TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 10, Octobe
r 20
14, pp. 7337
~ 734
2
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.519
7
7337
Re
cei
v
ed
No
vem
ber 2
3
, 2013; Re
vi
sed
Jul
y
11, 201
4
;
Accepte
d
Augu
st 5, 2014
A Fault Detection Mechanism of Tunnel based on
Artificial Neural
Liu Liu*, Ma Chen
gqian
Department of
Computer Sci
ence,
W
u
Ha
n Univers
i
t
y
of
T
e
chn
o
lo
g
y
*Corres
p
o
ndi
n
g
author, e-ma
i
l
:270
20
849
3@
qq.com
A
b
st
r
a
ct
T
h
is p
aper
h
a
s
mad
e
a
q
ual
itative
and
q
u
a
n
t
itativ
e a
nalys
i
s
by
establ
ishi
ng th
e tu
nne
l f
ault tre
e
and
givi
ng th
e
mi
ni
ma
l cut set
s
of the fau
l
ts i
n
tunn
el, a
nd t
e
sted th
e dat
a
in tun
nel
co
mb
i
ned w
i
th
artifici
al
neur
al netw
o
r
k
. T
he fault detection
mec
h
anis
m
in
this
article has b
een si
mu
late
d
by MatLab a
nd
process
ed
a l
o
t of the act
ual
data
t
h
rou
gh
the tun
nel
o
p
e
r
ating
histor
y.
Experi
m
ental
r
e
sults sh
ow
th
at:
T
h
is fault detec
tion mech
anis
m
is effective.
Ke
y
w
ords
:
arti
ficial n
eura
l
, fault detectio
n
of tunne
ls, fault tree
Co
p
y
rig
h
t
©
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
With the
a
c
celeratio
n
of
urba
nization,
urb
an tran
sport i
s
e
n
co
unterin
g a
n
i
n
crea
sing
pre
s
sure, whi
l
e the
tunn
el
unde
r th
e
city has
signi
ficantly ea
sed
t
he p
r
e
s
sure
on u
r
b
an t
r
af
fic.
But the tunn
el is
a very
compl
e
x envi
r
onm
ent. In
orde
r to
ma
ke the tun
nel
run
safely a
n
d
efficiently, and to avoid traffic accident
s, the
manag
ement of the tunnel
for its daily operati
on
requi
re
s a
n
e
ffective meth
od. Tun
nel in
telli
gent mo
nitoring
and
se
curity ma
nag
ement a
r
e
ne
w
r
e
sear
ch topic
s
.
This
pap
er
h
a
s
com
b
ine
d
with ma
ny tu
nnel
s in
ope
ration, Wuhan
Shuigu
ohu t
unnel, the
unde
rg
roun
d passa
ge of Wuh
an Han
k
ou Rail
way Station, Wuha
n Zhong
sh
an
Road u
nde
rpass
and tun
nel
s
of the Intern
ational Expo
Cente
r
in
Wuhan
and
so
on. A set of
effective fail
ure
detectio
n
me
cha
n
ism h
a
s
been e
s
tabli
s
hed in this a
r
ticle, and thi
s
pape
r also ha
s given a typical
failure
dete
c
ti
on p
r
o
c
e
s
s a
nd m
e
thod
s
by analy
z
ing
and
studyin
g the
a
c
tual
failure
dete
c
ti
on
mech
ani
sm.
2. The Fault
Tree of
the Intellig
ent Monitoring S
y
stem in Tunnel
2.1. The Con
ception o
f
th
e Fault Tre
e
Fault tre
e
a
n
a
lysis technol
ogy is a
co
m
p
lex te
chni
cal
tool, an
alyzi
ng thin
gs logi
cally a
nd
vividly. Fault tree an
alysi
s
is
calle
d F
T
A for sh
ort,
a techni
que
develope
d
by the Ameri
c
an
Teleg
r
ap
h Co
mpany Bell L
abs i
n
19
62. It has ve
ry dist
inctive features. It is not
o
n
ly intuitive, it’s
thinkin
g
cle
a
r, also its lo
gi
c rig
o
ro
us [1]
.
It can be u
s
ed to do n
o
t only qualitati
v
e analysi
s
, but
also
qua
ntitative analysi
s
. It is one
of the pr
im
a
r
y analysi
s
method
s for safety sy
stem
engin
eeri
ng.
From thi
s
study, we
can see it
s
cha
r
a
c
teri
stics, integ
r
ity, expan
sibility, and
abstractio
n
. In ad
dition, fault tree
an
a
l
ysis i
s
a
ma
jor
symbol
of the d
e
velop
m
ent of
safe
ty
system
s engi
neeri
ng.
2.2. The In
tr
oduction
o
f
the
Har
d
w
a
r
e
of
th
e Tu
nnel Moni
to
ring Sy
stem and
the
Fa
ult
Information of
the Dev
i
c
e
s
The device
s
, installed i
n
a typical t
unnel, are g
enerally divided into info
rmation
colle
ction
de
vices an
d co
ntrollabl
e d
e
vice
s. The
inf
o
rmatio
n
coll
ection
device
s
in
clu
de veh
i
cle
insp
ectio
n
, te
mperature
/
humidity sen
s
ors, lig
ht
se
nso
r
s,
win
d
speed
/ directi
on
sen
s
o
r
s a
n
d
fire alarms. T
he co
ntrolla
bl
e equipm
ent
covers fa
n
s
, driveway lights, inca
nde
sce
n
t / sodium a
nd
so on.
Vehicle
inspe
c
tion i
s
resp
onsi
b
le for collectin
g the
numbe
r a
nd
the sp
eed
of vehicle
s
throug
h the t
unnel. A tunn
el may be eq
uippe
d with a
few of vehicl
e insp
ectio
n
device
s
[2]. The
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 733
7
– 7342
7338
fault inform
ation of th
e vehicl
e i
n
sp
ectio
n
d
e
vice g
ene
rally inclu
d
e
s
power
out
age,
comm
uni
cati
on failures,
a
nd data
erro
rs. Temp
erat
u
r
e / humi
d
ity sen
s
o
r
s are u
s
ed to
gathe
r th
e
temperature
and hu
midity inside th
e tunnel, a tun
nel
is g
ene
ra
lly installed
with at lea
s
t two
sen
s
o
r
s. It
s f
ault informati
on g
ene
rally i
n
clu
d
e
s
p
o
wer o
u
tage,
co
mmuni
cation
failure, d
e
tect
ing
data errors. L
i
ght intensity sen
s
o
r
is respon
sible fo
r
collectin
g the light intensity insid
e
the tun
nel
in ord
e
r to
co
ntrast
with th
e external li
g
h
t intens
ity, a
nd its failu
re i
n
formatio
n g
enerally incl
u
des
power outa
g
e
s,
comm
uni
cation
failu
re
s, dete
c
ting
data e
r
rors.
Speed
/ di
re
ction
se
nsor is
respon
sibl
e for coll
ectin
g
the informatio
n of
the wind speed a
nd its dire
ction in
the tunnel, and
its failure inf
o
rmatio
n gen
erally incl
ude
s po
wer o
u
tage
s, comm
unication failure
s, and d
a
ta
acq
u
isitio
n e
r
rors. Fi
re
ala
r
m i
s
respon
sible
for
dete
c
ting th
e fire
safety
con
d
itions i
n
sid
e
t
he
tunnel, an
d it
s failu
re info
rmation g
ene
rally in
clu
d
e
s
power outag
es, commu
ni
cation
fail
ure
s
,
high rate of false p
o
sitive,
and so fault
informat
ion.
Fans a
r
e ma
inly used to redu
ce the CO
con
c
e
n
tration
inside th
e tunnel an
d to improve its vi
sibility, and can let off den
se
smo
k
e wh
en
the tunnel fires, and it
s failure i
n
form
ation
gen
eral
ly include
s p
o
we
r outa
g
e
s
, com
m
uni
cation
failure
s, p
r
od
ucin
g
wind
d
r
op
and
so
o
n
[3]. Lig
h
ting in tu
nnel
mainly p
r
ovid
es li
ght, an
d
its
failure info
rm
ation ge
nerall
y
includ
es
po
wer out
a
g
e
s
, comm
uni
cati
on failu
re
s. Driveway lig
hts is
the traffic
si
gnal of tu
nn
el, and it
s f
a
ilure
in
form
ation ge
ne
ral
l
y includ
es
power
outag
es,
c
o
mmunic
a
tion failures
.
2.3. The Esta
blishment of Fault Tree o
f
the Moni
tor
i
ng Sy
stem
The fa
ult tree
of the i
n
telli
gent mo
nitori
ng
system
m
a
ke
s
analy
s
is by u
s
ing
do
wn-way
quantity anal
ysis, getting
the minimal
cut sets
(M
CS). It is: {veh
icle in
sp
ectio
n
device po
wer
failure}, {v
eh
icle ins
p
e
c
tion dev
ice li
ne fault}, {v
ehicl
e insp
e
c
tion dev
ic
e
--- King View
comm
uni
cati
on failu
re},
{
v
ehicle
in
spe
c
tion
dev
ice d
a
t
a a
n
o
ma
ly}
,
{
C
O
/
VI s
e
ns
or
po
w
e
r
failure},
{
C
O /
VI se
nsor lin
e fault
}, {
C
O
/ VI
sen
s
or --- Ki
ng Vie
w
comm
uni
cati
on failu
re
}, {
C
O /
VI data ano
maly}, {fire sensor po
we
r
failure}, {fire sen
s
o
r
line f
ault}, {fire se
nso
r
--- King
View
comm
uni
cati
on failure
}, {fire ala
r
m false high rate
},
{FS / FX sensor po
we
r sup
p
ly failure}, {
F
S
/
FX sen
s
o
r
lin
e fault}, {FS
/ FX sen
s
o
r
--- King
View
comm
uni
cati
on failure}, {
F
S / FX sen
s
or
data anom
aly}, {fan power failure}, {fan
line fault}
,{fans
--- King View co
mmu
nicatio
n
failure},
{blo
wer
no a
c
hieving the
d
e
sired effe
cts}, {lighting
po
wer fail
ure}, {l
ighting line
fa
ult}, {lighting
---
King View
co
mmuni
cation
failure}, {li
ghti
ng no
rea
c
hi
ng the exp
e
cted re
sults}, {
d
riveway ligh
t
s
pow
er failu
re}
,
{lane light line failure
}, {d
riveway light
s --- King Vie
w
commu
nication failure
}.
The fault tre
e
, establi
s
he
d acco
rdin
g to the inform
ation of the hard
w
a
r
e in
side the
tunnel, is sho
w
n a
s
Figu
re
1:
Figure 1. Fau
l
t Tree of the Monitori
ng System
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Fault Detection Mech
ani
sm
of Tunnel based on Arti
ficial Ne
ural
(Liu Liu)
7339
3. Fault De
te
ction of the
Tunnel Ba
se
d on the Neu
r
al Net
w
o
r
k
s
3.1. The Principle of De
te
cting Data F
a
ilures of Ne
ural Ne
t
w
o
r
k
s
With the d
e
velopme
n
t an
d re
sea
r
ch of
Neu
r
al
Networks
(NN), it
is in
cre
a
si
ngl
y being
use
d
in the field of engin
e
e
ring
cont
rol
s
, based o
n
its inhere
n
t advantage
s,
esp
e
cially its sy
stem
identificatio
n, adaptive
co
ntrol,
mod
e
li
ng, and
oth
e
r a
r
ea
s. It coul
d solve
these
proble
m
s
approp
riately, the un
ce
rtai
nty, severe
n
on-lin
ea
rity, hystere
s
is, tim
e
-varyin
g
con
t
rol of compl
e
x
system
s, mo
deling
and
te
sting i
s
sue
s
.
Different
envi
r
onm
ental p
a
r
amete
r
s in
si
de the
tunnel,
the
impact of th
e cont
rollabl
e device o
n
environ
ment
a
l
param
eters are mutual
rest
raint, mut
ual
influen
ce [4]. For exam
ple:
CO con
c
ent
ration insi
de t
he tunnel
at a time dep
en
ds o
n
the time of
the fan
s
runn
ing on, th
e n
u
mbe
r
of the
fans tu
rne
d
o
n
, the wi
nd
speed
and
direction i
n
tun
n
e
l,
traffic and visibility of tunnels. The
relat
i
onship,
only
con
s
id
erin
g the CO
co
nce
n
tration a
nd the
traffic, is a substa
ntially linear
relatio
n
s
hip. T
he hi
g
h
traffic, the greate
r
CO concentratio
n
. On
the contra
ry,
the ru
nnin
g
ti
me of th
e ai
r blo
w
er
,
th
e numbe
r of
fa
ns,
a
nd
th
e CO co
ncentration
are i
n
versely
pro
portio
nal
to sp
eed. A
c
cordi
ng to t
he cha
r
a
c
teri
stics that va
rious
pa
ramet
e
rs
insid
e
the tunnel a
r
e mut
ual impa
ct a
nd mutual re
straint a
nd a
powe
r
ful sel
f
-adaptio
n an
d
learni
ng abilit
y, discriminability of
the Neural
Netw
orks, it is a good way to solve the problem of
data a
nom
aly testing
in
the
tunnel
by
usi
ng the
Neu
r
al
Net
w
o
r
k to
e
n
su
re
the
pro
c
e
s
s of te
stin
g
data in tunn
el. In order
to monitor the t
unnel int
e
lligently, col
l
ecting a
nd
pro
c
e
ssi
ng the
abno
rmal d
a
ta in tunnel are essential.
Take tu
nnel
vehicle in
sp
e
c
tion controll
er a
s
an exa
m
ple to illust
rate ho
w BP Neu
r
al
Networks det
ect data
colle
cting a
nomal
y of the v
ehi
cle in
sp
ecto
r.
Vehicl
e in
sp
ection i
s
m
a
i
n
ly
use
d
to coll
e
c
t the numb
e
r
of vehicle
s
enterin
g the tunnel a
nd ve
hicle
s
’ real-ti
m
e sp
eed at
a
certai
n time. In a perio
d o
f
time, the numbe
r of
vehicle
s
ente
r
in
g the tunnel,
can affe
ct the
con
c
e
n
tration
of
CO i
n
si
de
the tu
nnel,
wind
dire
ctio
n, win
d
spe
e
d
an
d vi
sibility. While
the
CO
con
c
e
n
tration
mainly d
epe
nds on th
e n
u
m
ber of fan
s
open
ed in
the
tunnel
and
th
e ru
nnin
g
tim
e
,
and the vi
sibil
i
ty has a
relat
i
onship
with the num
be
r
of
lights a
nd th
e lighting i
n
te
nsity. Therefo
r
e,
input the
con
c
entration
of
CO i
n
si
de th
e tunn
el
（）
, wind
speed
（）
, visibility
（
）
, the num
b
e
r of the fa
ns run
n
ing
at a
certai
n time
（）
, the run
n
ing time
of the fan
s
（）
, the n
u
m
ber
of the l
i
ghts tu
rnin
g
on
（）
, and
th
e
light intensity
in tunnel
（（
）
) into the
Neural
Net
w
orks. Th
at
is to
say, the input
vector of
BP Neural Networks is{
、、
、
、
、
、
}.
Refers t
o
the
sample
. In ord
e
r to redu
ce
data
e
rro
rs in th
e in
put vecto
r
, ta
ke th
e
averag
e of several d
a
ta o
f
each ve
ctor in t
unnel a
s
the actual i
n
put of Neu
r
al
Netwo
r
ks in
se
con
d
s. Fo
r
example: in
seco
nd
s,
,
in whi
c
h the
is the num
ber
of th
e
data of th
e
CO con
c
ent
rati
on in
tunn
el,
colle
cted
by t
he
CO
se
nso
r
. How to d
e
termin
e the
Value
of
should ma
ke some ap
p
r
op
riate adju
s
tments,
base
d
on the freq
uen
cy of collecting the CO
con
c
e
n
tration
,
the actual m
agnitud
e
of CO co
nc
entrati
on and th
e experim
ent effect of the actu
al
sampl
e
s.
Take the av
erag
e of sa
mples in a p
e
riod of
time
as the Inpu
t of the exp
e
rime
ntal
sampl
e
s. Th
e
n
the inputs
of the variabl
e in this peri
od ca
n be co
nsid
ere
d
to be homog
ene
ous.
So no larg
e
erro
rs occu
r in the exp
e
rime
nt
al dat
a becau
se o
f
an abno
rm
al testing da
ta
,
guaranteei
ng the
accuracy of
dat
a inputted. Take
as the output of Neu
r
al Netwo
r
ks.
3.2. The Desi
gn of the
Ne
ural Ne
t
w
o
r
k
s
Struc
t
ure
The de
sig
n
o
f
the Netwo
r
ks structu
r
e p
r
imar
ily cove
rs the nu
mbe
r
of hidden l
a
yers, the
numbe
r of n
euro
n
s i
n
ea
ch laye
r. The
o
ry proof tha
t
Netwo
r
ks,
with bia
s
, at least o
ne S-t
y
pe
hidde
n
laye
r and a
lin
ear output
laye
r, can
ap
prox
im
ate any
ratio
n
a
l fun
c
tion. B
u
t so
me
ran
d
o
m
uncertaintie
s
exist in the
control
syste
m
of t
he a
c
tual
tunnel
(Eg:
weath
e
r, p
e
o
p
le, environm
ent,
etc.). That is
to say, the model of tunn
el
inte
lligent m
onitorin
g
sy
stem is
an un
certain stocha
stic
system. The
r
efore, when t
r
ainin
g
the Networks by
u
s
ing o
n
ly one
hidden laye
r,
there is a
ne
ed
to increa
se
the nu
mbe
r
o
f
neuron
s in
this laye
r, on
ly in this
wa
y can th
e e
r
ror p
r
e
c
isi
on
b
e
()
co
t
()
f
st
()
vi
t
_(
)
f
j
c
ou
nt
t
_(
)
f
jt
i
m
e
t
_(
)
z
m
c
o
unt
t
()
g
qt
()
co
t
()
f
st
()
vi
t
_(
)
f
j
c
ou
nt
t
_(
)
f
j
tim
e
t
_(
)
z
m
c
o
unt
t
()
g
qt
tt
a
a
1
()
()
s
n
n
co
t
c
o
t
s
s
a
()
cl
l
t
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 733
7
– 7342
7340
improve
d
. But Actual Op
eration sho
w
s t
hat even
a
ddi
ng a lot n
e
u
r
ons i
n
to the
Neu
r
al
Netwo
r
ks
with only o
n
e
hidd
en lay
e
r, the requi
red preci
s
io
n
is still h
a
rdly
rea
c
he
d. Th
erefo
r
e, Neural
Networks t
r
ai
ning ne
ed
s to
sele
ct more than on
e hid
d
en layers, but
too much hid
den laye
rs
will
make
the
Neural
Netwo
r
ks be
com
e
v
e
ry compli
cat
ed. That
will
increa
se
bu
rden
on
Neu
r
al
Networks co
ntrol a
nd the
trainin
g
time
of t
he Neural Net
w
o
r
ks [
5
]. The data
abno
rmal
Ne
ural
Networks of
the intellige
n
t monito
ring
system i
n
th
e tunn
el will
use two
hid
den laye
rs a
fter
overall con
s
id
eration.
Acco
rdi
ng to
the con
c
ent
ration
of CO
（）
, win
d
speed
（）
, visibility
（
）
, the numb
e
r of the fans run
n
ing
（）
, the runni
ng time of the fans
（
）
, the numb
e
r of the ligh
t
s turnin
g on
（）
, and th
e light inten
s
i
t
y
（（
）
in a pe
riod of time, the traffic flo
w
in this pe
riod
of time can b
e
inferred. Th
erefo
r
e,
the numb
e
r o
f
neuro
n
s
of the input laye
r in the Ne
ural
Networks i
s
7, t
he numb
e
r
of neu
ron
s
of
the outp
u
t la
yer in
the
Neural
Netwo
r
ks is 1.
T
h
e
hidd
en l
a
yer of the
Net
w
orks structu
r
e is
prefe
r
ably 1
6
22 n
ode
s,
whi
c
h i
s
p
r
o
v
ed thro
ugh
trainin
g
diff
erent
neu
ral
sam
p
le
s an
d
comp
ari
ng th
e experim
ent
al data. As is
sho
w
n in Fig
u
re 2:
Figure 2. The
Neural Net
w
orks Mo
del of
the V
ehicle Insp
ectio
n
for Testing
Data
Erro
rs
3.3. The Selection of the
Activ
a
tion Function
Acco
rdi
ng to
the pa
rticul
ari
t
y and rand
o
m
un
ce
rtainti
e
s of
the Ne
ural Net
w
orks
mo
del,
the sele
ction
of the
sigm
oi
d fun
c
tion i
s
a relati
vely
g
ood ch
oice. The sigm
oid function
ha
s the
cha
r
a
c
teri
stics, smooth
n
e
s
s a
nd
rob
u
st
ness. O
n
th
e
othe
r h
and,
the de
rivative of the fu
ncti
on
can
be expressed by
a certai
n
ex
p
r
e
ssi
on of th
e
function
[6]. At
this
rate,
when us
ing the
function fo
r the actu
al cal
c
ulatio
n, the con
s
um
pt
ion
of a com
pute
r
syste
m
, in the pro
c
e
s
s
of
comp
uting
a
nd
storage,
can
be
effe
ctively so
lve
d
. The
n
u
s
i
ng the
sigm
oid fun
c
tion
can
signifi
cantly improve the
converg
e
n
c
e speed of the Networks in the
stru
ctural model of the sa
me
Networks, wh
ich is very im
portant to so
me more
co
mplex
Neu
r
al
Netwo
r
k
s
. In order to s
i
mplify
netwo
rk d
e
si
gn, the linear transfe
r
function (Pureli
n
) is use
d
into
the output la
yer, so that the
final output of the network
can b
e
any value.
3.4. The Selection of Sam
p
le Data Ba
s
e
d on the F
a
ult De
tec
t
ion
Model
The tun
nel d
e
tection
and
cont
rol i
s
a
very compl
e
x pro
c
e
s
s.
The am
ount
of data
c
o
llec
t
e
d
is ve
r
y
lar
g
e an
d th
e
s
e
le
c
t
ion o
f
th
e
sampl
e
in
the Ne
ural Netwo
r
ks
occupi
es a very
importa
nt po
sition. T
hat, wh
ether th
e data
of
t
he sampl
e
vectors enjo
y
s
po
pula
r
ity
and
rep
r
e
s
entatio
n or n
o
t, largely affect the learni
ng of
the Neu
r
al
Networks. It is an extrem
ely
()
co
t
()
f
st
()
vi
t
_(
)
f
j
c
ount
t
_(
)
f
jt
i
m
e
t
_(
)
z
m
c
o
unt
t
()
g
qt
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Fault Detection Mech
ani
sm
of Tunnel based on Arti
ficial Ne
ural
(Liu Liu)
7341
tediou
s wo
rk to sele
ct an app
rop
r
iat
e
trai
nin
g
sample
s, fro
m
massive
amount
s of data
gene
rated
by
in th
e tun
n
e
l
monito
ring
system. A
n
d
in th
e p
r
o
c
e
s
s of
gathe
ri
ng info
rmatio
n,
some
errors
may occur, e
v
en to be larg
e errors re
lati
ve to the real
data may be
erro
neo
us
d
a
ta
.
Removal
of
such
data i
s
t
he imp
o
rtant
part of th
e
work of the
sa
mple. Th
e tu
nnel m
onitori
ng
data ha
s the followin
g
ch
aracteri
stics:
1) Hu
ge am
o
unt of data
2) The
same
data ch
ang
ed
in relatively small extent
3) The d
a
ta may be wron
g data in a pe
riod of time
4) The d
a
ta h
a
s a big redu
ndan
cy
5) Detectin
g
data in a
ce
rtain ran
ge
(for
exam
ple: the CO con
c
e
n
tration h
a
s
a ran
g
e
and the teste
d
data are g
e
nerally not ou
tside this
ran
ge)
6) According
to the actual operating con
d
ition
s
of the tunnel, there is a wi
de gap
betwe
en traffi
c flows of different p
e
rio
d
s in a
day. (Eg: 7:00 to 9:00 (in the m
o
rning
)
, 11:00
to
14:00 (at no
o
n
) and 1
7
:00
to 21:00 (in ni
ght) have la
rg
e traffic flow)
Acco
rdi
ng to
the above f
eature
s
, the
sele
cti
on of t
he traini
ng sample of the
Neu
r
al
Networks
should follow the
following principles:
1) Trai
ning
sample
s is ra
ndomly sel
e
cted:
for a huge amou
nt of monitoring
data, if
taking
all
the
monitori
ng
da
ta as the
trai
ning
sa
mple
s, that will
g
r
e
a
tly increa
se
the trai
ning
time
of the neural
netwo
rk a
nd i
t
is not nece
s
sary to
collect
such sa
mple
s in the traini
ng pro
c
e
s
s.
2) The
stand
ardi
zation of trainin
g
sa
mpl
e
s:
Since the
tunnel monito
ring data i
s
chang
ed
in a certai
n range, there is a need to st
anda
rdi
z
e
so
me sam
p
le d
a
ta to redu
ce
the erro
r cau
s
ed
by the sele
ction of the erro
r data.
3) Acco
rdin
g
to the testing data in t
unnel
with the characte
ri
st
ics of the
smalle
r
amplitude of
variation, so
me good trai
ning data i
s
better to take more b
e
fore and after t
he
sampl
e
, whi
c
h is ba
sed o
n
the neare
s
t n
e
ighb
or rule.
4) Sel
e
ct th
e
sa
mple
s
of the d
e
tecte
d
data
wi
thin
a
pe
riod
of tim
e
sele
cted, a
nd the
n
obtain the av
erag
e value to redu
ce the
error cau
s
ed
by the erro
r d
a
ta to network trainin
g
.
5) Sele
ctively sele
ct the sample in eve
r
y per
io
d of time of every
day, to incre
a
se th
e
sele
ction
cov
e
rag
e
of the sample an
d im
prove the
qu
a
lity of
training
of the Neural
Netwo
r
ks [7].
3.5. The Mat
Lab Simulation Experime
nt
The histo
r
i
c
al
data
of t
he a
c
tual
tra
ffic, colle
cted
in
Wu
han
Shuigu
ohu
tun
nel in
a
month, ma
ke
s the
Ne
ural
Networ
ks
do self-a
daptive training.
M
a
ke
the Neu
r
al Networks
do self-
adaptive trai
n
i
ng, usin
g a l
o
t of the histo
r
ical
data,
the
model of trai
ning the n
eural netwo
rk is
as
sho
w
n in 3:
Figure 3. The
Structure of
the Ne
ural
Ne
tworks
The si
mulati
on of the sy
stem takes
Wuh
an Shui
guoh
u Tun
n
e
l
as research obje
c
ts,
Wuh
an Sh
ui
guoh
u tunn
el
is a
bi
dire
ctional an
d fo
ur la
ne tu
nn
el, with a
len
g
th of ove
r
1
700
meters. The t
unnel
can
be
divided into t
h
ree
se
ction
s
. There i
s
so
me ope
n spa
c
e left bet
we
en
each sectio
n, whi
c
h i
s
effectively reg
u
l
a
ting
the tun
nel environm
ent. The ma
x spee
d in t
h
e
tunnel i
s
8
0
KM / H. Th
e n
u
mbe
r
of the
vehicle
s
th
ro
ugh the
tunn
el is
ro
ughly
35 eve
r
y min
u
te,
and
rou
ghly
75 p
e
r mi
nut
e at p
eak tim
e
s. Ba
se
d
on
the a
c
tual
condition
of Wuhan
Shuig
u
ohu
tunnel, the n
eural
network training b
e
lo
w ha
s be
en
d
one, u
s
ing a l
o
t of the actu
al data throu
gh
the tunnel op
erating hi
story. T
hen use the Neu
r
al Ne
tworks fini
shi
ng to do the testing of data
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 733
7
– 7342
7342
anomaly of th
e traffic flo
w
colle
cted
by the vehi
cl
e in
spe
c
tion. Afte
r a la
rge
num
ber of hi
sto
r
ical
data a
nd
rep
eated trainin
g
of the
Net
w
o
r
ks, p
r
edi
ct
th
e traffic flow
of the tun
nel
and fin
a
lly draw
a set of curv
es for th
e co
mpari
s
o
n
bet
wee
n
t
he a
c
tual value a
n
d
the predi
cted value of t
h
e
traffic flow, as is sho
w
n in F
i
gure 4:
Figure 4. The
Compa
r
i
s
on
Cha
r
t of Traffic Forecast
s
In the ab
ove
figure, th
e
solid line
re
prese
n
ts
th
e a
c
tual traffic fl
ow
of the
sa
mple in
a
minute, whil
e
the dotted lin
e rep
r
e
s
e
n
ts
the pre
d
icte
d
traffic flow
of the Ne
ural
Networks, train
ed
by the seven
input ve
ctors u
s
ing
the
same
samp
l
e
, the CO
con
c
entration, lig
ht intensity, the
time of the fa
n ope
ned, fa
ns’
numb
e
r,
visibility,
win
d
speed, li
ght
ing nu
mbe
r
i
n
the tun
nel.
Not
difficult to find
, from the
co
mpari
s
o
n
cha
r
t, that the de
gree
of fitting
of the solid li
ne is better th
an
the da
sh
ed li
ne. It ca
n
ge
nerally
be
co
nsid
ere
d
that
v the m
odel
of data
ano
m
a
ly dete
c
tion
for
vehicle in
spe
c
tion devi
c
e can pro
p
e
r
ly predi
ct traffic flow.
4. The Summar
y
The
simul
a
tio
n
results
sh
o
w
that th
e
co
mbinat
ion
of t
he fault t
r
ee
analysi
s
and
the fault
detectio
n
me
cha
n
ism
of artificial Neu
r
al
Networks
ca
n
prop
erly pre
d
ict accid
ents in the tunnel
to
occur. T
h
e
r
e
is no
dou
bt that the an
alysis
ha
s a
bro
ad a
ppli
c
ation
pro
s
p
e
ct in the
safety
manag
eme
n
t of the tunnel.
Referen
ces
[1]
Yin Hu
anfe
ng.
Researc
h
an
d app
lic
ation
of the V
entil
ating C
ontrol te
chno
log
y
intun
nels. W
uha
n
:
Master'
s
degre
e
thesis of W
uhan U
n
iv
ers
i
t
y
of
technol
og
y. 201
2;
6.
[2]
Z
hang T
i
anze
n
g
. Researc
h
an
d Desi
gn for Automatic Edg
e
S
y
stem Bas
ed
on Ne
ural Mo
d
e
l. Don
gbe
i:
Master'
s
degre
e
thesis of Do
n
g
Bei Un
iversit
y
. 20
08.
[3]
Lia
ng Z
he. Res
earch for Veh
i
c
l
e S
y
st
em of Elec
tromag
netic
Comp
atibi
lit
y
B
a
sed o
n
F
ault
T
r
ee Model.
Xi A
n
: Master'
s
degre
e
thesis
of Xi'
a
n El
ectro
n
ic an
d Eng
i
ne
erin
g Univ
ersit
y
. 20
12.
[4]
Ma Che
n
g
q
ia
n
,
Ren g
u
ish
an.
Intellig
ent C
o
ntrol S
y
stem
for Urban
Ro
a
d
T
unnel Bas
e
d on N
eur
al
Mode
l.
Co
mput
er & Digita
l
En
gin
eeri
n
g
. 20
0
8
: 46(2): 8-14.
[5]
Pham DT
, Liu
X. T
r
aini
ngof E
l
man
net
w
o
rks and
d
y
n
a
mo
ic s
y
stem
m
o
d
e
li
ng.
Intern
atio
n
gal J
our
nal
of
System
s Science
. 1996.
[6]
F Liu, GS Ng.
Artificial Venti
l
atio
n Mode
lin
g
usin
g Neur
o
-
F
u
zz
y
Hybr
id
System
. International Joint
Confer
ence
on
Neura
l
Net
w
or
ks. 2006.
[7]
Liu Da
w
e
i. C
o
ntrol strateg
y
of
T
unnel Mo
nito
ri
ng S
y
ste
m
. Shang Ha
i
:
Master’s deg
ree thesis of
shan
gh
ai mariti
me univ
e
rsit
y
.
200
5; 9.
Evaluation Warning : The document was created with Spire.PDF for Python.