Geodesy and Geodynamics 2012,3 ( 1) :52 -56
http://www. jgg09. com
Doi:10.3724/SP.J.1246.2012.00052
Signal prediction based on empirical mode decomposition and
artificial neural networks
Wang Y ong1 , Liu Y anping2 and Yang Jing3
1 School of Surveying
& Umd lnfonnation EngiMering, Henan Polyteclmic University, Ji.oozuo 454000, China 2
School of Civil Engineering, Central South University, Changsha 410075, China 3
College of Mining EngiMering, Hebei United University, Tangslw.n 063009 , Chino
Abstract: In view of the usefulness of Empirical Mode Decomposition ( EMD) , Artificial Neural Networks ( ANN) , and Most Relevant Matching Extension ( MRME) methods in dealing with nonlinear signals , we pro-pose a new way of combining these methods to deal with signal prediction. We found the results of combining EMD with either ANN or MRME to have higher prediction precision for a time series than the result of using EMD alone.
Key words: EMD (Empirical Mode Decomposition); ANN (Artificial Neural Networks); MRME (Most Rel-evant Matching Extension) ; IMF (Intrinsic Mode Function) ; endpoint problem; RBF ( Radial Basis Func-tion)
1 Introduction
Empirical mode decomposition ( EMD) is a method of transforming an empirical time series into a few Intrin-sic Mode Function ( IMF) components and a tendency term, which is the final drab and smooth part of the o-riginal sequence['-'l. It is usually applied to deal with some nonlinear or non stationary series. Because of its certain characteristics , such as parallel processing, self adaptivity, self-organization, associative memory,
fault
tolerance , robustness , it is suitable for application to prediction studies.In this paper, we show how to use EMD to decom-pose a simulation signal into several IMF components and a tendency , how to treat the endpoint problem in two ways, how to do signal prediction by using RBF
Received,2012.()2.05; Aooepted,2012.Q2·12
Corresponding author: Tel: + 86-13785091437, E-mail: Wangyongiz@ 126. com
This work was supporteal. by the Notional Natural Scince Foundation of
Hebei Provinoe ( 0201000921)
(Radial Basis Function) neural network for each com-ponent separately, and how to reconstroct the final pre-diction results.
2 EMD and endpoint problem
During the EMD decomposition, the resultant IMF components must meet the following conditions : 1 ) The number of maximum and minimum points and the number of zero-crossing in different directions must be approximately equal; 2) the mean value of the
maxi-mum and minimaxi-mum at any point must be zero. The decomposition process is as follows[3J :
1) For a signal x ( t) , connecting all the maxima
with a 3 -order spline curve to get the upper envelope , and from the minima to get a lower envelope similarly. Generating a new signal by subtracting the mean of the upper and lower envelope from the original signal.
2) Checking whether the new signal meets the a-bove-mentioned basic requirement of IMF, or whether the residual r is a monotonic function. IT not, repeating step1).
No.1
3) At the ead, s(a) is
deeompaeed intD
a time
oe-riel ~
of
Rm....J.er
<4 JMF
O"C!!pp"'IIllllaDd
r(a).
1D
~toaborteD
du:
~p:oee.., a
SD
pa-- i l
defined. Whell
SD
il ~ dwla-ram
•al-ue,
U8lolallr
choeea
tDbe 0.
2 -
0.
3,
the
~ ~eada.
Since thealgorithm
UJM a cubic lplineiDI£rpcWdim>.
..m..a
the IWIIIher rL - ul""" il... Ibm
two.the
pzooeHaho end..
Cubie
lp1iDe
iD-IlllpaloliOD
illtllldu-
...mrwloiah
willopread
to theill-tcmal data
aad.&et tbe
br-frequmcr data,
e~ the~ endpointp!Ob1em.
To
deal with
dlia
prob-1-, ...., apply the
rwlialbaoie fimrtioo _,...)
...m.k
aad the
wa•e£orm
metcbiqmetbdl.
2.1Dlftd
EMD mrt. rJFint
we
lhow how
to apply
EMD
clizectly to a
time -ri ..S (I) by
uaiJII
tDcJboK
inMaiJab
p~.SiDce ill
the mil
woild,
a data ~u.uall.y
ecm-taina noile, 110 we
add
aome
whill& CauMian
noile 11(r)
with SNR
=
10 dBintD
our
simulalioD
oipal,
21r
(2'1r""
(
2'1r )
s(r) =0.7C!08(
10
r) +O.&coa
30
. , +coa
fiJ'
+
I
.:~
53
0.
6oin(::,~>
u(l)
,1=
[5,95]
(1)
...L.:re
s(
a)
.,.,..;m. afthe
faii....U..
mnr-
aadthe
21r (2'1r ) (2"' )
:r
1=O.
1 -
10
a
.~=O. Boo.
30
•
,:r, = -fiJ' •
:r..
=0.61in(:~r)
,n(a)
(Z)
l'igure
I
abowa the
llimulelioD
aigual
aDd
the
DIF
<'"'~•>nenlll aft£rdimct
EMD.
It maybe """" that
IMF1
i1 dzawD
out u
noi.e,111111
the
remailling
compo-nealll
Ullld
tDbe
b&oically the - aa theOCimllpOIId-iag
oridAal
~ eompoz~a~t., a<>ept at du: eDb,o-peeially
after
the third-order.
2.2 EMD with ANN
""""hqd
JW...,.,[
4]
aaed a
flirlP
layer,oin&le
ll.llU1'CIG,ad
linear
neural
~~elw<llk toperatc
twoaclditional
CK·a-eme valuea at
flllda
of dte -wonof
the date.
Inthil
~,....,
uoe a rwlialb&oie &mc1ion _,...)
networlt, aeleet
IS
conaeeutive
data
aampl•
u illput!6o
UlO
100
:
~
~=,l=il=l
:
.
·
:
~
~
~
~
~
~
~
·
~:
:
~
::
:
1
·
.
:
·
·
:n:II:Li
:
~
:
t:
li
:~
:::
I
~
~~
~
~
---TI
111•1
'
I
·
1
11
1
11111
.
4 111:*~W:
:
.
"
"
"
·2 10:w
so
40 ~ 60 10 110oo
uio
'
r
p,nlllllliillltil
~~~~ ~"
2""~' '"
'
"~~.,
.
~~.
'jjUI ,.
.
1111
IiiiiI
0- , , - - ' ' ' ' III\~
-1 10 20 30 40 $0 60 70 10 90 100and die-
12
ump'lee
Illdie
oulpllt,and by
rUera·
tioa i«*al&
60
poupe <I~ympl". We
the
IIC!ect S
(!ll)llpl utcort
umpla
ad fcecl lhem
illtothe
lllllllab
JU:Wal
aetwolk
t.oon-.
A&z
tcstiq diff!ftat~
TWea we
ehooee
the the
~~.etwozkw!W:h
b.
the
hm perf.omw>c>e
to peqlc eoh-ftluN
atboth
end.
of the data.
A. . .
_
in 6pue2. allir
tbe
tre•hn..,t.
the
clill'ero--
betwCCD tbe
DIF compoae«dt lllldlhe
real
slpalIa
algnl6cantly mluced, mdlcaling that the
DeUJal
Del·'ling
tbe
pmlilfm.2. 3 EMD
wllll
MIIMB IDCibol
Rued
oa ...re...,.,.
[5,6],...,
p~lb&
I!IIDIII;ond&o-~ (HRME) mellmd u :li>Dmn:
1)
lfmk
all
mnjmo£n-
left to ,qltlM M1,JI,. ... ,
111
wilh
ounap.mdin1 timeami•
uT
M,T
Jlt, •••,7' •;
mad.
all
oflite
millimumpoint u N
1 ,N
2 , ••• ,N
1ad
tho
~"I timeIND'i.ee
aa 1'111 ,TMZ, •••,
T,.,
mad.
left
point
oflhe
,qnal
us, and lite
~nila-
lle!i•u
r,.
l.ellhe Wlmlf'orm
&om "• to
N,
bo .,,
and
aet ila ~u
N
1
let
lite
~lrom "•
toIll,
bo
~
"'
U:
41Jl,
andN>JI.
2)
Hake
r.~~~ulhe
- - p o i n t
inthe wamozm
lllld 1M
11ft
mo-.e
tothe
ri&l4
toovedap
2'
14tmd
2'
.Ill Icalculate Qlllllld
ilaeozzdatiOD
codlicicatm61jo
3) Take
the
,....efOZIIl
tollteoJKIII"'ing
tothe
lllllli--
eone1ation eoeft'ieient u tho
- 1 n:cenlwave.
Doote tltc
eoneepoading
mayimumnlue
<I M
0 at tb!.time u
M,,
UJd
e.lco•Jate tbe
~between
.11
0aDd M.,....J
demote
tbe diffcreaee •
'I•'l'ben tbe
~~ ... ~to ...
should be
Add to thc
a
cla!a and e:dzll1dtho
actual .... -.elimnto
the
lall.4) Emmel
the
llllhlal
wa .. &vm to theri&l4
oimilady.
Cboooe
0. 9
ar '"""' .. oarn!l.dionCXHJ!Iim..m.o
senw-ly.H
tbe..,...;-t
ia Dill """"• ...., may .td cliD!ally the ...ft1oe
m
three
...a,hbo!iq
fldleme valu .. atthe
Mdpoinll,111 tb!. way, -
C>all ptlhe IMF
llOIIIp"'Wl'" ohOWII in fipm3.
FlomIIIia
qu-
we
may-!hat the dilr.
eDCe ~ee~~
IMF
aompo!MDIIIand
the IIUeojp
ia100
'2 10 40 50 60 10 80 ~ 100
-t
10 100~2
r-pm-
tl.J-Ittfmo!pll (d-.IIlDea) with Ia llD' "'}'1!-
(tlllaliDco) a f t r r -w!1b -doe ri ... !l.d!oallNo.1
~f
.0.) 10 20 JO <40 50 60 70 80 90
100
"1 10 20 30 <40 .50 60 70 80 90
10 20 30 40 50 60 70 80 90
tOo
l
=
Ill IHI':
21'1'~11 p!lttl~?ll!!l!oi:'PI"
:
II
Ill!I~
L
.
:
- -
- - - -
II II!I:·
II
'
ill~?
·I 10 20 30 <40 SO 60 70 80 90
tOo
. , _ 3 Campori-<i .;mo~,;., q,..l ( d....a lmoo) 'lrilh IMF 1 " , ..
(llliD liDoo) . . - 'lrilh tbo 111110 ...
Reamlllttllmb ,. ,. rm !lldlud DID
0.719 0.955 0.953 0.961 0.984 0,!)76 IMF4 0.!12S 0.983
O.Sl&S
0.430.97
0.962IIDI4IIller
thaD
in fi&un
1 ,
mel
aho
the u.ld linM fit
helt.ar.
'fbi. ..,_,.. the
<llfeoti'l'lln-«
tiWI
method
in
-1-rinJ
the
fmdpoimprcblrml.
Tllllk 2 Oaopilwa d -ollllllrwol
.,.
...
AJ!.r
the ""'....!an by tbe cllfl'eren?:
methoda
ami. """""'"IIIUclilla
ci
tbe pm!ieted val11e11
of each
coapmc:Dt,cllfl'creat
mulb
ue
~A.
thown in IDle 2,
che
belt
r.ult
il
olui-'
by che
ANN
me!hod
wilh an
mean cllfl'crenae
«
Cll!y
0. 45.
Prediction
&emthe
EMD deomq~.Gtico....Wt
il
ri~UifioantlyLeila'
than
dJD
prediatina
of the
dinHJt....Will,
al1lt4>111h '""""
iodi--.idual emn (ouch u poiDI9) ..., Imp.
'lite,...;,.
1!1111111 cmJI'
in the
rmJll.
of lhe
twomc:tbod8
oJoe.
DOlcsceed
2.
+1
whm:a the maimlllll
eD:or<Jf the dizo:et
pmlictico
~3.17.
1 1 3 4s
6 7a
9 10 11 12. Rllll2.1
1.156 Z.OOil 3.3&9 2.11711.656
1.6111.186
1.445 2,15 1.lilll 3.171 2.0730.164
0.3$6
0.219
0.136
o.m
0.271
0.319
0.455 1.06 1,180.709
0.153
0.45 O.QS3o.
006 0.:101 0.133 0.159 0.515 0.991 0.84S 1.883 1,333 1.771 1.137 0.1364
Conclusion
The neural network extension and the most relevant match extension methods are both good solutions to the endpoint-effect problem. EMD decomposition can sup-ply input variables with higher quality to the RBF neu-ral network. The new prediction method presented here can achieve higher precision.
References
[ 1 ] Huang N E , Shen Z , Long S R , et al. The empirical mode
de-composition and the Hilbert spectrum for nonlinear non-stationary time series analysis. Proceedings of the Poyal. Society A, 1998 , 454(1971)' 903-995.
[2] Zhen Zuguang and Liu Hongli. The empirical mode decomposition
[3]
and wavelet analysis and application. Beijing: China
Meteorologi-cal Press, 2010, 1 -83. (in Chinese)
Zhu Jinlong and Qiu Xiaohui. Application with orthogonal
polyno-mial algorithm in the empirical mode decomposition endpoint problem. Computer Engineering and Application, 2006, 23 ( 2) : 72
-74. (in Chinese)
[ 4] Deng Yongjun, Wang Wei, Qian Chengchun, et al. Treatment the
boundary problem in the empirical mode decomposition and the
Hilbert transfonn. Science Bulletin, 2001 , 46 ( 3) : 257 - 263.
( in Chinese)
[5] Shao Chenxi, Wang Jian, Fan Jinfeng, et al. An adaptive
exten-sion method of the empirical mode decomposition endpoint. Elec-tronics Journal, 2007, 35 (10) : 1944 -1948. (in Chinese) [ 6] Wang Ting, Yang Shenyuan and li Bingbing. A new method to
improve the empirical mode decomposition endpoing effect. Harbin Univemity of Science Journal, 2009, 14 (5) : 23 -26. (in Chi-nese)
[7] Dong Changhong. Matlab neuml network and application. Beijing: National Defence Industry Press, 2005: 121 -142. (in Chinese)