el e
n
g
i
nes us
in
g
instanta
n
e
ous ang
ular
s
p
e
ed.
Journ
a
l of Me
chan
ical Sc
ien
c
e an
d T
e
chn
o
l
ogy
. 2
012; 2
6
(
8
): 241
3–
242
3.
[5]
Lee S, H
e
in
bu
ch D. T
r
aining
a ne
ural
net
w
o
rk-
base
d
intrus
ion
detector to
recog
n
ize
nov
el attacks
.
IEEE Trans. S
ystem
s, Man and Cyber
netics
Part A: Syste
m
s and Humans
. 2001; 31: 29
4-29
9.
[6]
T
enenba
um J, Silva V, Lan
gford J. A glo
bal
g
eometric
frame
w
ork for
nonl
ine
a
r dim
ensi
ona
lit
y
reducti
on.
Science
. 200
0; 290
: 2319–
23
23.
[7]
Ro
w
e
is S, Sau
l
L. Nonl
in
ear
dime
nsio
nal
it
y
reducti
on b
y
l
o
call
y li
ne
ar em
bed
din
g
.
Science
. 200
0;
290: 23
23-
232
6.
[8]
Belki
n
M, Ni
yogi P. La
plac
i
an ei
ge
nmaps
fo
r dimensi
o
n
a
lit
y
r
e
d
u
ction
and d
a
ta rep
r
esentati
o
n
.
Neur
al Com
put
. 2003; 15: 13
7
3
-13
96.
[9]
Jian
g Q, Jia
M
,
Hu J,
Xu F
.
Machi
ner
y fa
ul
t dia
gnos
is
usi
ng s
uperv
i
se
d
manifo
ld
lear
ni
ng.
Mec
h
.
Syst. Signal Pr
oces
. 200
9; 23
: 2301–
23
11.
[10]
Li Z
,
Yan X,
T
a
in Z
,
Y
uan C, Peng Z
.
Blind
vibration com
pon
ent sep
a
rat
i
on a
nd no
nli
n
ear featur
e
extracti
on a
p
p
lied to the n
o
n
station
a
r
y
v
i
b
r
ati
on si
gna
ls for the gearb
o
x mu
lti-fault
dia
gnos
is.
Measur
e
m
ent
. 201
3; 46: 259
–
271.
[11]
Z
hu S, Hu
B.
H
y
brid fe
ature
selectio
n bas
ed on im
prove
d
GA for the i
n
trusio
n detect
i
on s
y
stem.
TEL
K
OMNIKA
. 2012; 1
1
(4): 1
725
–1
730.
[12]
Che
n
X, L
i
F
,
W
ang
L.F
ault
-
tolera
nt me
th
od for
detecti
ng c
o
vera
ge
hol
es of
w
i
re
l
e
ss se
nso
r
net
w
o
rks.
TELKOMNIKA
. 2012; 10(4): 8
76–
882.
Evaluation Warning : The document was created with Spire.PDF for Python.