Co
m
pu
t
er
Science
a
nd
I
nfo
r
m
a
t
io
n T
ec
hn
o
lo
g
ies
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
,
p
p
.
2
5
6
~
2
7
0
I
SS
N:
2722
-
3
2
2
1
,
DOI
:
1
0
.
1
1
5
9
1
/csi
t
.
v
7
i
3
.
p
2
5
6
-
2
7
0
256
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
p
r
ime.
co
m/in
d
ex
.
p
h
p
/csi
t
C
o
nv
o
lutiona
l neural network
a
nd
lo
ng
sho
rt
-
t
erm
memo
ry
foreca
sting
and
v
a
ria
tiona
l auto
enco
der
a
no
ma
ly
det
ection i
n
4
G
cellular
net
wo
rks
Ruv
a
ra
s
he
C.
H
o
v
e
1
,
E
ng
M
a
info
rd
M
uta
nd
a
v
a
ri
2
1
D
e
p
a
r
t
me
n
t
o
f
D
a
t
a
S
c
i
e
n
c
e
a
n
d
A
n
a
l
y
t
i
c
s,
S
c
h
o
o
l
o
f
I
n
f
o
r
ma
t
i
o
n
S
c
i
e
n
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
H
a
r
a
r
e
I
n
st
i
t
u
t
e
o
f
T
e
c
h
n
o
l
o
g
y
,
H
a
r
a
r
e
,
Zi
m
b
a
b
w
e
2
D
e
p
a
r
t
me
n
t
o
f
D
a
t
a
S
c
i
e
n
c
e
a
n
d
B
u
s
i
n
e
ss S
y
st
e
ms,
S
R
M
I
n
st
i
t
u
t
e
o
f
S
c
i
e
n
c
e
a
n
d
T
e
c
h
n
o
l
o
g
y
,
K
a
t
t
a
n
k
u
l
a
t
h
u
r
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
May
5
,
2
0
2
6
R
ev
is
ed
J
u
n
12
,
2
0
2
6
Acc
ep
ted
J
u
l
13
,
2
0
2
6
S
p
e
c
tru
m
m
o
n
i
to
ri
n
g
in
c
e
ll
u
lar
n
e
two
rk
s
wit
h
li
m
it
e
d
re
so
u
rc
e
s
re
m
a
in
s
p
re
d
o
m
i
n
a
n
tl
y
re
a
c
ti
v
e
b
e
c
a
u
se
c
a
p
a
c
it
y
p
la
n
n
i
n
g
a
n
d
fa
i
lu
re
d
e
tec
ti
o
n
o
c
c
u
r
a
fter
n
e
two
rk
c
o
n
g
e
stio
n
,
wh
il
e
sp
e
c
tru
m
m
o
n
it
o
ri
n
g
d
a
ta
a
re
ra
re
ly
a
v
a
il
a
b
le
fo
r
re
p
r
o
d
u
c
ib
le
re
se
a
rc
h
.
Ex
isti
n
g
st
u
d
ies
a
d
d
re
ss
sh
o
rt
-
term
traffic
fo
re
c
a
stin
g
a
n
d
a
n
o
m
a
ly
d
e
tec
ti
o
n
a
s
se
p
a
ra
te
tas
k
s
a
n
d
larg
e
l
y
d
e
p
e
n
d
o
n
d
e
n
se
ly
sa
m
p
led
o
p
e
r
a
to
r
tele
m
e
try
d
a
ta
th
a
t
a
re
in
a
c
c
e
ss
ib
le
to
a
c
a
d
e
m
ia
a
n
d
re
g
u
lato
rs
in
e
m
e
rg
in
g
m
a
rk
e
ts.
Th
is
stu
d
y
in
teg
r
a
tes
b
o
t
h
tas
k
s
in
t
o
a
u
n
ifi
e
d
d
e
e
p
-
lea
rn
i
n
g
p
ip
e
li
n
e
fo
r
a
4
G
c
e
ll
u
lar
e
n
v
iro
n
m
e
n
t.
Th
e
Zi
m
b
a
b
we
s
p
e
c
tru
m
d
a
tas
e
t
c
o
n
tai
n
s
3
1
5
,
2
4
7
h
o
u
rl
y
m
e
a
su
re
m
e
n
ts
c
o
ll
e
c
ted
fro
m
1
3
c
e
ll
sites
,
t
h
re
e
o
p
e
ra
to
rs,
a
n
d
fiv
e
fre
q
u
e
n
c
y
b
a
n
d
s,
a
n
n
o
tate
d
with
fo
u
r
o
p
e
ra
to
r
-
d
e
fi
n
e
d
a
n
o
m
a
ly
c
las
se
s
re
p
re
se
n
ti
n
g
2
.
0
3
%
o
f
a
ll
m
e
a
su
re
m
e
n
ts.
A
h
y
b
rid
o
n
e
-
d
ime
n
sio
n
a
l
(1
-
D)
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
-
l
o
n
g
sh
o
r
t
-
term
m
e
m
o
ry
(CNN
–
LS
TM
)
m
o
d
e
l
u
se
s
a
7
2
-
h
o
u
r
traffic
win
d
o
w
t
o
fo
re
c
a
st
th
e
n
e
x
t
six
h
o
u
rs
o
f
a
g
g
re
g
a
te
n
e
two
rk
traffic
,
wh
il
e
a
v
a
riatio
n
a
l
a
u
t
o
e
n
c
o
d
e
r
(VA
E)
train
e
d
e
x
c
l
u
siv
e
l
y
o
n
n
o
rm
a
l
re
c
o
rd
s
d
e
tec
ts
a
n
o
m
a
li
e
s
wh
e
n
re
c
o
n
str
u
c
ti
o
n
e
rro
r
e
x
c
e
e
d
s
th
e
9
9
th
-
p
e
rc
e
n
ti
le
v
a
li
d
a
ti
o
n
th
re
sh
o
l
d
.
O
n
th
e
h
e
ld
-
o
u
t
tes
t
se
t,
th
e
m
o
d
e
l
a
c
h
iev
e
d
a
m
e
a
n
a
b
so
lu
te
e
rro
r
(
M
AE)
o
f
1
5
8
.
9
7
G
B,
ro
o
t
m
e
a
n
sq
u
a
r
e
e
rro
r
(RM
S
E)
o
f
2
3
0
.
4
1
G
B
,
a
n
d
m
e
a
n
a
b
s
o
l
u
t
e
p
e
r
c
e
n
t
a
g
e
e
r
r
o
r
(
M
A
P
E
)
o
f
5
2
.
9
3
%
,
o
u
t
p
e
r
f
o
r
m
i
n
g
a
s
e
a
s
o
n
a
l
-
n
a
i
v
e
b
a
s
e
l
i
n
e
(
M
A
E
=
2
1
6
.
4
3
G
B
,
M
A
P
E
=
6
6
.
7
7
%
,
p
<
0
.
0
0
1
,
D
i
e
b
o
l
d
-
M
a
r
i
a
n
o
t
e
s
t
)
.
T
h
e
s
t
u
d
y
c
o
n
t
r
i
b
u
t
e
s
a
p
u
b
l
i
c
l
y
a
v
a
i
l
a
b
l
e
l
o
c
a
l
i
z
e
d
s
y
n
t
h
e
t
i
c
d
a
t
a
s
e
t
,
a
re
p
r
o
d
u
c
i
b
l
e
e
n
d
-
to
-
e
n
d
f
o
r
e
c
a
s
t
i
n
g
a
n
d
a
n
o
m
a
l
y
d
e
t
e
c
t
i
o
n
b
a
s
e
l
i
n
e
,
a
n
d
a
p
e
r
-
c
la
s
s
a
n
o
m
a
l
y
a
n
a
l
y
s
i
s
.
K
ey
w
o
r
d
s
:
An
o
m
aly
d
etec
tio
n
C
ellu
lar
tr
af
f
ic
f
o
r
ec
asti
n
g
C
NN
–
L
STM
R
ep
r
o
d
u
cib
le
r
esear
c
h
R
F sp
ec
tr
u
m
m
an
ag
em
en
t
Sy
n
th
etic
d
ataset
Var
iatio
n
al
au
to
en
co
d
er
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
R
u
v
ar
ash
e
C
.
Ho
v
e
Dep
ar
tm
en
t o
f
Data
Scien
ce
an
d
An
aly
tics
,
Sch
o
o
l
o
f
I
n
f
o
r
m
atio
n
Scien
ce
an
d
T
ec
h
n
o
lo
g
y
Har
ar
e
I
n
s
titu
te
o
f
T
ec
h
n
o
l
o
g
y
P.O.
B
o
x
B
E
2
7
7
,
B
elv
ed
er
e
,
Har
ar
e,
Z
im
b
ab
we
E
m
ail: h
2
4
0
6
1
8
f
@
h
it.a
c.
zw
1.
I
NT
RO
D
UCT
I
O
N
R
ad
io
f
r
eq
u
en
cy
(
R
F)
s
p
ec
tr
u
m
is
a
f
in
ite
r
eso
u
r
ce
,
co
n
tr
o
lled
b
y
ju
r
is
d
ictio
n
a
n
d
t
h
e
u
tili
za
tio
n
ef
f
icien
cy
h
as
a
d
ir
ec
t
r
elatio
n
s
h
ip
with
th
e
s
er
v
ice
co
s
t,
q
u
a
lity
an
d
co
v
er
ag
e
a
r
ea
o
f
ce
llu
lar
s
er
v
ice.
Mo
b
ile
d
ata
tr
af
f
ic
h
as
in
c
r
ea
s
ed
s
ig
n
if
ican
tly
o
v
er
th
e
last
d
ec
ad
e,
f
u
eled
b
y
th
e
p
e
n
etr
atio
n
o
f
4
G
h
a
n
d
s
ets,
m
o
r
e
b
a
n
d
wid
th
-
h
u
n
g
r
y
ap
p
li
ca
tio
n
s
,
an
d
t
h
e
s
h
if
t
o
f
v
o
ice
tr
af
f
ic
to
d
ata
-
b
ea
r
in
g
s
er
v
ic
es,
ac
r
o
s
s
em
er
g
in
g
m
ar
k
ets.
I
n
Z
im
b
ab
we,
th
e
Z
im
b
ab
we
Natio
n
al
Fre
q
u
en
cy
Allo
ca
tio
n
Plan
(
Z
NFAP)
[
1
]
allo
ca
tes
th
e
s
p
ec
tr
u
m
an
d
i
n
th
e
Z
im
b
a
b
w
ea
n
m
o
b
ile
s
ec
to
r
,
m
o
b
ile
d
at
a
tr
af
f
ic
h
as
b
ee
n
s
tead
ily
g
r
o
win
g
o
v
er
t
h
e
last
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
C
o
n
vo
lu
tio
n
a
l
n
eu
r
a
l n
etw
o
r
k
a
n
d
lo
n
g
s
h
o
r
t
-
term me
mo
r
y
f
o
r
ec
a
s
tin
g
a
n
d
…
(
R
u
v
a
r
a
s
h
e
C
.
Ho
ve
)
257
f
ew
y
ea
r
s
in
ter
m
s
o
f
p
er
ce
n
ta
g
e
an
d
v
o
lu
m
e
as
d
o
c
u
m
en
ted
in
th
e
ab
r
i
d
g
ed
Po
s
tal
an
d
T
e
leco
m
m
u
n
icatio
n
s
Secto
r
Per
f
o
r
m
an
ce
R
ep
o
r
t
(
P
T
SP
R
s
)
[
2
]
an
d
it
h
as
b
ee
n
n
o
ted
th
at
a
g
r
o
win
g
p
er
ce
n
tag
e
o
f
th
is
tr
af
f
ic
is
o
n
th
e
lo
wer
e
n
d
o
f
t
h
e
s
p
ec
tr
u
m
b
an
d
s
a
n
d
th
at
th
e
q
u
ality
o
f
s
er
v
ice
(
Qo
S)
h
as
b
ee
n
u
n
d
er
s
tr
ain
o
n
a
r
eg
u
lar
b
asis
o
n
th
e
u
r
b
an
ce
ll.
Fix
ed
-
allo
ca
tio
n
r
eg
im
es
p
r
o
v
i
d
e
lit
tle
in
f
o
r
m
atio
n
to
o
p
er
ato
r
s
a
n
d
r
eg
u
lato
r
s
ab
o
u
t
em
er
g
in
g
c
o
n
g
esti
o
n
,
in
te
r
f
er
en
ce
an
d
u
n
au
th
o
r
ized
tr
an
s
m
is
s
io
n
s
u
n
til
it
af
f
ec
ts
s
u
b
s
cr
ib
er
s
an
d
cu
r
ate
d
o
p
er
ato
r
tr
ac
es th
at
ca
n
b
e
m
a
d
e
av
ailab
le
f
o
r
ac
ad
em
ic
s
tu
d
y
ar
e
less
co
m
m
o
n
.
T
h
er
e
ar
e
th
en
two
d
is
tin
ct
b
u
t
in
ter
d
ep
en
d
en
t o
p
e
r
atio
n
al
ta
s
k
s
wh
ich
ar
e
v
er
y
im
p
o
r
tan
t b
o
th
to
th
e
r
eg
u
lato
r
a
n
d
to
th
e
o
p
er
ato
r
.
T
h
e
f
ir
s
t
is
s
h
o
r
t
-
h
o
r
izo
n
f
o
r
ec
asti
n
g
o
f
n
etwo
r
k
t
r
af
f
ic,
to
b
ase
d
ec
is
io
n
s
o
n
ca
p
ac
ity
p
lan
n
in
g
,
s
ch
ed
u
lin
g
an
d
lo
ad
-
b
alan
cin
g
o
n
a
q
u
a
n
titativ
e
lo
ad
esti
m
ate
in
s
tead
o
f
th
e
o
p
er
ato
r
'
s
in
tu
itio
n
[
3
]
–
[
5
]
.
T
h
e
s
ec
o
n
d
is
an
o
m
aly
d
etec
tio
n
o
f
t
h
e
s
am
e
telem
etr
y
with
o
u
t
an
y
lab
elled
tr
ain
in
g
d
ata
,
s
u
ch
th
at
in
ter
f
er
en
ce
,
tr
af
f
ic
s
u
r
g
es,
s
ig
n
al
d
r
o
p
o
u
ts
a
n
d
u
n
a
u
th
o
r
ized
em
is
s
io
n
s
ca
n
b
e
id
en
tifie
d
f
o
r
r
e
v
iew
[
6
]
–
[
9
]
.
T
h
ese
two
task
s
ar
e
u
s
u
ally
d
o
n
e
b
y
two
d
if
f
er
en
t
p
ip
elin
es
n
o
wad
ay
s
,
th
ey
g
iv
e
p
o
in
t
f
o
r
ec
asts
o
r
a
d
etec
tio
n
alg
o
r
ith
m
aler
ts
b
u
t
d
o
es
n
o
t
e
x
p
lain
wh
y
r
ea
lit
y
d
o
es
n
o
t
m
atch
p
r
e
d
ictio
n
an
d
d
o
es
n
o
t
g
iv
e
an
y
an
ticip
ato
r
y
lo
a
d
s
ig
n
al.
I
n
em
er
g
in
g
m
ar
k
ets,
with
lim
i
ted
o
n
-
ca
ll
e
n
g
in
ee
r
in
g
ca
p
a
city
,
o
p
er
at
o
r
s
ar
e
esp
ec
ially
in
n
ee
d
o
f
a
u
n
if
ied
v
iew
to
p
r
esen
t
th
e
lear
n
t
n
o
r
m
al
r
eg
im
e
as
well
as
d
e
v
iatio
n
f
r
o
m
t
h
a
t
r
e
g
im
e
in
a
s
in
g
le
d
ec
is
io
n
lo
o
p
.
T
h
r
ee
g
a
p
s
o
f
co
n
cr
ete
r
em
a
in
.
First,
th
e
h
y
b
r
id
d
ee
p
f
o
r
ec
asti
n
g
ar
ch
itectu
r
es
p
r
esen
t
ed
in
th
e
wir
eless
-
n
etwo
r
k
in
g
liter
atu
r
e
[
4
]
,
[
5
]
,
[
1
0
]
,
[
1
1
]
h
av
e
p
r
ed
o
m
in
an
tly
b
ee
n
ass
ess
ed
o
n
d
e
n
s
e
o
p
e
r
ato
r
-
in
ter
n
al
tr
ac
es
wh
er
e
th
e
y
ar
e
u
n
av
ailab
le
to
t
h
e
g
en
e
r
al
r
ese
ar
ch
co
m
m
u
n
ity
,
m
ak
in
g
it
d
if
f
icu
lt
to
r
ep
r
o
d
u
ce
ev
en
s
o
lid
id
ea
s
.
Seco
n
d
,
r
ec
o
n
s
tr
u
ctio
n
-
b
ased
an
o
m
aly
d
et
ec
to
r
s
,
[
6
]
–
[
9
]
,
[
1
2
]
–
[
1
6
]
ten
d
to
b
e
ev
alu
ated
o
n
b
en
ch
m
ar
k
s
th
at
lu
m
p
d
e
tecto
r
p
er
f
o
r
m
an
ce
to
g
eth
e
r
with
d
ataset
id
io
s
y
n
cr
asy
,
an
d
v
er
y
f
ew
s
tu
d
ies
r
ep
o
r
t
r
ec
all
p
er
an
o
m
aly
class
,
th
e
m
ea
s
u
r
e
o
f
in
ter
est
wh
en
tr
iag
in
g
o
p
e
r
atio
n
s
.
T
h
ir
d
,
d
ata,
in
f
r
astru
ctu
r
e
an
d
p
o
licy
ar
e
cited
co
n
s
is
ten
tly
as
th
e
“f
iv
e
en
f
o
r
ce
r
s
”
o
f
Af
r
ican
an
d
o
th
er
em
er
g
i
n
g
m
ar
k
et
d
ep
lo
y
m
e
n
ts
in
ar
tific
ial
in
tellig
en
ce
(
AI
)
-
f
o
r
-
s
p
ec
tr
u
m
-
m
a
n
ag
em
en
t
s
u
r
v
e
y
s
[
1
7
]
–
[
2
0
]
,
with
v
er
y
f
ew
p
u
b
lis
h
ed
p
i
p
elin
es
ca
lib
r
a
ted
ag
ain
s
t
p
u
b
licly
v
e
r
if
iab
le
n
atio
n
al
-
le
v
el
s
tatis
tic
s
an
d
r
elea
s
ed
en
d
to
en
d
.
T
h
is
p
ap
er
d
i
r
ec
tly
tack
les th
o
s
e
th
r
ee
g
ap
s
.
I
n
th
is
p
ap
er
,
a
d
ee
p
lear
n
in
g
b
ased
two
-
c
o
m
p
o
n
en
t
p
ip
elin
e
is
d
ev
elo
p
e
d
an
d
ev
alu
ated
.
T
h
e
f
o
r
ec
aster
m
o
d
el
u
s
ed
is
a
h
y
b
r
id
m
o
d
el
o
f
one
-
d
im
en
s
io
n
al
(
1
-
D
)
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN
)
an
d
lo
n
g
s
h
o
r
t
-
ter
m
m
em
o
r
y
(
L
STM
)
wh
ich
f
o
r
ec
asts
tr
af
f
ic
f
o
r
th
e
n
ex
t
6
h
o
u
r
s
g
iv
en
7
2
h
o
u
r
s
s
lid
in
g
win
d
o
w
in
p
u
t.
T
h
e
tr
ain
ed
v
ar
iatio
n
al
au
to
en
c
o
d
er
(
VAE
)
is
o
n
ly
tr
ain
ed
o
n
n
o
r
m
al
r
ec
o
r
d
s
an
d
th
e
n
d
etec
ts
th
e
an
o
m
alies
u
s
in
g
th
e
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
,
u
s
in
g
th
e
9
9
th
p
er
ce
n
tile
o
f
th
e
v
alid
atio
n
r
e
co
n
s
tr
u
ctio
n
e
r
r
o
r
s
as
a
f
ix
ed
d
ec
is
io
n
th
r
esh
o
l
d
.
B
o
th
m
o
d
els
ar
e
ass
e
s
s
ed
a
g
ain
s
t
th
e
Z
im
b
ab
we
s
p
ec
tr
u
m
d
ataset
wh
ich
is
a
s
y
n
th
etic
d
ataset
d
ev
elo
p
ed
with
a
f
o
o
tp
r
in
t
r
ef
lectin
g
th
e
1
3
-
to
wer
,
5
-
b
an
d
,
3
-
o
p
er
ato
r
s
p
ec
tr
u
m
in
Z
im
b
ab
we
an
d
ca
lib
r
ate
d
u
s
in
g
th
e
p
u
b
licly
v
er
if
iab
le
Po
s
tal
an
d
T
elec
o
m
m
u
n
ic
atio
n
s
R
eg
u
lato
r
y
Au
t
h
o
r
ity
o
f
Z
im
b
ab
we
(
POTRAZ)
s
ec
to
r
s
tatis
tics
.
I
t
is
an
en
d
-
to
-
en
d
r
e
p
r
o
d
u
cib
le
b
aselin
e,
n
o
t
a
d
etec
to
r
s
u
itab
le
f
o
r
d
ep
lo
y
m
e
n
t
-
th
is
ca
n
d
o
r
is
a
k
e
y
asp
ec
t o
f
th
e
co
n
tr
ib
u
tio
n
.
T
wo
ch
ar
ac
ter
is
tics
ar
e
k
ey
to
u
n
d
e
r
s
tan
d
in
g
th
e
n
o
v
elty
v
alu
e
o
f
th
is
wo
r
k
,
r
elat
iv
e
to
th
e
C
NN
–
L
STM
f
o
r
ec
aster
s
[
5
]
,
[
1
1
]
an
d
th
e
p
r
ev
io
u
s
r
ec
o
n
s
tr
u
ctio
n
-
b
ased
d
etec
to
r
s
[
9
]
,
[
1
4
]
,
[
1
5
]
,
[
2
1
]
:
i)
f
o
r
ec
asti
n
g
an
d
d
etec
tio
n
in
o
n
e
b
o
x
,
wh
er
e
th
e
o
u
tp
u
ts
o
f
th
e
h
y
b
r
id
f
o
r
ec
aster
ar
e
in
ten
d
ed
to
b
e
co
n
s
u
m
ed
in
a
s
in
g
le
o
p
e
r
atio
n
s
d
ash
b
o
ar
d
,
r
ath
er
th
a
n
two
an
d
ii)
d
etec
tab
ilit
y
d
ec
o
m
p
o
s
itio
n
p
er
class
,
r
ev
ea
lin
g
wh
ich
o
f
t
h
e
f
o
u
r
an
o
m
aly
class
es
ca
n
o
r
ca
n
n
o
t
b
e
r
ec
o
v
er
ed
th
r
o
u
g
h
th
e
m
an
if
o
ld
-
d
ep
ar
tu
r
e
ass
u
m
p
tio
n
,
with
p
r
ac
tical
o
p
e
r
atio
n
al
im
p
licatio
n
s
f
o
r
h
y
b
r
i
d
f
ir
s
t
-
p
ass
-
f
ilter
d
ep
lo
y
m
en
ts
.
T
h
is
p
ap
er
m
a
k
es
th
e
f
o
llo
win
g
co
n
tr
ib
u
tio
n
s
:
i)
r
ep
r
o
d
u
cib
le
lo
ca
lized
d
ataset
,
w
e
d
o
cu
m
e
n
t,
p
ar
am
eter
ize
an
d
p
u
b
lis
h
a
s
y
n
th
etic
s
et
o
f
3
1
5
,
2
4
7
r
ec
o
r
d
s
f
r
o
m
1
3
ce
ll
s
ites
,
co
v
er
in
g
th
r
ee
o
p
er
at
o
r
s
in
f
iv
e
b
a
n
d
s
,
with
f
o
u
r
o
p
er
at
o
r
s
ty
le
an
o
m
aly
class
es
with
a
to
tal
r
ate
o
f
2
.
0
3
%,
b
ased
o
n
p
u
b
licly
av
ailab
le
POTRAZ
s
ec
to
r
s
tati
s
tics
;
ii)
b
aselin
e
f
o
r
en
d
-
to
-
en
d
fo
r
e
ca
s
tin
g
an
d
d
etec
tio
n
,
w
e
co
m
b
in
e
a
f
o
r
ec
aster
b
ased
o
n
a
h
y
b
r
id
C
NN
–
L
STM
ap
p
r
o
ac
h
with
a
d
etec
to
r
b
ased
o
n
a
VAE
in
th
e
s
am
e
s
tr
ea
m
in
g
p
ip
elin
e,
p
r
o
v
id
i
n
g
p
r
ep
r
o
ce
s
s
in
g
,
c
h
r
o
n
o
lo
g
ical
s
p
lits
,
f
ix
ed
th
r
esh
o
ld
s
,
an
d
f
u
ll
r
ep
o
r
tin
g
o
f
m
etr
ics,
h
y
p
er
p
ar
am
et
er
s
,
s
en
s
itiv
ity
an
aly
s
is
,
an
d
co
n
f
u
s
io
n
m
atr
ices
all
in
a
r
ep
r
o
d
u
cib
le
m
an
n
er
;
iii)
c
o
m
p
r
e
h
en
s
iv
e
b
aselin
e
c
o
m
p
ar
is
o
n
.
W
e
co
m
p
ar
e
th
e
C
NN
–
L
STM
to
n
aiv
e
f
o
r
ec
aster
s
,
s
ea
s
o
n
al
-
n
aiv
e
f
o
r
ec
aster
s
,
au
to
r
e
g
r
ess
iv
e
(
AR
)
(
2
4
)
f
o
r
ec
aster
s
an
d
f
ee
d
-
f
o
r
war
d
d
ee
p
n
e
u
r
al
n
etwo
r
k
(
DNN
)
f
o
r
ec
aster
s
(
wh
ich
ex
clu
d
es
L
STM
s
an
d
g
ated
r
ec
u
r
r
en
t
u
n
it
(
GR
U
)
)
an
d
p
er
f
o
r
m
p
air
ed
s
tatis
ti
ca
l
test
s
an
d
9
5
%
co
n
f
id
en
ce
in
ter
v
als
f
o
r
th
e
L
STM
-
o
n
ly
,
GR
U
an
d
C
NN
-
o
n
ly
f
o
r
ec
aster
s
.
Similar
ly
,
we
co
m
p
ar
e
th
e
VAE
with
s
tati
s
tical
th
r
esh
o
ld
in
g
,
I
s
o
latio
n
Fo
r
est
,
o
n
e
-
class
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM
)
an
d
v
an
illa
au
to
en
co
d
er
.
W
e
also
co
m
p
ar
e
o
u
r
r
esu
lts
with
th
e
r
ec
en
t
liter
atu
r
e
C
o
n
v
L
STM
an
d
T
r
an
s
f
o
r
m
e
r
b
ased
f
o
r
ec
ast
b
aselin
es
[
2
2
]
,
[
2
3
]
;
i
v
)
t
r
u
t
h
f
u
l r
ep
r
esen
tatio
n
an
d
d
ec
lar
atio
n
o
f
tr
a
d
e
-
o
f
f
s
.
W
e
r
ep
o
r
t
th
e
b
eh
a
v
io
u
r
o
f
th
e
VAE
i
n
a
h
ig
h
p
r
ec
is
io
n
/lo
w
r
ec
all
r
eg
im
e
an
d
th
e
p
er
f
o
r
m
a
n
ce
p
er
an
o
m
al
y
ty
p
e,
in
clu
d
i
n
g
b
o
o
ts
tr
ap
co
n
f
id
en
ce
i
n
ter
v
als,
an
d
d
is
cu
s
s
th
e
lim
its
o
n
s
y
n
th
etic
lo
ca
lis
atio
n
,
v
iewin
g
th
e
s
y
s
tem
as
a
f
ir
s
t
-
p
ass
o
p
er
atio
n
al
f
ilter
r
ath
e
r
th
a
n
a
d
ep
lo
y
m
en
t
-
g
r
ad
e
d
etec
to
r
.
Ad
d
itio
n
ally
,
we
s
u
g
g
est
s
p
ec
if
ic
ap
p
r
o
ac
h
es
t
o
b
o
o
s
t
r
ec
all
f
o
r
th
e
m
o
r
e
d
if
f
icu
lt
class
es
u
n
au
th
o
r
ized
an
d
s
ig
n
al
d
r
o
p
o
u
t
;
an
d
v
)
b
r
o
ad
er
c
o
m
p
u
ter
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
5
6
-
270
258
s
cien
ce
(
CS
)
r
elev
an
ce
.
Giv
e
n
th
e
h
o
u
r
l
y
ag
g
r
eg
atio
n
o
f
k
ey
p
er
f
o
r
m
a
n
ce
in
d
icat
o
r
(
KP
I
)
s
tr
ea
m
s
an
d
th
e
lig
h
t
weig
h
t
-
o
u
t
p
e
r
s
am
p
le,
th
e
p
ip
elin
e
ca
n
b
e
d
ep
lo
y
ed
i
n
a
d
is
tr
ib
u
ted
way
ac
r
o
s
s
o
p
e
r
ato
r
an
d
r
eg
u
lato
r
y
b
o
u
n
d
ar
ies
(
ed
g
e
-
p
r
e
-
f
ilter
/cl
o
u
d
-
tr
ai
n
er
s
p
lit),
b
e
ex
ten
d
ed
in
a
f
ed
er
ated
way
p
r
eser
v
in
g
o
p
e
r
ato
r
co
n
f
id
en
tiality
,
an
d
ca
n
b
e
in
teg
r
ated
with
in
ter
n
et
o
f
th
i
n
g
s
(
I
o
T
)
b
ased
wid
e
-
a
r
ea
s
p
ec
tr
u
m
m
o
n
ito
r
in
g
co
n
n
ec
tio
n
s
we
d
e
v
elo
p
f
r
o
m
s
ec
tio
n
5.
T
h
e
r
est
o
f
th
e
p
ap
er
is
ar
r
a
n
g
ed
as
f
o
llo
ws.
A
r
ev
iew
o
f
r
elev
an
t
wo
r
k
in
th
e
ar
ea
s
o
f
ce
llu
lar
f
o
r
ec
asti
n
g
,
d
ee
p
wir
eless
p
r
e
d
ictio
n
,
r
ec
o
n
s
tr
u
ctio
n
-
b
ased
an
o
m
aly
d
etec
tio
n
,
AI
f
o
r
s
p
e
ctr
u
m
m
an
a
g
em
en
t
,
an
d
s
y
n
th
etic
lo
ca
lized
d
atase
ts
is
p
r
o
v
id
ed
in
th
e
n
e
x
t
s
ec
tio
n
.
Sectio
n
2
in
tr
o
d
u
ce
s
d
eta
ils
o
n
th
e
d
ataset,
lo
ca
lizatio
n
p
r
o
ce
d
u
r
e,
a
r
ch
i
tectu
r
e
o
f
C
NN
–
L
STM
an
d
VAE
,
h
y
p
er
p
ar
am
eter
s
elec
tio
n
,
tr
ain
i
n
g
a
n
d
ev
alu
atio
n
p
r
o
ce
d
u
r
e,
an
d
r
e
p
r
o
d
u
cib
ilit
y
s
ettin
g
.
R
esu
lts
o
f
f
o
r
ec
asti
n
g
a
n
d
a
n
o
m
aly
d
etec
tio
n
,
s
tatis
tical
an
aly
s
is
,
s
en
s
it
iv
ity
an
aly
s
is
,
an
d
a
d
is
cu
s
s
io
n
ab
o
u
t
th
e
d
etec
tab
ilit
y
p
er
class
ar
e
p
r
esen
ted
in
s
ec
tio
n
3
.
T
h
e
lim
itatio
n
s
ar
e
lis
ted
in
s
ec
tio
n
4
,
an
d
th
e
c
o
n
clu
s
io
n
is
s
u
m
m
ar
ized
in
s
ec
tio
n
5
with
f
u
tu
r
e
w
o
r
k
s
to
war
d
s
5
G,
I
o
T
an
d
f
ed
er
ated
ex
ten
s
io
n
s
.
−
Hy
b
r
id
d
ee
p
m
o
d
el
f
o
r
f
o
r
ec
asti
n
g
ce
llu
lar
tr
af
f
ic
T
h
e
m
u
lti
-
s
ca
le
tem
p
o
r
al
s
tr
u
ctu
r
e
ass
o
ciate
d
with
ce
llu
lar
tr
af
f
ic
-
th
e
d
iu
r
n
al
c
y
cle
with
in
-
d
ay
,
wee
k
ly
cy
cle
,
h
o
lid
ay
ef
f
ec
ts
,
an
d
lo
n
g
er
-
ter
m
g
r
o
wth
-
is
p
ar
tially
d
escr
ib
ed
b
y
class
ical
s
tati
s
tical
m
eth
o
d
s
lik
e
au
to
r
e
g
r
ess
iv
e
in
teg
r
ated
m
o
v
i
n
g
av
er
ag
e
(
AR
I
MA
)
a
n
d
e
x
p
o
n
en
tial
s
m
o
o
t
h
in
g
[
3
]
.
Mo
d
els
with
d
ee
p
r
ec
u
r
r
en
t
s
tr
u
ctu
r
es
(
L
STM
,
GR
U)
[
1
0
]
,
[
1
1
]
,
h
av
e
n
o
w
b
ec
o
m
e
s
tan
d
a
r
d
b
u
ild
in
g
b
lo
ck
s
an
d
h
y
b
r
id
C
NN
–
L
STM
m
o
d
els
h
av
e
d
e
m
o
n
s
tr
ated
im
p
r
o
v
em
e
n
t
o
v
er
eith
er
co
m
p
o
n
e
n
t
alo
n
e
o
n
l
o
n
g
-
ter
m
ev
o
lu
tio
n
(
LTE
)
tr
af
f
ic
a
n
d
m
u
lti
-
ce
ll
lo
ad
[
4
]
,
[
5
]
.
C
o
n
v
o
lu
tio
n
al
lay
er
s
ca
p
tu
r
e
lo
ca
l
s
h
ap
es
m
o
r
n
in
g
r
am
p
s
,
ev
e
n
in
g
p
ea
k
s
,
s
h
o
r
t
b
u
r
s
ts
as
a
s
m
all
k
er
n
el
ar
e
r
o
lled
ac
r
o
s
s
th
e
tim
e
ax
is
,
an
d
r
ec
u
r
r
en
t
lay
er
s
ar
e
u
s
ed
to
p
ass
lo
n
g
er
-
r
an
g
e
tem
p
o
r
al
c
o
n
tex
t
th
r
o
u
g
h
th
e
7
2
-
h
o
u
r
in
p
u
t
win
d
o
w.
I
n
o
th
er
ar
ea
s
,
s
p
atio
tem
p
o
r
al
ex
ten
s
io
n
s
lik
e
C
o
n
v
L
STM
[
2
2
]
h
av
e
b
e
en
p
r
o
p
o
s
ed
an
d
th
ey
ar
e
p
ar
ti
cu
lar
ly
s
u
itab
le
f
o
r
m
u
lti
-
ce
ll
g
r
id
s
;
tr
an
s
f
o
r
m
er
-
b
ased
f
o
r
ec
aster
s
[
2
3
]
h
a
v
e
d
em
o
n
s
tr
ated
ef
f
ec
tiv
e
p
er
f
o
r
m
an
ce
o
n
l
o
n
g
-
h
o
r
iz
o
n
s
er
ies
in
o
th
er
f
ield
s
,
b
u
t
th
ey
r
eq
u
ir
e
m
u
ch
m
o
r
e
d
ata
an
d
co
m
p
u
tin
g
p
o
wer
th
an
t
h
e
s
co
p
e
o
f
th
is
p
ap
er
.
T
h
e
C
NN
–
L
STM
d
esig
n
wh
ich
is
u
tili
ze
d
h
er
e
is
in
th
e
tr
ad
itio
n
o
f
h
y
b
r
id
d
esig
n
s
,
with
th
e
tw
o
co
n
v
o
lu
tio
n
al
b
lo
ck
s
f
o
llo
wed
b
y
two
-
s
tack
L
STM
.
−
R
ec
o
n
s
tr
u
ctio
n
-
b
ased
an
o
m
al
y
d
etec
tio
n
Op
er
ato
r
an
o
m
aly
in
v
e
n
to
r
ies
ar
e
n
o
t o
f
te
n
co
m
p
letely
la
b
eled
,
an
d
e
v
en
wh
en
t
h
ey
ar
e,
m
an
y
lab
els
ten
d
to
b
e
n
o
is
y
an
d
/o
r
b
iase
d
to
war
d
s
v
is
u
al
f
ailu
r
e
m
o
d
e
s
.
T
h
u
s
th
e
liter
atu
r
e
h
as
s
ettled
to
u
n
s
u
p
er
v
is
ed
r
ec
o
n
s
tr
u
ctio
n
b
ased
d
etec
to
r
s
[
6
]
–
[
9
]
.
I
n
lin
e
with
th
is
,
th
e
VAE
o
f
Kin
g
m
a
a
n
d
W
ellin
g
[
6
]
en
c
o
u
r
a
g
es
th
e
laten
t
s
p
ac
e
to
b
e
clo
s
er
to
a
Gau
s
s
ian
p
r
io
r
,
wh
ile
th
e
r
ec
o
n
s
tr
u
ctio
n
-
p
r
o
b
ab
ilit
y
f
r
a
m
in
g
p
r
o
p
o
s
ed
b
y
I
s
lam
et
a
l
.
[
7
]
f
o
r
5
G
is
p
r
ev
alen
t
in
th
e
ce
llu
lar
d
o
m
ain
[
1
4
]
,
[
1
5
]
.
R
ec
en
tly
,
m
u
ltiv
ar
i
ate
tim
e
s
er
ies
h
av
e
s
ee
n
co
m
p
etitiv
e
p
er
f
o
r
m
a
n
c
e
b
y
an
o
m
aly
d
etec
to
r
s
s
u
ch
as
th
e
a
n
o
m
al
y
t
r
a
n
s
f
o
r
m
er
[
1
3
]
s
ig
n
if
ica
n
tly
h
ig
h
er
co
m
p
u
te
co
s
t.
A
co
n
s
is
ten
t
ex
p
er
im
e
n
tal
f
in
d
in
g
o
v
er
all
th
e
m
eth
o
d
s
in
r
ec
o
n
s
tr
u
ctio
n
-
b
ased
m
eth
o
d
s
is
th
at
h
ig
h
p
e
r
ce
n
tile
th
r
esh
o
ld
s
lead
to
h
ig
h
p
r
ec
is
io
n
a
n
d
lo
w
(
well
l
o
wer
th
a
n
wh
at
a
s
u
p
er
v
is
ed
m
eth
o
d
with
th
e
s
am
e
ev
i
d
en
ce
c
o
u
ld
ac
h
iev
e)
r
ec
all
[
1
6
]
,
[
1
7
]
.
W
e
m
ak
e
t
h
e
p
atter
n
e
x
p
licit
f
o
r
ce
llu
lar
a
n
o
m
alies
b
y
o
u
r
p
er
-
class
an
aly
s
is
in
s
e
ctio
n
3
.
2
.
−
R
ep
r
o
d
u
cib
le
em
e
r
g
in
g
-
m
ar
k
e
t d
atasets
an
d
AI
f
o
r
s
p
ec
tr
u
m
m
an
ag
em
en
t
Sp
ec
tr
u
m
s
en
s
in
g
an
d
d
y
n
am
ic
s
p
ec
tr
u
m
ac
ce
s
s
f
o
r
AI
s
y
s
tem
s
m
ak
e
th
e
ca
s
e
f
o
r
h
av
i
n
g
s
m
all,
ad
ap
ted
m
o
d
els
with
a
s
tr
o
n
g
f
o
c
u
s
o
n
r
eg
u
lato
r
y
c
o
m
p
lian
ce
an
d
i
n
f
r
astru
ctu
r
e
lim
itatio
n
s
[
1
8
]
–
[
2
0
]
.
Av
ailab
ilit
y
o
f
d
ata,
h
eter
o
g
e
n
eity
o
f
in
f
r
astru
ctu
r
e
an
d
p
o
licy
ar
e
th
e
b
in
d
in
g
co
n
s
tr
ain
ts
in
all
em
p
ir
ical
s
tu
d
ies
o
f
th
e
d
ep
lo
y
m
en
t
o
f
Af
r
ican
an
d
o
t
h
er
em
e
r
g
in
g
-
m
ar
k
et
n
etwo
r
k
s
[
1
9
]
,
[
2
1
]
,
[
2
4
]
,
[
2
5
]
an
d
ev
e
n
ca
p
ac
ity
-
b
u
ild
in
g
ac
tiv
ities
b
y
I
n
te
r
n
atio
n
al
T
elec
o
m
m
u
n
i
ca
tio
n
Un
io
n
Dev
elo
p
m
en
t
Secto
r
(
I
T
U
-
D)
h
a
v
e
in
clu
d
ed
th
em
as
f
ir
s
t
-
o
r
d
er
r
eq
u
ir
em
en
ts
in
s
y
s
tem
s
a
im
ed
at
in
f
o
r
m
in
g
r
e
g
u
lato
r
y
d
ec
is
io
n
s
[
1
9
]
.
I
n
n
etwo
r
k
r
esear
ch
wh
e
r
e
p
r
o
d
u
ctio
n
t
r
ac
es
ar
e
k
ep
t
c
o
n
f
i
d
en
tial
an
d
b
e
n
ch
m
a
r
k
d
ataset
s
g
et
o
u
td
ated
v
e
r
y
r
ap
id
ly
,
th
e
u
s
e
o
f
s
y
n
th
etic
d
atasets
h
as
a
lo
n
g
tr
ad
itio
n
in
r
esear
ch
:
an
y
s
y
n
th
etic
d
ataset
s
h
o
u
ld
b
e
ca
lib
r
ated
to
v
er
if
iab
le
r
ea
l
-
wo
r
ld
d
ata,
b
e
r
ep
r
o
d
u
cib
le,
an
d
b
e
clea
r
ly
d
esig
n
ate
d
as
s
y
n
th
etic
tr
ac
es
o
f
th
e
o
p
e
r
ato
r
s
[
1
]
,
[
2
]
,
[
1
9
]
.
T
h
e
f
r
a
m
ewo
r
k
d
escr
ib
ed
in
th
is
p
ap
er
f
its
in
th
is
tr
a
d
itio
n
wh
ile
b
ein
g
b
ased
o
n
p
u
b
licly
av
ailab
le
s
ec
to
r
-
lev
el
s
tatis
tics
(
POT
R
AZ
'
s
t
r
ac
es)
in
s
tead
o
f
c
o
n
f
id
e
n
tia
l
o
p
er
at
o
r
-
in
ter
n
al
tr
ac
es
an
d
is
ea
s
ily
s
ca
lab
l
e
to
d
is
tr
ib
u
te
d
an
d
f
e
d
er
a
ted
d
ep
lo
y
m
en
ts
ac
r
o
s
s
o
p
er
ato
r
b
o
u
n
d
ar
ies,
an
in
cr
ea
s
in
g
ly
im
p
o
r
tan
t
ar
ch
itectu
r
al
ch
allen
g
e
as
5
G
an
d
I
o
T
er
a
m
u
lti
-
s
tak
eh
o
l
d
e
r
s
p
ec
tr
u
m
s
h
a
r
in
g
ev
o
lv
es
[
2
4
]
,
[
2
5
]
.
2.
M
E
T
H
O
D
2
.
1
.
No
t
a
t
io
n
a
nd
pro
blem
s
t
a
t
em
ent
L
et
x
(
c,
b
,
o
,
t)
∈
ℝ
ᵈ
d
en
o
te
t
h
e
f
ea
tu
r
e
v
ec
to
r
o
b
s
er
v
ed
at
ce
ll
c,
b
a
n
d
b
,
o
p
er
ato
r
o
,
an
d
d
is
cr
ete
h
o
u
r
t.
T
h
e
co
m
p
o
n
en
ts
o
f
x
ar
e
r
ec
eiv
ed
s
ig
n
al
s
tr
en
g
th
in
d
icato
r
(
R
SS
I
)
d
B
m
,
ch
an
n
el
o
cc
u
p
an
cy
(
%),
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
C
o
n
vo
lu
tio
n
a
l
n
eu
r
a
l n
etw
o
r
k
a
n
d
lo
n
g
s
h
o
r
t
-
term me
mo
r
y
f
o
r
ec
a
s
tin
g
a
n
d
…
(
R
u
v
a
r
a
s
h
e
C
.
Ho
ve
)
259
tr
af
f
ic
v
o
lu
m
e
(
GB
/h
o
u
r
)
,
a
n
d
in
ter
f
er
en
ce
lev
el
(
d
B
)
,
a
u
g
m
en
ted
with
cy
clic
tim
e
-
of
-
d
ay
an
d
lag
f
ea
t
u
r
es
d
escr
ib
ed
in
s
ec
tio
n
2
.
3
.
4
.
T
wo
co
m
p
lem
en
tar
y
task
s
ar
e
th
en
d
ef
in
ed
o
n
th
is
r
ep
r
esen
tatio
n
.
T
h
e
f
o
r
ec
asti
n
g
task
lear
n
s
a
m
ap
p
in
g
:
ŷ
(
+
1
:
+
)
=
(
(
−
+
1
:
)
)
w
ith
in
p
u
t
win
d
o
w
L
=7
2
h
o
u
r
s
an
d
f
o
r
ec
ast
h
o
r
iz
o
n
H=
6
h
o
u
r
s
.
C
o
n
ce
p
tu
ally
,
th
e
f
o
r
ec
as
ter
r
ea
d
s
th
r
ee
f
u
ll
d
ay
s
o
f
h
is
to
r
y
an
d
p
r
o
jects
th
e
n
ex
t
q
u
ar
ter
-
d
ay
o
f
n
etwo
r
k
lo
ad
.
T
h
e
an
o
m
aly
task
lear
n
s
an
u
n
s
u
p
er
v
is
ed
r
ec
o
n
s
tr
u
ctio
n
g
ϕ
s
u
c
h
th
at
t
h
e
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
is
s
m
all
o
n
n
o
r
m
al
r
ec
o
r
d
s
an
d
ex
ce
e
d
s
a
f
ix
ed
th
r
esh
o
l
d
τ
o
n
a
n
o
m
alo
u
s
o
n
es.
I
n
tu
itiv
ely
,
e(
x
)
m
ea
s
u
r
es
h
o
w
f
ar
a
r
ec
o
r
d
s
its
f
r
o
m
th
e
m
a
n
if
o
ld
o
f
n
o
r
m
al
o
p
er
atio
n
th
at
th
e
VAE
h
as
lear
n
ed
to
r
ep
r
o
d
u
ce
,
a
n
d
th
e
th
r
esh
o
l
d
τ
is
th
e
d
ec
is
io
n
b
o
u
n
d
ar
y
wh
ich
a
r
ec
o
r
d
is
f
lag
g
ed
.
Sectio
n
2
.
5
f
ix
es
τ
at
th
e
9
9
th
p
er
ce
n
tile
o
f
v
alid
atio
n
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
s
.
C
h
o
o
s
in
g
τ
f
r
o
m
v
alid
atio
n
r
ath
e
r
th
an
test
d
ata
f
o
llo
ws s
tan
d
ar
d
r
ec
o
n
s
tr
u
cti
o
n
-
b
ased
p
r
ac
tice
[
6
]
,
[
7
]
.
(
)
=
‖
−
‖
²
2
.
2
.
Sy
s
t
e
m
o
v
er
v
iew
Fig
u
r
e
1
s
u
m
m
ar
izes
th
e
en
d
-
to
-
en
d
p
ip
elin
e
,
an
d
Fig
u
r
e
2
d
etails
th
e
lo
ca
l
izatio
n
s
tag
e
th
at
p
r
o
d
u
ce
s
th
e
in
p
u
t
d
ata.
T
h
e
p
ip
elin
e
co
n
tain
s
f
o
u
r
s
tag
es:
i)
in
g
esti
o
n
o
f
th
e
Z
im
b
a
b
w
e
s
p
ec
tr
u
m
d
atase
t
d
escr
ib
ed
in
s
ec
tio
n
2
.
3
.
1
;
ii
)
p
r
e
p
r
o
ce
s
s
in
g
a
n
d
s
lid
in
g
-
win
d
o
w
co
n
s
tr
u
ctio
n
d
escr
ib
ed
in
s
ec
tio
n
2
.
3
.
4
;
iii)
p
ar
allel
C
NN
–
L
STM
f
o
r
ec
asti
n
g
(
s
ec
tio
n
2
.
4
)
an
d
VAE
an
o
m
aly
d
etec
tio
n
(
s
ec
tio
n
2
.
5
)
;
an
d
i
v
)
a
u
n
if
ied
r
ep
o
r
tin
g
lay
er
th
at
c
o
n
s
u
m
e
s
b
o
th
m
o
d
el
o
u
tp
u
ts
in
to
a
s
in
g
le
s
et
o
f
o
p
er
atio
n
al
ar
te
f
ac
ts
(
s
ec
tio
n
2
.
6
)
.
T
h
e
two
m
o
d
el
b
r
an
c
h
es
s
h
ar
e
th
e
s
am
e
p
r
ep
r
o
ce
s
s
ed
f
ea
tu
r
e
r
ep
r
e
s
en
tatio
n
b
u
t
u
s
e
in
d
ep
en
d
e
n
t
tr
ain
in
g
lo
o
p
s
;
th
is
d
esig
n
c
h
o
ice
is
d
elib
er
ate,
as
it
allo
ws
th
e
f
o
r
ec
aster
to
b
e
r
etr
ain
ed
o
n
ag
g
r
eg
ate
tr
af
f
ic
alo
n
e
wh
ile
th
e
d
etec
to
r
co
n
tin
u
es to
o
p
er
ate
o
n
th
e
p
er
-
r
ec
o
r
d
f
ea
t
u
r
e
s
p
ac
e.
Fig
u
r
e
1
.
E
n
d
-
to
-
en
d
f
o
r
ec
asti
n
g
an
d
an
o
m
aly
-
d
etec
tio
n
p
ip
elin
e.
I
n
p
u
ts
f
lo
w
f
r
o
m
th
e
Z
i
m
b
ab
we
s
p
ec
tr
u
m
d
ataset
th
r
o
u
g
h
p
r
e
p
r
o
ce
s
s
in
g
in
to
two
p
ar
allel
b
r
an
ch
es a
C
NN
–
L
STM
f
o
r
ec
aster
o
n
t
h
e
n
etwo
r
k
-
ag
g
r
eg
ate
h
o
u
r
ly
s
er
ies an
d
a
VAE
d
etec
to
r
o
n
t
h
e
p
er
-
r
ec
o
r
d
f
ea
tu
r
e
s
p
ac
e
wh
o
s
e
o
u
t
p
u
ts
f
ee
d
a
u
n
i
f
ied
r
ep
o
r
tin
g
lay
e
r
Fig
u
r
e
2
.
Sy
n
th
etic
lo
ca
lizatio
n
wo
r
k
f
l
o
w
ag
ain
s
t p
u
b
licly
p
u
b
lis
h
ed
POTRAZ
s
ec
to
r
s
tati
s
tics
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
5
6
-
270
260
2
.
3
.
Da
t
a
s
et
a
nd
s
y
nthet
ic
lo
ca
liza
t
io
n wo
rkf
lo
w
2
.
3
.
1
.
Da
t
a
s
et
o
v
er
v
iew
T
h
e
Z
i
m
b
a
b
we
s
p
ec
tr
u
m
d
at
aset
c
o
m
p
r
is
es
3
1
5
,
2
4
7
h
o
u
r
l
y
r
e
c
o
r
d
s
s
p
a
n
n
in
g
1
3
c
ell
s
ites
at
t
h
e
co
u
n
t
r
y
’
s
m
ai
n
u
r
b
a
n
a
n
d
r
eg
i
o
n
a
l
c
en
tr
es
(
Ha
r
a
r
e
C
B
D,
H
ar
ar
e
W
est
,
H
a
r
a
r
e
E
a
s
t,
B
u
law
a
y
o
C
B
D,
B
u
law
a
y
o
S
o
u
t
h
,
C
h
i
tu
n
g
wi
z
a,
M
u
ta
r
e
C
B
D,
Gw
er
u
C
B
D,
Kwe
k
w
e
C
B
D
,
Ka
d
o
m
a
C
B
D,
Ma
s
v
i
n
g
o
C
B
D,
Vict
o
r
ia
Fal
ls
,
a
n
d
L
i
v
i
n
g
s
t
o
n
e)
,
t
h
r
ee
m
o
b
il
e
o
p
e
r
at
o
r
s
(
E
co
n
e
t,
Net
On
e,
a
n
d
T
ele
ce
l)
,
an
d
f
i
v
e
f
r
e
q
u
e
n
c
y
b
a
n
d
s
(
7
0
0
,
8
5
0
,
9
0
0
,
1
8
0
0
,
an
d
2
6
0
0
MH
z
)
.
E
a
ch
r
e
co
r
d
r
e
p
o
r
ts
a
t
im
esta
m
p
,
c
ell
t
o
wer
,
o
p
er
at
o
r
,
b
a
n
d
,
d
e
n
s
it
y
t
y
p
e
(
u
r
b
a
n
/s
u
b
u
r
b
a
n
/
r
u
r
al
)
,
t
r
a
f
f
ic
v
o
lu
m
e
i
n
GB
,
R
SS
I
in
d
B
m
,
c
h
a
n
n
el
o
cc
u
p
a
n
cy
(
%
)
,
in
te
r
f
er
e
n
ce
lev
el
,
h
o
u
r
o
f
d
a
y
,
d
a
y
o
f
wee
k
,
a
n
a
n
o
m
al
y
f
l
ag
,
a
n
d
a
f
o
u
r
-
class
a
n
o
m
al
y
t
y
p
e
l
ab
el
.
Desc
r
i
p
ti
v
e
s
t
atis
t
ics
f
o
r
th
e
p
r
im
ar
y
m
o
d
el
li
n
g
f
e
at
u
r
es
a
r
e
r
e
p
o
r
t
ed
i
n
T
a
b
l
e
1
,
a
n
d
t
h
e
a
n
o
m
al
y
c
o
m
p
o
s
it
io
n
is
r
ep
o
r
te
d
i
n
T
a
b
l
e
2
.
T
ab
le
1
.
Descr
ip
tiv
e
s
tatis
tics
o
f
p
r
im
a
r
y
m
o
d
ellin
g
f
ea
tu
r
es
F
e
a
t
u
r
e
M
e
a
n
S
t
d
.
D
e
v
.
M
i
n
M
a
x
Tr
a
f
f
i
c
(
G
B
/
h
o
u
r
)
4
2
3
.
9
3
2
9
7
.
2
0
3
0
.
3
1
6
,
0
3
1
.
4
6
R
S
S
I
(
d
B
m)
−
8
5
.
2
4
2
0
.
4
9
−
1
4
0
.
0
0
−
5
0
.
0
0
C
h
a
n
n
e
l
o
c
c
u
p
a
n
c
y
(
%)
5
2
.
8
3
2
4
.
6
9
1
0
.
0
0
1
0
0
.
0
0
T
ab
le
2
.
An
o
m
aly
co
u
n
ts
an
d
o
v
er
all
r
ate
A
n
o
m
a
l
y
c
l
a
ss
C
o
u
n
t
S
h
a
r
e
o
f
a
n
o
m
a
l
i
e
s
(
%)
S
h
a
r
e
o
f
t
o
t
a
l
(
%)
U
n
a
u
t
h
o
r
i
z
e
d
1
,
6
6
0
2
5
.
9
4
0
.
5
2
7
S
i
g
n
a
l
D
r
o
p
o
u
t
1
,
6
2
3
2
5
.
3
6
0
.
5
1
5
Tr
a
f
f
i
c
S
u
r
g
e
1
,
5
9
9
2
4
.
9
8
0
.
5
0
7
I
n
t
e
r
f
e
r
e
n
c
e
1
,
5
1
8
2
3
.
7
2
0
.
4
8
2
To
t
a
l
a
n
o
mal
i
e
s
6
,
4
0
0
1
0
0
.
0
0
2
.
0
3
0
To
t
a
l
r
e
c
o
r
d
s
3
1
5
,
2
4
7
1
0
0
.
0
0
2
.
3
.
2
.
Sy
nthet
ic
lo
ca
liza
t
io
n
wo
rk
f
lo
w
T
h
e
d
ataset
is
v
alid
ated
with
o
f
f
icial
POTRAZ
s
ec
to
r
al
d
ata
[
1
]
,
[
2
]
.
T
h
e
f
iv
e
co
m
p
o
n
en
ts
ar
e:
i)
u
r
b
a
n
/s
u
b
u
r
b
an
/r
u
r
al
m
ix
b
ased
o
n
c
o
v
er
a
g
e
s
tatis
tics
;
ii)
weig
h
ts
attac
h
ed
to
r
ec
o
r
d
-
co
u
n
ts
b
ased
o
n
p
u
b
lis
h
ed
u
tili
za
tio
n
tab
les;
iii)
a
b
i
-
m
o
d
al
d
iu
r
n
al
en
v
el
o
p
e
(
p
ea
k
ar
o
u
n
d
1
0
:0
0
,
tr
o
u
g
h
ar
o
u
n
d
2
1
:0
0
,
p
ea
k
-
to
-
tr
o
u
g
h
r
atio
~
2
.
6
,
we
ek
en
d
am
p
litu
d
e
~0
.
8
8
)
b
ased
o
n
th
e
q
u
alitativ
e
Qo
S
co
m
m
en
tar
y
f
r
o
m
t
h
e
s
ec
to
r
r
ep
o
r
ts
;
iv
)
af
f
in
e
r
es
ca
lin
g
o
f
tr
a
f
f
ic
an
d
ac
tiv
e
-
u
s
er
f
ea
tu
r
es
to
m
atch
t
h
e
m
ar
g
in
al
m
ea
n
s
an
d
d
is
p
er
s
io
n
s
o
f
th
e
s
ec
to
r
r
ep
o
r
t
ag
g
r
eg
ates;
an
d
v
)
r
ejec
tio
n
s
am
p
lin
g
to
a
d
ju
s
t
th
e
o
v
e
r
all
an
o
m
aly
r
ate
to
2
.
0
0
%.
T
h
e
p
r
o
ce
d
u
r
e
is
d
eter
m
in
is
tically
r
ec
o
r
d
ed
i
n
Alg
o
r
ith
m
1
.
Alg
o
rit
hm
1
.
L
o
ca
lized
s
y
n
t
h
etic
d
ataset
co
n
s
tr
u
ctio
n
I
n
p
u
t:
Kag
g
le
4
G
b
ase
D₀;
P
OT
R
AZ
s
tati
s
tic
s
P; r
an
d
o
m
s
ee
d
s
s
Ou
tp
u
t: lo
ca
lized
s
y
n
th
etic
d
ataset
D
1
.
I
n
itialize
r
an
d
o
m
s
tate
with
s
ee
d
s
s
2
.
Stra
tify
ce
lls
in
to
u
r
b
an
/s
u
b
u
r
b
a
n
/r
u
r
al
u
s
in
g
d
e
n
s
ity
weig
h
ts
f
r
o
m
P
3
.
R
ewe
ig
h
t r
ec
o
r
d
s
b
y
b
an
d
{7
0
0
,
8
5
0
,
9
0
0
,
1
8
0
0
,
2
6
0
0
}
MH
z
4
.
Fit
d
iu
r
n
al/wee
k
ly
en
v
elo
p
e
E
(
t)
f
r
o
m
p
ea
k
-
h
o
u
r
p
atter
n
s
in
P
5
.
Ap
p
ly
m
u
ltip
licativ
e
s
ca
lin
g
:
tr
af
f
ic
←
tr
af
f
ic
·
E
(
t)
6
.
Af
f
in
e
-
r
escale
tr
af
f
ic
an
d
a
ctiv
e
-
u
s
er
f
ea
tu
r
es
7
.
I
n
ject
4
-
class
an
o
m
alies a
t 2
.
0
0
% r
ate
v
ia
r
ejec
tio
n
s
am
p
l
in
g
8
.
Ap
p
e
n
d
cy
clic
a
n
d
lag
f
ea
t
u
r
es (
s
ec
tio
n
2
.
3
.
4
)
9
.
Per
s
is
t D
with
m
etad
ata
(
s
ee
d
s
,
weig
h
ts
,
en
v
elo
p
es,
s
ca
les)
1
0
.
R
etu
r
n
D
A
2
%
an
o
m
aly
r
ate
is
u
s
ed
,
w
h
ich
n
ee
d
s
to
b
e
ex
p
licitly
ju
s
tifie
d
.
T
h
e
p
u
b
lis
h
ed
in
cid
e
n
t
-
r
ep
o
r
tin
g
tab
les
th
at
ar
e
u
s
ed
to
cr
ea
t
e
th
e
lo
ca
lizatio
n
in
d
icate
a
co
n
s
is
ten
t
p
atter
n
o
f
(
lo
ca
lis
ed
)
co
n
g
esti
o
n
an
d
d
r
o
p
p
ed
ca
ll
ev
en
ts
an
d
q
u
ality
-
of
-
s
er
v
ice
in
cid
e
n
ts
th
at,
wh
en
ex
p
r
ess
ed
as
a
p
r
o
p
o
r
ti
o
n
o
f
th
e
ce
ll
-
h
o
u
r
s
m
o
n
ito
r
ed
,
f
all
with
in
a
b
an
d
o
f
1
.
5
-
2
.
5
% a
cr
o
s
s
all
r
ep
o
r
ted
q
u
ar
ter
s
as d
etailed
in
POT
R
AZ
q
u
ar
ter
ly
s
ec
to
r
r
ep
o
r
ts
[
2
]
.
C
h
o
o
s
in
g
a
r
ejec
tio
n
s
am
p
lin
g
r
ate
o
f
2
.
0
0
%
(
2
.
0
3
0
%
af
ter
class
b
alan
cin
g
)
f
o
r
ce
s
th
e
s
y
n
th
etic
st
r
ea
m
to
b
e
with
i
n
th
e
r
an
g
e
th
at
is
lik
ely
to
b
e
e
n
co
u
n
ter
ed
in
a
n
o
p
er
atio
n
s
ce
n
tr
e
wh
en
u
s
in
g
r
ea
l
telem
etr
y
,
b
u
t
s
till
r
ar
e
en
o
u
g
h
to
test
th
e
ass
u
m
p
tio
n
o
f
u
n
s
u
p
er
v
is
ed
d
etec
tio
n
.
T
h
e
f
o
u
r
class
es
o
f
an
o
m
aly
(
in
ter
f
er
en
ce
,
tr
af
f
ic
s
u
r
g
e,
s
ig
n
al
d
r
o
p
o
u
t
,
an
d
u
n
au
t
h
o
r
iz
ed
)
an
d
th
e
ap
p
r
o
x
im
ate
b
ala
n
ce
b
etwe
en
th
em
m
ir
r
o
r
th
o
s
e
f
o
u
n
d
in
o
p
er
at
o
r
-
ty
p
e
a
n
o
m
aly
in
v
en
t
o
r
ies
an
d
th
e
o
n
es
th
at
th
e
r
eg
u
lato
r
will
m
o
s
t
lik
ely
r
eq
u
ir
e
to
b
e
f
lag
g
ed
.
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
C
o
n
vo
lu
tio
n
a
l
n
eu
r
a
l n
etw
o
r
k
a
n
d
lo
n
g
s
h
o
r
t
-
term me
mo
r
y
f
o
r
ec
a
s
tin
g
a
n
d
…
(
R
u
v
a
r
a
s
h
e
C
.
Ho
ve
)
261
2
.
3
.
3
.
E
t
hica
l a
nd
priv
a
cy
co
ns
idera
t
io
ns
All
in
p
u
ts
to
th
is
s
tu
d
y
ar
e
p
u
b
lic.
No
p
er
s
o
n
ally
id
en
tif
iab
le
in
f
o
r
m
atio
n
,
cu
s
to
m
e
r
id
en
tifie
r
s
,
b
illi
n
g
d
ata,
o
r
o
p
er
at
o
r
-
in
ter
n
al
tr
ac
es
wer
e
ac
ce
s
s
ed
at
a
n
y
s
tag
e.
C
ell
id
en
tifie
r
s
a
r
e
f
o
r
p
u
b
licly
k
n
o
w
n
city
-
lev
el
lo
ca
tio
n
s
an
d
n
o
t
f
o
r
co
n
f
i
d
en
tial,
o
p
er
ato
r
-
i
n
ter
n
al
ce
ll
I
D
s
.
Fu
r
th
er
eth
ics,
p
r
iv
ac
y
an
d
r
eg
u
lato
r
y
r
ev
iew
will b
e
n
ee
d
ed
f
o
r
an
y
ex
p
an
s
io
n
o
f
th
e
m
eth
o
d
o
lo
g
y
to
o
p
er
at
o
r
in
ter
n
al
tr
ac
es.
2
.
3
.
4
.
P
re
pro
ce
s
s
ing
Fo
r
war
d
-
a
n
d
b
ac
k
war
d
-
f
ill
a
r
e
u
s
ed
f
o
r
s
h
o
r
t
g
ap
s
,
lin
ea
r
in
ter
p
o
latio
n
is
u
s
ed
f
o
r
lo
n
g
er
g
a
p
s
.
R
ec
o
r
d
s
th
at
ar
e
d
u
p
licated
o
r
in
clu
d
e
im
p
r
o
b
a
b
le
o
u
tlier
s
ar
e
d
is
ca
r
d
ed
.
T
im
estam
p
s
ar
e
b
r
o
k
e
n
ap
ar
t
in
to
co
m
p
o
n
en
ts
o
f
h
o
u
r
-
of
-
d
a
y
,
d
ay
-
of
-
wee
k
an
d
m
o
n
th
-
of
-
y
e
a
r
an
d
s
en
t
th
r
o
u
g
h
s
in
e
a
n
d
c
o
s
in
e
tr
an
s
f
o
r
m
s
t
o
r
etain
th
eir
cy
clic
n
atu
r
e.
As
co
m
m
o
n
tim
e
-
s
er
ies
f
ea
tu
r
e
-
e
n
g
in
ee
r
in
g
p
r
ac
tice
[
2
6
]
,
th
e
l
ag
f
ea
tu
r
es
in
clu
d
e
f
ea
tu
r
es
f
r
o
m
th
e
p
r
e
v
io
u
s
h
o
u
r
,
th
e
s
am
e
h
o
u
r
th
e
d
a
y
b
ef
o
r
e,
an
d
th
e
s
am
e
d
ay
th
e
wee
k
b
ef
o
r
e.
Z
s
co
r
e
is
u
s
ed
to
n
o
r
m
alize
all
n
u
m
er
ic
f
ea
tu
r
es.
T
h
e
tim
e
s
er
ies
is
r
e
-
ar
r
an
g
e
d
in
a
s
u
p
er
v
is
ed
m
an
n
er
,
with
a
7
2
-
h
o
u
r
tim
e
win
d
o
w
an
d
a
6
-
h
o
u
r
f
o
r
ec
ast
h
o
r
izo
n
,
to
p
er
f
o
r
m
th
e
f
o
r
ec
asti
n
g
task
a
n
d
e
n
s
u
r
e
n
o
in
f
o
r
m
atio
n
f
r
o
m
th
e
f
u
tu
r
e
is
leak
ed
in
to
th
e
tr
a
in
in
g
.
2
.
4
.
CNN
–
L
ST
M
f
o
re
ca
s
t
er
T
h
e
f
o
r
e
ca
s
t
er
is
a
1
-
D
C
N
N
–
L
STM
h
y
b
r
i
d
,
as
s
h
o
w
n
in
F
i
g
u
r
e
3
,
wi
th
h
y
p
e
r
p
ar
am
ete
r
s
s
h
o
wn
in
T
a
b
le
3
.
Na
tu
r
a
ll
y
,
l
o
c
al
te
m
p
o
r
al
f
ea
t
u
r
es
li
k
e
m
o
r
n
i
n
g
r
a
m
p
s
,
e
v
e
n
i
n
g
p
e
ak
s
a
n
d
s
h
o
r
t
b
u
r
s
ts
ar
e
e
x
t
r
ac
te
d
wit
h
tw
o
1
-
D
co
n
v
o
l
u
ti
o
n
al
l
a
y
e
r
s
(
wit
h
6
4
an
d
3
2
f
i
lte
r
s
,
k
e
r
n
el
s
i
ze
3
,
r
e
cti
f
i
ed
li
n
ea
r
u
n
it
(
R
eL
U
)
ac
t
iv
ati
o
n
,
an
d
a
s
a
m
e
p
ad
d
i
n
g
)
a
n
d
b
a
t
ch
n
o
r
m
al
iz
ati
o
n
[
2
7
]
r
es
p
e
cti
v
el
y
.
T
h
e
i
d
ea
is
s
i
m
p
le
:
a
s
m
al
l
k
er
n
e
l
m
o
v
i
n
g
alo
n
g
t
h
e
h
o
u
r
l
y
a
x
is
is
k
i
n
d
o
f
li
k
e
a
n
“
e
y
e
”
t
h
at
ca
n
b
e
tr
ai
n
e
d
t
o
d
ete
ct
t
h
ese
r
e
cu
r
r
i
n
g
s
u
b
-
d
a
y
m
o
t
if
s
,
a
n
d
p
r
ec
is
el
y
t
h
at
is
t
h
e
s
o
r
t
o
f
t
h
i
n
g
th
at
p
u
r
e
ly
r
e
cu
r
r
e
n
t
m
o
d
e
l
s
c
an
b
e
s
l
o
w
t
o
c
ap
tu
r
e
.
A
f
te
r
th
e
c
o
n
v
o
l
u
t
io
n
a
l
b
l
o
c
k
,
t
h
e
r
e
is
a
m
a
x
-
p
o
o
li
n
g
(
2
,
2
)
a
n
d
a
d
r
o
p
o
u
t
(
0
.
2
)
la
y
e
r
.
T
h
e
lo
n
g
e
r
-
r
a
n
g
e
d
e
p
e
n
d
e
n
cie
s
ar
e
th
en
c
ap
tu
r
ed
ac
r
o
s
s
t
h
e
7
2
-
h
o
u
r
wi
n
d
o
w
b
y
tw
o
s
t
ac
k
e
d
L
S
T
M
la
y
er
s
wit
h
6
4
an
d
3
2
h
i
d
d
en
u
n
its
,
r
es
p
e
cti
v
e
ly
,
a
n
d
0
.
2
d
r
o
p
o
u
t
u
s
e
d
af
te
r
ea
c
h
.
T
h
e
6
-
h
o
u
r
ly
f
o
r
e
ca
s
ts
a
r
e
o
b
t
ai
n
e
d
i
n
a
s
in
g
l
e
f
e
ed
-
f
o
r
w
ar
d
p
a
s
s
u
s
i
n
g
a
3
2
-
u
n
it
R
eL
U
d
en
s
e
la
y
e
r
f
o
ll
o
w
ed
b
y
a
6
-
u
n
i
t
l
in
ea
r
o
u
t
p
u
t
h
ea
d
.
T
r
ai
n
i
n
g
wi
th
A
d
a
m
(
l
ea
r
n
i
n
g
r
ate
1
×
1
0
⁻³
)
,
m
ea
n
s
q
u
a
r
e
d
er
r
o
r
(
MS
E
)
l
o
s
s
,
e
a
r
l
y
s
to
p
p
i
n
g
(
p
at
ie
n
c
e
1
0
)
,
a
n
d
R
ed
u
c
eL
R
O
n
Pl
ate
a
u
(
f
ac
to
r
0
.
5
,
p
at
ie
n
c
e
5
)
.
Fig
u
r
e
3
.
Hy
b
r
id
1
-
D
C
NN
–
L
STM
ar
ch
itectu
r
e
f
o
r
s
ix
-
h
o
u
r
-
ah
ea
d
tr
af
f
ic
f
o
r
ec
asti
n
g
.
C
o
n
v
o
lu
tio
n
al
b
lo
ck
s
(
lef
t)
ca
p
tu
r
e
lo
ca
l w
ith
in
-
d
a
y
s
h
ap
e;
s
tack
ed
L
STM
lay
er
s
(
ce
n
tr
e)
ca
r
r
y
lo
n
g
er
-
r
a
n
g
e
tem
p
o
r
al
co
n
tex
t;
a
d
en
s
e
h
ea
d
(
r
ig
h
t
)
p
r
o
d
u
ce
s
th
e
s
ix
-
s
tep
f
o
r
ec
ast
Ta
b
le
3
.
C
NN
–
L
STM
h
y
p
e
r
p
a
r
am
eter
s
an
d
s
ea
r
ch
s
p
ac
e
C
o
m
p
o
n
e
n
t
F
i
n
a
l
s
e
t
t
i
n
g
S
e
a
r
c
h
s
p
a
c
e
(
v
a
l
i
d
a
t
i
o
n
o
n
l
y
)
I
n
p
u
t
w
i
n
d
o
w
L
7
2
h
o
u
r
s
{2
4
,
4
8
,
7
2
,
9
6
,
1
6
8
}
F
o
r
e
c
a
st
h
o
r
i
z
o
n
H
6
h
o
u
r
s
F
i
x
e
d
b
y
t
a
s
k
C
o
n
v
1
D
f
i
l
t
e
r
s
6
4
→3
2
{3
2
,
6
4
,
9
6
}
C
o
n
v
1
D
k
e
r
n
e
l
/
p
a
d
d
i
n
g
/
a
c
t
i
v
a
t
i
o
n
3
/
s
a
me
/
R
e
LU
F
i
x
e
d
B
a
t
c
h
n
o
r
ma
l
i
z
a
t
i
o
n
a
f
t
e
r
e
a
c
h
C
o
n
v
1
D
{o
n
,
o
f
f
}
M
a
x
P
o
o
l
/
d
r
o
p
o
u
t
P
o
o
l
2
/
D
r
o
p
o
u
t
0
.
2
{0
.
1
,
0
.
2
,
0
.
3
,
0
.
4
}
LSTM
h
i
d
d
e
n
u
n
i
t
s
6
4
→3
2
{3
2
,
6
4
,
9
6
}
D
e
n
se
h
e
a
d
D
e
n
se
(
3
2
,
R
e
LU
)
→D
e
n
se
(
6
,
l
i
n
e
a
r
)
F
i
x
e
d
O
p
t
i
mi
z
e
r
/
l
e
a
r
n
i
n
g
r
a
t
e
A
d
a
m
/
1
×
1
0
⁻
³
{1
×
1
0
⁻
³,
5
×
1
0
⁻
⁴
,
1
×
1
0
⁻
⁴}
Lo
ss
M
e
a
n
sq
u
a
r
e
d
e
r
r
o
r
F
i
x
e
d
B
a
t
c
h
si
z
e
64
{3
2
,
6
4
,
1
2
8
}
M
a
x
e
p
o
c
h
s/
e
a
r
l
y
st
o
p
p
i
n
g
5
0
/
p
a
t
i
e
n
c
e
1
0
F
i
x
e
d
LR
s
c
h
e
d
u
l
e
R
e
d
u
c
e
L
R
O
n
P
l
a
t
e
a
u
(
f
a
c
t
o
r
0
.
5
,
p
a
t
i
e
n
c
e
5
,
m
i
n
l
r
1
×
1
0
⁻
⁶)
F
i
x
e
d
R
a
n
d
o
m s
e
e
d
(
TF)
42
F
i
x
e
d
Valid
atio
n
o
n
ly
g
r
id
s
ea
r
ch
ac
r
o
s
s
th
e
s
ea
r
ch
s
p
ac
e
s
u
m
m
ar
i
ze
d
in
T
ab
le
3
(
r
ig
h
t
co
lu
m
n
)
was
u
s
ed
to
ch
o
o
s
e
h
y
p
er
p
ar
a
m
eter
s
.
I
n
p
ar
ticu
lar
,
we
s
ea
r
ch
ed
o
v
er
3
2
,
6
4
a
n
d
9
6
co
n
v
o
lu
tio
n
al
f
i
lter
co
u
n
ts
,
3
2
,
6
4
,
an
d
9
6
L
STM
h
i
d
d
en
u
n
its
,
0
.
1
,
0
.
2
,
0
.
3
,
an
d
0
.
4
d
r
o
p
o
u
t
r
ates
an
d
1
×1
0
⁻³,
5
×1
0
⁻⁴,
1
×
1
0
⁻⁴
lear
n
in
g
r
ates,
test
in
g
ea
ch
o
f
th
e
ca
n
d
id
a
te
m
o
d
el
co
n
f
ig
u
r
atio
n
s
o
n
th
e
1
5
%
ch
r
o
n
o
lo
g
ical
v
al
id
atio
n
s
p
lit
o
n
ly
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
5
6
-
270
262
T
h
e
r
ep
o
r
ted
ar
ch
itectu
r
e
(
6
4
-
6
4
co
n
v
,
6
4
-
3
2
L
STM
,
d
r
o
p
o
u
t
0
.
2
,
an
d
lr
=
1
×1
0
⁻³
)
is
th
e
s
i
m
p
lest
m
o
d
el
t
o
b
e
ju
d
g
ed
b
y
th
e
o
n
e
s
tan
d
a
r
d
er
r
o
r
d
o
wn
cr
iter
i
o
n
.
No
o
t
h
er
p
a
r
titi
o
n
s
o
f
th
e
d
ata
wer
e
u
s
ed
in
th
e
test
p
ar
titi
o
n
o
th
er
th
an
o
n
ce
at
t
h
e
en
d
f
o
r
h
y
p
er
p
ar
am
eter
s
elec
tio
n
.
A
d
r
o
p
o
u
t
s
en
s
itiv
ity
s
wee
p
s
o
v
er
{0
.
1
,
0
.
2
,
0
.
3
,
an
d
0
.
4
}
o
n
th
e
v
alid
atio
n
p
ar
titi
o
n
p
r
o
d
u
ce
d
test
-
eq
u
iv
alen
t
m
ea
n
ab
s
o
lu
te
er
r
o
r
(
MA
E
)
v
alu
es
with
in
a
n
ar
r
o
w
b
an
d
o
f
ap
p
r
o
x
im
ately
1
5
6
–
1
6
3
GB
,
with
th
e
lo
west
d
is
p
er
s
io
n
at
0
.
2
an
d
clea
r
u
n
d
e
r
f
itti
n
g
at
0
.
4
.
T
h
e
f
o
r
ec
aster
is
th
er
e
f
o
r
e
r
o
b
u
s
t
to
s
m
all
p
er
tu
r
b
atio
n
s
in
d
r
o
p
o
u
t
b
u
t
d
o
es
r
eq
u
i
r
e
s
o
m
e
r
e
g
u
lar
izatio
n
to
av
o
i
d
m
e
m
o
r
izin
g
th
e
s
tr
o
n
g
d
iu
r
n
al
p
atter
n
.
T
h
e
in
p
u
t
win
d
o
w
L
=7
2
h
o
u
r
s
was
s
elec
ted
af
ter
also
ev
a
lu
atin
g
L
∈
{2
4
,
4
8
,
9
6
,
an
d
1
6
8
}:
th
e
7
2
-
h
o
u
r
win
d
o
w
ca
p
tu
r
ed
b
o
th
d
i
u
r
n
a
l
an
d
wee
k
en
d
-
wee
k
d
ay
s
tr
u
c
tu
r
e
with
th
e
lo
west
v
alid
atio
n
MA
E
,
wh
ile
th
e
168
-
h
o
u
r
wee
k
ly
win
d
o
w
ad
d
ed
m
ar
g
in
al
ac
c
u
r
ac
y
at
s
u
b
s
tan
tially
h
ig
h
er
p
er
-
b
atch
co
m
p
u
te
co
s
t.
2
.
5
.
Va
ri
a
t
io
na
l
a
ut
o
enco
der
T
h
e
d
etec
to
r
f
o
llo
ws
th
e
VAE
f
o
r
m
u
latio
n
o
f
Kin
g
m
a
an
d
W
ellin
g
[
6
]
an
d
th
e
r
ec
o
n
s
tr
u
ctio
n
-
p
r
o
b
a
b
il
ity
f
r
am
in
g
u
s
ed
in
5
G
an
o
m
aly
d
etec
tio
n
b
y
I
s
la
m
et
a
l.
[
7
]
.
T
h
e
en
c
o
d
er
m
a
p
s
an
8
-
d
im
e
n
s
io
n
al
in
p
u
t
to
a
4
-
d
im
en
s
io
n
al
laten
t
s
p
ac
e
p
ar
am
eter
ized
b
y
m
ea
n
μ
an
d
lo
g
-
v
ar
ian
ce
lo
g
σ
²;
t
h
e
d
ec
o
d
e
r
m
ir
r
o
r
s
th
e
en
co
d
er
.
T
r
ain
in
g
m
in
im
i
ze
s
th
e
s
tan
d
ar
d
ev
id
en
ce
lo
wer
b
o
u
n
d
(
E
L
B
O
)
o
b
jectiv
e
r
ec
o
n
s
tr
u
ctio
n
MSE
p
lu
s
a
KL
-
d
iv
er
g
en
ce
ter
m
a
g
ain
s
t
a
s
tan
d
ar
d
Gau
s
s
ian
p
r
io
r
o
n
n
o
r
m
al
r
ec
o
r
d
s
o
n
ly
.
T
h
e
r
ec
o
n
s
tr
u
ctio
n
-
er
r
o
r
d
ec
is
io
n
r
u
le
th
en
ask
s
,
f
o
r
a
n
y
n
ew
r
ec
o
r
d
x
,
wh
et
h
er
x
lo
o
k
s
lik
e
a
s
am
p
le
d
r
awn
f
r
o
m
th
e
s
am
e
d
is
tr
ib
u
tio
n
as
th
e
t
r
ain
in
g
n
o
r
m
als:
if
x
lies
f
ar
f
r
o
m
th
at
lear
n
ed
m
an
if
o
ld
,
its
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
is
lar
g
e
an
d
th
e
r
ec
o
r
d
is
f
lag
g
e
d
.
T
h
e
th
r
esh
o
ld
τ
is
f
ix
e
d
at
t
h
e
9
9
th
p
e
r
ce
n
tile
o
f
v
alid
atio
n
r
e
co
n
s
tr
u
ctio
n
er
r
o
r
s
,
y
ield
in
g
τ
=0
.
7
9
7
o
n
th
is
d
ata
s
et.
T
h
e
f
u
ll
ar
ch
itec
tu
r
e
an
d
h
y
p
er
p
ar
am
eter
s
ar
e
lis
ted
in
T
ab
le
4
an
d
s
h
o
wn
s
ch
em
atica
lly
in
Fig
u
r
e
4
.
T
ab
le
4
.
VAE
h
y
p
er
p
ar
am
eter
s
C
o
m
p
o
n
e
n
t
S
e
t
t
i
n
g
I
n
p
u
t
d
i
me
n
si
o
n
d
8
La
t
e
n
t
d
i
m
e
n
s
i
o
n
4
En
c
o
d
e
r
D
e
n
se
(
3
2
,
R
e
LU
)
→D
e
n
se
(
1
6
,
R
e
L
U
)
→
(
μ,
l
o
g
σ²
)
∈
ℝ⁴
D
e
c
o
d
e
r
D
e
n
se
(
1
6
,
R
e
LU
)
→D
e
n
se
(
3
2
,
R
e
L
U
)
→
D
e
n
se
(
8
,
l
i
n
e
a
r
)
R
e
p
a
r
a
m
e
t
e
r
i
z
a
t
i
o
n
z=
μ+
σ
⊙
ε
,
ε
∼
(
0
,
I)
O
p
t
i
mi
z
e
r
A
d
a
m (
l
r
=
1
×
1
0
⁻
³)
Lo
ss
M
S
E
r
e
c
o
n
s
t
r
u
c
t
i
o
n
+
K
L
d
i
v
e
r
g
e
n
c
e
(
β=1
.
0
)
B
a
t
c
h
si
z
e
1
2
8
M
a
x
e
p
o
c
h
s/
e
a
r
l
y
st
o
p
p
i
n
g
5
0
/
p
a
t
i
e
n
c
e
4
Th
r
e
s
h
o
l
d
τ
0
.
7
9
7
(
9
9
t
h
p
e
r
c
e
n
t
i
l
e
o
f
v
a
l
i
d
a
t
i
o
n
M
S
E)
R
a
n
d
o
m s
e
e
d
(
TF)
42
Fig
u
r
e
4
.
VAE
f
o
r
u
n
s
u
p
er
v
is
ed
an
o
m
al
y
d
etec
tio
n
.
T
h
e
e
n
c
o
d
er
co
m
p
r
ess
es th
e
8
-
d
im
e
n
s
io
n
al
r
ec
o
r
d
in
to
a
4
-
d
im
en
s
io
n
al
Gau
s
s
ian
laten
t; th
e
d
ec
o
d
er
r
ec
o
n
s
tr
u
cts it.
R
ec
o
n
s
tr
u
ctio
n
er
r
o
r
e(
x
)
is
c
o
m
p
ar
ed
ag
ain
s
t
a
f
ix
ed
th
r
esh
o
ld
τ
d
er
i
v
ed
f
r
o
m
th
e
9
9
t
h
p
er
ce
n
tile o
f
v
alid
a
tio
n
er
r
o
r
s
A
s
wee
p
o
f
th
r
esh
o
ld
s
en
s
iti
v
ity
was
also
p
er
f
o
r
m
ed
o
n
t
h
e
v
alid
atio
n
p
ar
titi
o
n
a
n
d
r
a
n
g
ed
f
r
o
m
9
5
th
to
th
e
9
9
.
9
t
h
p
e
r
ce
n
tiles
.
T
r
an
s
itio
n
s
b
etwe
en
9
5
th
a
n
d
9
9
th
p
er
ce
n
tiles
r
esu
lted
in
a
l
o
s
s
o
f
r
ec
all
(
f
r
o
m
0
.
6
6
to
0
.
5
0
)
b
u
t
an
in
cr
ea
s
e
in
p
r
ec
is
io
n
(
f
r
o
m
0
.
3
1
t
o
0
.
4
9
)
;
w
h
ile
tr
an
s
itio
n
s
to
v
er
y
ex
tr
em
e
9
9
.
9
th
p
er
ce
n
tile
r
esu
lted
in
h
ig
h
p
r
ec
is
io
n
-
lo
w
r
ec
all
r
e
g
im
e
s
u
itab
le
f
o
r
v
er
y
co
n
s
er
v
ativ
e
aler
tin
g
.
T
h
e
9
9
th
p
er
ce
n
tile
s
ettin
g
p
r
esen
ted
in
T
ab
le
4
h
as
b
ee
n
ch
o
s
en
as
a
s
u
itab
le
o
p
er
atin
g
p
o
i
n
t
an
d
to
b
e
co
n
s
is
ten
t
with
th
e
liter
atu
r
e
f
r
o
m
th
e
r
ec
o
n
s
tr
u
ctio
n
p
er
s
p
ec
tiv
e.
T
h
is
s
en
s
itiv
ity
s
wee
p
i
s
ex
ten
d
ed
to
th
e
test
s
et
an
d
th
e
p
r
ec
is
io
n
-
r
ec
all
-
F1
tr
a
d
e
-
o
f
f
is
ex
p
licitly
s
h
o
wn
in
s
ec
tio
n
3
.
2
.
2
.
6
.
I
m
ple
m
ent
a
t
io
n,
e
v
a
lua
t
io
n,
a
nd
ba
s
elines
E
v
alu
atio
n
m
etr
ics
:
f
o
r
ec
asti
n
g
is
r
ep
o
r
ted
as M
AE
(
GB
)
,
r
o
o
t m
ea
n
s
q
u
ar
e
e
r
r
o
r
(
R
MSE
)
(
GB
)
,
an
d
m
ea
n
ab
s
o
lu
te
p
er
ce
n
tag
e
er
r
o
r
(
MA
PE
)
(
%).
Dete
ctio
n
is
r
ep
o
r
te
d
u
s
in
g
p
r
ec
is
io
n
,
r
e
ca
ll,
F1
,
ac
cu
r
ac
y
,
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
C
o
n
vo
lu
tio
n
a
l
n
eu
r
a
l n
etw
o
r
k
a
n
d
lo
n
g
s
h
o
r
t
-
term me
mo
r
y
f
o
r
ec
a
s
tin
g
a
n
d
…
(
R
u
v
a
r
a
s
h
e
C
.
Ho
ve
)
263
th
e
f
u
ll
co
n
f
u
s
io
n
m
atr
i
x
,
p
r
ec
is
io
n
-
r
ec
all
ar
ea
u
n
d
e
r
th
e
cu
r
v
e
(
PR
-
AUC
)
,
an
d
r
ec
eiv
er
o
p
er
atin
g
ch
ar
ac
ter
is
tic
-
ar
ea
u
n
d
er
th
e
cu
r
v
e
(
R
OC
-
AUC
)
.
W
e
also
r
ep
o
r
t
9
5
%
b
o
o
ts
tr
ap
co
n
f
i
d
e
n
ce
in
t
er
v
als
(
5
0
0
iter
atio
n
s
)
an
d
p
air
wis
e
s
tatis
t
ical
test
s
Dieb
o
ld
-
Ma
r
ian
o
[
2
8
]
o
n
s
q
u
a
r
ed
f
o
r
ec
ast
lo
s
s
f
o
r
th
e
f
o
r
ec
aster
,
an
d
Mc
Nem
ar
f
o
r
th
e
d
etec
to
r
a
g
a
in
s
t e
ac
h
b
aselin
e.
R
ep
r
o
d
u
cib
ilit
y
:
all
c
o
d
e
an
d
d
ata
ar
e
r
elea
s
ed
at
h
ttp
s
://g
ith
u
b
.
c
o
m
/r
u
v
a
r
ash
eh
o
v
e
-
s
v
g
/s
m
ar
t
-
rf
-
s
p
ec
tr
u
m
-
zw.
T
h
e
r
ep
o
s
ito
r
y
in
clu
d
es
th
e
d
atas
et,
th
e
lo
ca
lizatio
n
s
cr
ip
t
(
Alg
o
r
ith
m
1
)
,
th
e
p
r
ep
r
o
ce
s
s
in
g
p
ip
elin
e,
th
e
C
NN
–
L
STM
an
d
VAE
tr
ain
in
g
n
o
teb
o
o
k
s
,
b
as
elin
e
r
u
n
n
e
r
s
,
an
d
th
e
ev
al
u
atio
n
h
ar
n
ess
u
s
ed
to
p
r
o
d
u
ce
ev
er
y
tab
le
an
d
f
ig
u
r
e.
T
h
e
C
NN
–
L
STM
an
d
VAE
m
o
d
els
wer
e
i
m
p
lem
en
ted
u
s
in
g
T
en
s
o
r
Fl
o
w/Ker
as
[
2
9
]
.
E
x
p
er
im
en
tal
s
etu
p
:
th
e
h
ea
d
lin
e
f
o
r
ec
aster
o
p
e
r
ates
o
n
th
e
n
etwo
r
k
-
ag
g
r
eg
ate
h
o
u
r
ly
s
er
ies
o
b
tain
ed
b
y
av
er
ag
i
n
g
tr
af
f
ic
-
v
o
lu
m
e
-
GB
ac
r
o
s
s
th
e
1
3
to
wer
s
,
th
r
ee
o
p
er
at
o
r
s
,
an
d
f
iv
e
b
an
d
s
at
ea
ch
tim
estam
p
,
p
r
o
d
u
cin
g
1
2
4
,
4
1
6
s
am
p
les
(
8
7
,
0
9
1
tr
ai
n
,
1
8
,
6
6
2
v
al,
1
8
,
6
6
3
test
u
n
d
e
r
s
tr
ict
c
h
r
o
n
o
lo
g
ical
s
p
lit).
An
o
m
aly
d
etec
tio
n
o
p
er
ates
o
n
th
e
f
u
ll
3
1
5
,
2
4
7
-
r
ec
o
r
d
p
er
-
r
ec
o
r
d
d
ataset
u
s
in
g
a
7
0
/1
5
/1
5
r
a
n
d
o
m
s
p
lit.
T
h
e
C
NN
–
L
STM
is
co
m
p
ar
ed
ag
ain
s
t
n
aiv
e
last
v
alu
e,
s
e
aso
n
al
n
aiv
e
(
2
4
-
h
o
u
r
lag
)
,
A
R
(
2
4
)
R
id
g
e
as
an
AR
I
MA
/
s
ea
s
o
n
al
au
to
r
eg
r
ess
iv
e
in
teg
r
ate
d
m
o
v
in
g
av
er
a
g
e
(
SAR
I
MA
)
p
r
o
x
y
,
an
d
f
ee
d
-
f
o
r
war
d
DNNs
s
ized
96
-
6
4
,
8
0
-
6
4
,
an
d
1
2
8
as
co
m
p
u
te
-
m
atch
e
d
p
r
o
x
ies
f
o
r
L
STM
-
o
n
ly
,
GR
U,
an
d
C
NN
-
o
n
ly
ar
c
h
itectu
r
es.
W
e
co
n
tex
tu
alize
th
ese
r
esu
lts
ag
ain
s
t
p
u
b
lis
h
ed
C
o
n
v
L
ST
M
[
2
2
]
an
d
tr
an
s
f
o
r
m
er
-
b
ased
f
o
r
ec
asti
n
g
r
esu
lts
[
2
3
]
in
s
ec
tio
n
3
.
1
;
f
u
ll
ar
c
h
itectu
r
es
with
eq
u
al
co
m
p
u
te
b
u
d
g
ets
ar
e
r
eser
v
ed
f
o
r
th
e
jo
u
r
n
al
-
v
er
s
io
n
ex
ten
s
io
n
.
T
h
e
VAE
i
s
c
o
m
p
ar
ed
a
g
ain
s
t
s
tatis
tical
3
-
σ
t
h
r
esh
o
ld
in
g
,
I
s
o
latio
n
Fo
r
est
(n
-
esti
m
ato
r
s
=1
0
0
,
co
n
tam
in
atio
n
=
0
.
0
2
)
,
one
-
class
S
VM
(
r
ad
ial
b
asis
f
u
n
ctio
n
o
r
R
B
F
k
er
n
el,
ν
=0
.
0
2
,
o
n
an
8
,
0
0
0
-
s
am
p
l
e
s
u
b
s
am
p
le
f
o
r
tr
ac
ta
b
ilit
y
)
,
an
d
a
v
an
illa sh
allo
w
au
to
e
n
co
d
er
.
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
3
.
1
.
F
o
re
ca
s
t
ing
re
s
ults
T
r
ain
in
g
l
o
s
s
co
n
v
e
r
g
ed
s
m
o
o
th
ly
with
ea
r
ly
s
to
p
p
i
n
g
.
On
th
e
h
eld
-
o
u
t
test
p
ar
titi
o
n
,
th
e
C
NN
–
L
STM
ac
h
iev
es
M
AE
=1
5
8
.
9
7
GB
,
R
MSE
=2
3
0
.
4
1
GB
,
a
n
d
MA
PE=
5
2
.
9
3
%
a
g
ain
s
t
an
ag
g
r
e
g
ate
m
ea
n
lo
ad
o
f
4
2
4
.
2
6
GB
/h
o
u
r
.
T
ab
le
5
r
ep
o
r
ts
th
e
f
u
ll
b
aselin
e
p
an
el.
T
h
e
C
NN
–
L
STM
r
ed
u
ce
s
MA
PE
b
y
ap
p
r
o
x
im
ately
1
9
–
4
6
%
co
m
p
ar
ed
with
th
e
class
ical
b
aselin
es,
in
clu
d
i
n
g
a
1
9
%
r
ed
u
ctio
n
in
MA
PE
r
elativ
e
to
th
e
s
tr
o
n
g
est
class
ical
b
asel
in
e,
AR
(
2
4
)
R
id
g
e;
all
im
p
r
o
v
em
en
ts
o
v
er
class
ical
b
aselin
es
ar
e
s
ig
n
if
ican
t
at
p
<0
.
0
0
1
b
y
p
air
e
d
Dieb
o
ld
–
Ma
r
ian
o
test
[
2
8
]
.
Ag
ain
s
t
th
e
co
m
p
u
te
-
m
atch
ed
f
ee
d
-
f
o
r
wa
r
d
DNN
p
r
o
x
ies
f
o
r
L
STM
-
o
n
ly
,
GR
U
-
o
n
ly
,
a
n
d
C
NN
-
o
n
ly
ar
ch
itectu
r
es,
th
e
C
NN
–
L
STM
i
s
s
tati
s
tically
i
n
d
is
tin
g
u
is
h
ab
le
in
MA
E
b
u
t
ex
h
ib
its
th
e
lo
west
MA
PE
o
f
th
e
d
ee
p
m
o
d
els,
s
u
g
g
esti
n
g
th
at
th
e
h
y
b
r
id
d
esig
n
is
at
leas
t
as
co
m
p
u
te
-
e
f
f
icien
t
as
an
y
s
in
g
l
e
-
f
am
ily
n
eu
r
al
b
aselin
e
at
th
i
s
d
ataset
s
ize.
T
h
is
p
atter
n
is
co
n
s
is
ten
t
with
th
e
C
o
n
v
L
STM
liter
atu
r
e
o
n
m
u
l
ti
-
ce
ll
lo
ad
f
o
r
ec
asti
n
g
[
2
2
]
,
wh
er
e
th
e
co
n
v
o
l
u
tio
n
al/r
ec
u
r
r
en
t
s
p
lit
y
ield
s
its
lar
g
est
g
ain
s
o
n
m
u
lti
-
ch
an
n
el
in
p
u
ts
;
o
n
a
s
in
g
le
n
etwo
r
k
-
ag
g
r
eg
ate
tar
g
et
th
e
m
ar
g
i
n
al
g
ain
is
s
m
aller
.
T
r
an
s
f
o
r
m
e
r
-
b
ased
f
o
r
ec
aster
s
[
2
3
]
wo
u
ld
in
p
r
i
n
cip
le
co
m
p
ete
h
er
e,
b
u
t
th
eir
r
e
p
o
r
te
d
g
ai
n
s
in
th
e
p
u
b
lis
h
ed
ce
llu
lar
liter
atu
r
e
co
m
e
at
s
u
b
s
tan
tially
h
ig
h
er
p
ar
am
eter
a
n
d
co
m
p
u
te
b
u
d
g
ets th
an
th
is
wo
r
k
tar
g
ets.
T
o
u
n
d
er
s
tan
d
th
e
s
ig
n
i
f
ican
c
e
o
f
th
e
MA
PE
v
alu
e
(
5
2
.
9
3
%),
it
is
im
p
o
r
tan
t
to
co
n
s
id
er
th
e
ce
llu
lar
f
o
r
ec
asti
n
g
r
esu
lts
th
at
h
a
v
e
b
ee
n
p
u
b
lis
h
ed
.
T
h
e
MA
PE
v
alu
es
r
e
p
o
r
ted
b
y
W
a
n
g
et
a
l.
[
4
]
f
o
r
a
s
p
atio
tem
p
o
r
al
d
ee
p
m
o
d
el
o
n
ce
ll
-
lev
el
tr
af
f
ic
d
ata
f
r
o
m
a
m
ajo
r
m
etr
o
p
o
litan
d
ataset
f
all
b
etwe
en
3
5
%
an
d
6
0
%
an
d
th
e
af
o
r
em
en
tio
n
ed
s
u
r
v
e
y
b
y
Z
h
an
g
et
a
l.
[
3
]
s
h
o
ws
th
at
MA
PE
v
alu
es
r
an
g
e
f
r
o
m
ar
o
u
n
d
2
5
%
f
o
r
d
en
s
e
ce
ll
-
lev
el
tr
ac
es
in
m
atu
r
e
m
a
r
k
ets
to
m
o
r
e
th
an
7
0
%
f
o
r
ag
g
r
eg
ated
,
l
ess
d
en
s
e
tr
ac
es
in
em
er
g
in
g
m
ar
k
ets.
T
h
is
MA
P
E
v
alu
e
f
alls
o
n
th
e
h
ig
h
en
d
o
f
th
at
r
an
g
e,
wh
ich
is
th
e
r
ea
s
o
n
ab
le
v
alu
e
f
o
r
a
n
etwo
r
k
-
ag
g
r
eg
ate
tar
g
et
f
o
r
a
lo
c
alize
d
s
et
o
f
m
o
d
er
ate
d
e
n
s
ity
.
T
h
e
C
NN
–
L
STM
’
s
1
9
%
MA
PE
r
ed
u
ctio
n
r
elativ
e
to
AR
(
2
4
)
R
id
g
e
is
co
m
p
ar
ab
le
to
t
h
e
im
p
r
o
v
e
m
en
ts
r
ep
o
r
ted
b
y
L
iv
ier
is
et
a
l.
[
5
]
an
d
th
e
g
en
er
alize
d
L
STM
f
o
r
ec
asti
n
g
im
p
r
o
v
e
m
en
ts
r
ep
o
r
ted
b
y
Pr
ater
et
a
l
.
[
1
0
]
.
Fo
r
ec
asti
n
g
in
p
r
ac
tice
:
th
e
C
NN
–
L
STM
h
as a
r
elativ
e
MA
E
o
f
3
7
% a
g
ain
s
t th
e
av
er
ag
e
lo
ad
o
f
4
2
4
GB
/h
o
u
r
.
Fig
u
r
e
5
c
o
m
p
ar
es
t
h
e
p
r
ed
icted
t
o
ac
tu
al
h
o
u
r
ly
tr
af
f
ic
o
n
th
e
test
p
ar
titi
o
n
,
an
d
d
em
o
n
s
tr
ates
an
o
p
er
atio
n
al
co
n
s
eq
u
en
ce
:
t
h
e
m
o
d
el
ca
p
tu
r
es
d
iu
r
n
al
a
n
d
wee
k
ly
s
tr
u
ctu
r
e
well,
th
o
u
g
h
o
cc
asio
n
ally
it
is
ca
u
g
h
t u
n
awa
r
e
at
a
d
r
am
atic
s
u
r
g
e
o
r
d
r
o
p
in
th
e
ac
tu
al
tr
af
f
ic.
Fro
m
th
e
o
p
er
atio
n
al
p
o
in
t
o
f
v
iew,
if
th
e
ce
ll
s
ite
was
ac
tu
ally
tr
an
s
m
itti
n
g
4
0
0
GB
/h
o
u
r
,
th
e
n
a
o
n
e
s
tan
d
ar
d
d
e
v
iatio
n
f
o
r
ec
ast
in
ter
v
al
wo
u
ld
b
e
ap
p
r
o
x
im
ately
2
4
1
-
5
5
9
GB
/h
o
u
r
.
T
h
e
R
MSE
v
alu
e
o
f
2
3
0
GB
(
5
4
%
o
f
th
e
m
ea
n
)
m
ea
n
s
t
h
at
u
n
ex
p
ec
ted
s
u
r
g
es
an
d
/o
r
s
h
ar
p
d
r
o
p
s
i
n
er
r
o
r
ar
e
ab
o
u
t
1
.
5
x
o
f
th
e
ty
p
ical
er
r
o
r
.
He
r
e'
s
h
o
w
it
will
d
ir
ec
tly
af
f
ec
t
ca
p
ac
ity
p
lan
n
in
g
:
2
0
0
+
GB
o
f
u
n
d
er
-
p
r
o
v
is
io
n
in
g
ca
n
b
e
s
ee
n
im
p
ac
tin
g
u
s
er
ex
p
e
r
ien
ce
d
u
r
in
g
ev
en
i
n
g
p
ea
k
s
,
an
d
o
v
er
-
p
r
o
v
is
io
n
in
g
was
te
leased
b
ac
k
h
au
l.
T
h
e
u
s
e
o
f
th
e
f
o
r
ec
aster
in
co
n
ju
n
ctio
n
with
th
e
an
o
m
aly
d
etec
to
r
(
s
ec
tio
n
3
.
2
)
is
ac
tu
ally
d
esig
n
ed
t
o
d
e
al
with
s
u
r
g
e
ca
s
e
as
ev
en
ts
th
at
th
e
f
o
r
ec
aster
u
n
d
er
-
p
r
ed
icts
will
g
en
er
ally
also
in
cr
ea
s
e
th
e
VAE
r
ec
o
n
s
tr
u
ctio
n
er
r
o
r
an
d
b
e
d
etec
te
d
in
th
e
co
m
b
in
e
d
o
u
tp
u
t r
at
h
er
th
a
n
b
e
o
v
er
lo
o
k
ed
b
y
t
h
e
f
o
r
ec
ast alo
n
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
7
2
2
-
3
2
2
1
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
,
Vo
l.
7
,
No
.
3
,
No
v
em
b
er
20
26
:
2
5
6
-
270
264
T
ab
le
5
.
Fo
r
ec
asti
n
g
b
aselin
e
co
m
p
ar
is
o
n
(
n
etwo
r
k
h
o
u
r
ly
s
er
ies,
test
s
et)
M
o
d
e
l
M
A
E
(
G
B
)
R
M
S
E
(
G
B
)
M
A
P
E
(
%)
N
o
t
e
s
N
a
i
v
e
l
a
s
t
v
a
l
u
e
2
7
6
.
0
0
3
8
4
.
4
6
9
8
.
2
1
ŷ
(
t
+
h
)
=
y
(
t
)
S
e
a
so
n
a
l
n
a
i
v
e
(
2
4
h
)
2
1
6
.
4
3
3
2
2
.
4
2
6
6
.
7
7
ŷ
(
t
+
h
)
=
y
(
t
+
h
−
2
4
)
A
R
(
2
4
)
R
i
d
g
e
(
S
A
R
I
M
A
p
r
o
x
y
)
1
7
6
.
3
3
2
4
8
.
0
9
6
5
.
3
0
R
i
d
g
e
α
=
1
.
0
o
n
l
a
s
t
2
4
l
a
g
s
FF
-
D
N
N
(
9
6
-
6
4
)
[
LST
M
p
r
o
x
y
]
1
5
8
.
7
1
2
3
0
.
3
9
5
2
.
7
5
2
-
h
i
d
d
e
n
-
l
a
y
e
r
R
e
LU
M
LP
FF
-
DNN
(
8
0
-
6
4
)
[
G
R
U
p
r
o
x
y
]
1
5
9
.
2
9
2
3
0
.
8
1
5
3
.
0
4
2
-
h
i
d
d
e
n
-
l
a
y
e
r
R
e
LU
M
LP
FF
-
D
N
N
(
1
2
8
)
[
C
N
N
p
r
o
x
y
]
1
6
1
.
1
6
2
3
2
.
0
7
5
4
.
0
3
1
-
h
i
d
d
e
n
-
l
a
y
e
r
R
e
LU
M
LP
C
N
N
–
LST
M
(
t
h
i
s w
o
r
k
)
1
5
8
.
9
7
2
3
0
.
4
1
5
2
.
9
3
H
y
b
r
i
d
1
-
D
C
N
N
+
2
-
s
t
a
c
k
LST
M
Fig
u
r
e
5
.
P
r
ed
ict
ed
v
s
.
a
ct
u
al
h
o
u
r
l
y
n
etw
o
r
k
-
ag
g
r
e
g
a
te
t
r
a
f
f
i
c
o
n
t
h
e
t
est
p
ar
tit
io
n
.
T
h
e
C
N
N
–
L
STM
t
r
a
ck
s
d
i
u
r
n
al
a
n
d
w
ee
k
l
y
s
tr
u
c
tu
r
e
cl
o
s
el
y
;
v
is
i
b
l
e
r
es
id
u
al
er
r
o
r
co
n
c
en
tr
ates
o
n
s
h
a
r
p
s
u
r
g
es
an
d
d
r
o
p
-
o
f
f
s
,
wh
ic
h
is
p
r
ec
is
el
y
th
e
r
e
g
i
m
e
w
h
e
r
e
t
h
e
p
a
r
al
lel
VA
E
d
ete
ct
o
r
(
s
ec
ti
o
n
3
.
2
)
is
i
n
te
n
d
e
d
t
o
r
ais
e
a
f
l
ag
.
Op
er
ati
o
n
a
ll
y
,
th
e
g
a
p
b
etw
ee
n
th
e
p
r
e
d
i
cte
d
an
d
ac
tu
al
c
u
r
v
es s
h
o
w
n
h
e
r
e
co
r
r
es
p
o
n
d
s
to
t
h
e
u
n
d
er
-
o
r
o
v
e
r
-
p
r
o
v
is
i
o
n
in
g
th
a
t
a
ca
p
a
cit
y
p
la
n
n
e
r
w
o
u
l
d
h
a
v
e
t
o
b
u
d
g
et
f
o
r
if
r
el
y
i
n
g
o
n
th
e
f
o
r
ec
ast
al
o
n
e
3
.
2
.
Ano
m
a
ly
det
ec
t
io
n r
esu
lt
s
VAE
tr
ain
in
g
lo
s
s
d
ec
r
ea
s
e
d
s
m
o
o
th
ly
o
v
er
5
0
ep
o
c
h
s
with
a
s
tead
y
tr
ain
–
v
alid
atio
n
g
ap
,
th
e
ex
p
ec
ted
b
eh
a
v
io
u
r
o
f
a
VAE
tr
ain
ed
o
n
th
e
n
o
r
m
al
s
u
b
s
et
o
n
ly
[
7
]
.
At
th
e
v
alid
ati
o
n
-
d
er
i
v
ed
th
r
esh
o
l
d
τ
=0
.
7
9
7
,
th
e
m
o
d
el
p
r
o
d
u
ce
s
p
r
ec
is
io
n
=0
.
4
8
9
,
r
ec
all=0
.
5
0
2
,
F1
=0
.
4
9
5
,
an
d
ac
cu
r
ac
y
=0
.
9
7
9
,
with
co
n
f
u
s
io
n
m
atr
ix
(
T
P=4
9
1
,
FP
=5
1
3
,
T
N=
4
5
,
7
9
6
,
an
d
FN=4
8
8
)
r
ep
o
r
ted
in
T
ab
le
6
.
P
R
-
AUC=0
.
5
3
8
an
d
R
OC
-
AU
C
=0
.
8
6
8
o
v
er
t
h
e
f
u
ll
th
r
esh
o
ld
s
wee
p
.
T
h
e
b
aselin
e
p
an
el
f
o
r
th
e
d
etec
to
r
is
r
e
p
o
r
ted
in
T
ab
le
7
.
Fig
u
r
e
6
s
h
o
ws
th
e
p
er
-
s
am
p
l
e
ab
s
o
lu
te
f
o
r
ec
ast
-
er
r
o
r
d
is
tr
ib
u
tio
n
th
at
m
o
tiv
ates
a
p
ar
allel
d
etec
to
r
.
Fig
u
r
e
7
v
is
u
alize
s
th
e
VAE
r
ec
o
n
s
tr
u
ctio
n
-
er
r
o
r
d
is
tr
ib
u
tio
n
s
f
o
r
n
o
r
m
al
an
d
a
n
o
m
alo
u
s
r
ec
o
r
d
s
an
d
ex
p
o
s
es
th
e
o
v
er
lap
r
e
g
io
n
th
at
lim
its
an
y
r
ec
o
n
s
tr
u
ctio
n
-
o
n
ly
s
co
r
e.
A
cc
u
r
ac
y
at
9
7
.
9
%
is
m
is
lead
in
g
u
n
d
er
th
e
2
.
0
3
%
an
o
m
aly
r
ate:
a
tr
iv
ial
all
-
n
eg
ativ
e
class
if
i
er
wo
u
ld
ac
h
iev
e
9
7
.
9
7
%.
Pre
cisi
o
n
an
d
r
ec
all
a
r
e
th
e
o
p
er
atio
n
ally
r
elev
an
t
m
et
r
ics,
an
d
th
e
p
r
ec
is
io
n
–
r
ec
all
tr
ad
e
-
o
f
f
in
Fig
u
r
e
8
an
d
th
r
esh
o
ld
s
en
s
itiv
ity
c
u
r
v
e
in
Fig
u
r
e
9
ar
e
wh
er
e
an
y
d
ep
lo
y
m
en
t d
ec
is
io
n
s
h
o
u
ld
b
e
m
ad
e
.
T
ab
le
6
.
VAE
an
o
m
aly
-
d
etec
t
io
n
p
er
f
o
r
m
a
n
ce
at
τ
=0
.
7
9
7
M
e
t
r
i
c
V
a
l
u
e
P
r
e
c
i
s
i
o
n
0
.
4
8
9
R
e
c
a
l
l
0
.
5
0
2
F1
-
sc
o
r
e
0
.
4
9
5
A
c
c
u
r
a
c
y
0
.
9
7
9
Tr
u
e
p
o
si
t
i
v
e
s
(
TP)
4
9
1
F
a
l
se
p
o
si
t
i
v
e
s (F
P
)
5
1
3
Tr
u
e
n
e
g
a
t
i
v
e
s (TN
)
4
5
,
7
9
6
F
a
l
se
n
e
g
a
t
i
v
e
s (F
N
)
4
8
8
PR
-
AUC
0
.
5
3
8
R
O
C
-
AUC
0
.
8
6
8
T
ab
le
7
.
An
o
m
aly
-
d
etec
tio
n
b
aselin
e
co
m
p
ar
is
o
n
M
o
d
e
l
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
PR
-
AUC
R
O
C
-
AUC
S
t
a
t
i
st
i
c
a
l
t
h
r
e
sh
o
l
d
i
n
g
(
3
-
σ)
0
.
9
8
3
0
.
3
7
8
0
.
5
4
6
0
.
5
7
8
0
.
9
4
4
I
so
l
a
t
i
o
n
F
o
r
e
st
0
.
1
5
8
0
.
1
8
4
0
.
1
7
0
0
.
1
1
1
0
.
8
3
0
O
n
e
-
c
l
a
ss SV
M
0
.
3
2
6
0
.
6
7
8
0
.
4
4
1
0
.
6
2
9
0
.
9
1
9
V
a
n
i
l
l
a
a
u
t
o
e
n
c
o
d
e
r
0
.
4
9
6
0
.
4
8
0
0
.
4
8
8
0
.
5
2
6
0
.
8
9
7
V
A
E
(
t
h
i
s w
o
r
k
)
0
.
4
8
9
0
.
5
0
2
0
.
4
9
5
0
.
5
3
8
0
.
8
6
8
Evaluation Warning : The document was created with Spire.PDF for Python.
C
o
m
p
u
t Sci
I
n
f
T
ec
h
n
o
l
I
SS
N:
2722
-
3
2
2
1
C
o
n
vo
lu
tio
n
a
l
n
eu
r
a
l n
etw
o
r
k
a
n
d
lo
n
g
s
h
o
r
t
-
term me
mo
r
y
f
o
r
ec
a
s
tin
g
a
n
d
…
(
R
u
v
a
r
a
s
h
e
C
.
Ho
ve
)
265
Fig
u
r
e
6
.
Per
-
s
am
p
le
ab
s
o
lu
te
f
o
r
ec
ast er
r
o
r
d
is
tr
ib
u
tio
n
o
n
t
h
e
test
p
ar
titi
o
n
(
T
h
e
lo
n
g
r
ig
h
t ta
il r
ep
r
esen
ts
th
e
s
u
r
g
e/d
r
o
p
-
o
f
f
r
e
g
im
e
wh
e
r
e
f
o
r
ec
ast an
d
d
etec
to
r
s
ig
n
als ar
e
in
ten
d
ed
t
o
b
e
c
o
m
b
in
e
d
)
Fig
u
r
e
7
.
VAE
r
ec
o
n
s
tr
u
ctio
n
MSE
f
o
r
n
o
r
m
al
v
s
.
an
o
m
alo
u
s
r
ec
o
r
d
s
,
with
τ
=0
.
7
9
7
m
ar
k
e
d
.
T
h
e
s
ep
ar
a
b
le
u
p
p
er
tail c
o
r
r
esp
o
n
d
s
p
r
im
ar
i
ly
to
in
ter
f
er
e
n
ce
an
d
tr
af
f
ic
s
u
r
g
e
ev
e
n
ts
; th
e
o
v
er
la
p
r
eg
i
o
n
n
ea
r
τ
is
d
o
m
in
ated
b
y
s
ig
n
al
d
r
o
p
o
u
t a
n
d
u
n
a
u
th
o
r
ized
r
ec
o
r
d
s
(
cf
.
T
ab
le
8
)
Fig
u
r
e
8
.
Pre
cisi
o
n
-
r
ec
all
cu
r
v
e
o
v
er
a
t
h
r
esh
o
ld
s
wee
p
s
o
n
t
h
e
test
s
et.
PR
-
AU
C
=0
.
5
3
8
Fig
u
r
e
9
.
Pre
cisi
o
n
,
r
ec
all,
an
d
F1
as th
e
th
r
esh
o
ld
τ
v
a
r
ies ac
r
o
s
s
th
e
9
5
th
–
9
9
.
9
th
-
p
e
r
ce
n
tile r
an
g
e
o
f
v
alid
atio
n
r
ec
o
n
s
tr
u
ctio
n
e
r
r
o
r
.
T
h
is
cu
r
v
e
is
th
e
b
asis
f
o
r
th
e
co
s
t
-
s
en
s
itiv
e
o
p
er
atin
g
p
o
in
t
s
elec
tio
n
d
is
cu
s
s
ed
at
th
e
en
d
o
f
s
ec
tio
n
3
.
2
Per
-
class
d
etec
ti
o
n
p
er
f
o
r
m
a
n
ce
is
r
ep
o
r
ted
in
T
ab
le
8
,
wi
th
9
5
%
b
o
o
ts
tr
ap
co
n
f
id
e
n
ce
in
ter
v
als
(
B
=5
0
0
r
esam
p
les)
ad
d
ed
in
r
esp
o
n
s
e
to
a
r
ev
iewe
r
r
eq
u
est.
T
ab
le
9
th
en
p
r
o
v
id
es
a
p
h
en
o
m
e
n
o
lo
g
ical
in
ter
p
r
etatio
n
o
f
wh
y
ea
ch
cla
s
s
is
o
r
is
n
o
t
s
ep
a
r
ab
le
b
y
a
r
ec
o
n
s
tr
u
ctio
n
-
o
n
ly
s
co
r
e.
Fig
u
r
e
1
0
v
is
u
alize
s
th
e
p
er
-
class
r
ec
o
n
s
tr
u
ctio
n
-
er
r
o
r
d
is
tr
ib
u
tio
n
s
ex
p
licitly
,
s
h
o
win
g
h
o
w
i
n
ter
f
er
e
n
ce
r
ec
o
r
d
s
s
it
clea
n
ly
in
th
e
u
p
p
er
tail w
h
ile
u
n
au
th
o
r
ized
r
ec
o
r
d
s
o
v
er
lap
alm
o
s
t e
n
tire
ly
with
th
e
n
o
r
m
al
m
ass
.
Evaluation Warning : The document was created with Spire.PDF for Python.