基于逆积分方程的正弦波频率估计
第35卷第10期
2007年10月
华南理工大学(自然科学版)
JournalofSouthChinaUniversityofTechnology
(NaturalScienceEdition)
VO1.35No,10
October2007
ArticleID:1000-565X(2007)10—0147—05
SinusoidalFrequencyEstimationBasedonInversionIntegralEquation
WeiLiDao—yuanChenFang-jiong
(SchoolofElectronicandInformationEngineering,SouthChinaUniv,ofTech.,Guangzhou510640,Guangdong,China)
Abstract:Intraditionalfrequencyestimationalgorithms,thereexistsageneralproblemarisingfromthediscre—
pancybetweenthecostlycomputationandtheunsatisfactoryperformanceunderlowSNRs(Signal—to—NoiseRati—
os).Inordertosolvethisproblem,thispaperproposesanewfrequencyestimationalgorithmbasedonthelIE
(InversionIntegralEquation)approach,Inthisalgorithm,FFT(FastFourierTransform)isusedtoroughly
estimatethefrequency,basedonwhichanlIEisconstructedaccordingtothenalTOW—
bandsignalresultingfrom
FFT.Then.thefinefrequencyisestimatedbycalculatingtheparametersandcharacteristicfrequencyofthe
lIE.Simulatedresultsshowthattheproposedalgorithmisofgoodperformanceandmoderatecomputationalre—
quirementunderlowSNRs.
Keywords:frequencyestimation;integralequation;characteristicfrequency
CLCnumber:TN911Documentcode:A
0Introduction
Frequencyestimationisaclassicsignal—processing
problemthathasattractedresearchers'attentionforde—
cades.Ithasapplicationsinawiderangeofareassuch asradar,wirelesscommunicationsandspeechsignal processing.Thispaperaddressestheproblemofsinu—
soidalfrequencyestimationmodeledby
x(t)=CCOS(2t+)+n0(t)(1)
whereC,fc,anddenoterespectivelytheamplitude, frequencyandinitialphaseofthesinusoidalsignal, andn0()istheadditivenoise.Forderivationsimpli—
city,inEq.(1)weconsiderasingle—tonecase.We
willextendtomultiple—tonecaseintheseque1.Theo—
retically,thefrequencyfccanbedirectlyestimated fromtheFourierTransformofx(t).However.ade—
tailedperformanceanalysisshowsthattheFourier TransforiBmethodsuffersfromlowresolutionwhenonlv Receiveddate:March1,2007
木Foundationitem:SupportedbytheNationalNatureScience FoundationofChina(60402014,60625101)
Biography:WeiGang(bornin1963),male,professor,Ph.D. tutor,mainlyresearchesoncommunicationsystems.E—mail:
ecgwei@scut.edu.an
asmallsamplesetofx(t)isavailable….
Alotoffrequencyestimationmethodshavebeen putforwardwhichestimatethefrequencynotdirectly fromtheFourierspectrum,butfromsomesuitablepa-
rameters.SofartheoreticallytheML(MaximumLikeli—
hood)methodisknowntoshowthebestperforlnance. However,itneedsalotofcomputations[.Othersub—
optimalmethodsareeitherdeterministic—based[一4]or
statistics—basedRef.[5—6].Thestatistics—basedme—
thods,whichusuallyexploitthesecond—ordercorrelation
matrix,canachievestableestimation.butt}leyalso exhibit
bound
Bound
significantperformancegaptotheperformance i.e.,theso—calledCRB(Cramer—Rao
[7].Thedeterministicapproachesdonotneed toestimatethestatisticsand,therefore,usuallyrequire asmallerdataset.Forinstance.theLPfLiBearPre—
diction)deterministic—basedapproachcanachievethe
performanceboundwhenonly10samplesareavailable (howeveritsperformancebecomesworsethantheCRB whenalargedatasetisavailable).Recently,Fou—
tierspectrum—basedestimatorswerere—examined[驯.
whichmayobtaintheperformanceboundatmediumto highSNRs(Signal—to—NoiseRatios).KayandSaha
haveproposedacomputation—reducedve~ionofthe
,
,,
148JournalofSouthChinaUniversityofTechnology(NaturalScienceEdition)Vol_35
MLalg0rithm[.
Ab0utani0sandMulgrewhavepre.
sentedamethodthatattainstheperformanceboundat
lowSNRsinthesingletonecase[.
Generallyspeaking,currentfrequencyestimation methodsencountermainlytwoproblems:eitherthe performancedropsdramaticallyunderlowSNRs, orthe
computationalrequirementistoocostly.Inthispapera newfrequencyestimationmethodisproposedbasedon theIIEapproach.Firstly,theFourierTransformis appliedtogiveacoarsefrequencyestimate.Then,an IIEisconstructedbasedonthenarrow.bandsignalre. constructedfromtheFourierspectrumcenteredatthe estimatedcoarsefrequency.Finally.thefrequencyis estimatedbycalculatingthecharacteristicfrequencyof theIIE.Experimentalresultsshowthatournewalg0. rithmhastheadvantagesofboththeMLmethodand theparametricmethods.Itisofgoodperformancean- derlowSNRsatthecostofmoderatecomputati0ns. 1TheProposedMethod
Thetheoreticalbasisofouralgorithmisthatany sinusoidalsignalx(t)isasolutionofthefollowingin. tegralequation
()+口,f.
()d+口f.
J一()dd:0JJ(2)一?一?J一?,
wherea1,a2arerelatedwiththefrequenciesinthewav thatthefrequenciescanbecalculatedfr0mtheimagi. narypartoftherootsofitscharacteristicequati0nas 1+61s一+口2s一=0(3)
Soifwecanconstructanintegralequationfromthe sampledataset,thatis,theIIE.wecanestimatethe
frequencyfromitscharacteristicequationas :T~—4ac2一口(4)
where=2istheangularfrequency.Itsh0uldbe remarkedthattheoreticallyadifferentialequati0n. in.
steadoftheintegralequationshowninEq.
(2).c0uld
alsobeappliedtodescribethesinusoidalsigna1.H0w. 0er,inpractme.thesampledsignaliscorruptedby additivenoise,whichisusuallymoresensitivetothe differentialoperationthantheintegral0Derati0n .
Therefore,weprefertheintegralequationmode1 .The
keypointofouralgorithmistofindanefficientwayto constructtheIIE,i.e.toestimatea】,a,.Thisprob.
1emissolvedintwosteps.
First,applyingFourierTransformtoEq.(1),we canget(),thespectrumofx(t).Fromtheprinci. pieofMLmethod,weknowthatthesinusoida1fre. quencycanbedeterminedbyfindingthepeakvalue pointof()ifthefrequencyresolutionandtheSNR ishighenough.Whenonlyasmallsampledatasetis available,whichisusuallythecaseinpractice.weatta. chtheavailablesampleswithzerostomakethelength apowerof2,soastocomputeaFFT.AfterF不.
therewillbeanarrowbandenvelopethatincludesthe frequencypoint.Thoughinthiscasewecannotaccu. ratelyobtainthefrequency,wecanlocatetherangeof thefrequencypoint.Inotherwords
,wecangetana.
rrowbandsignalcontainingthesinusoidalsignalasis shownbelow:
l()=Acos(moo0t+咖)(5)
mrno—Am
where0isthefundamentalfrequencyoftheFourier Transform;m0denotesthepositionofthepeakva1ue: Amindicatesthefrequencyrange;Aanddenote theamplitudeandthephaseofthem.thharmonicsig. nal,respectively.Eq.(5)canalsobeinterpretedbv Fourierseries.Letx(t)denotethezer0.attached x(t).Performtheharmonicdecompositiononthepc- riodicalextensionof(t).Thespectrumlinesinthe frequencydomainwillhaveanenvelopecenteIledat m00,whichistheclosestspectrumlinetotherea1 value.InEq.(5)weinfactusethespectrumlines 盯."ndetoapproximate(t)?Wehavethefollowing remarksontheselectionofAmandm0.
Remark1:Letf,Tdenotethelengthsofft) and(,)respectively.(,)canbepresentedasf,):
x(t)w(t)(0?t<)wherew(,)istherectangular
window.Calculatingtheharmonicdecompositi0nof(,)
(0?,<T)withrespectto0=2rr/T,wehave c=_J01Tccos(,+咖.
)edc.
—
wocos(w—
oT)-j(mw0sin(wcT)+//20)0)
Lc一//2090c+B/co0).ej咖c
(6)
Fortheclosestspectrumlineto~o
c,
i.e.themnth
spectrumline,wehavef一m00f?0/2.N0te
thatwealsohaveI.一(m0?4)0I~9oJ0/2.Fr0m
Eq.(6),itiseasytoverifythattheenergy0fany spectrumlinedeclinesmorethan20dBwhenitsdis.
No.10WeiGangetal:SinusoidalFrequencyEstimationBasedonInversionIntegralEquatio
n149
tancetothemothspectrumlineisgreaterthan4atte' nuate.Therefore.itisarguedthat(t)canbeappro? ximatedbythespectrumbetween,一4tomo+4.Itis
0bviousthatthemainlodeofthespectrumofthewin. dowsigr~w(t)hasabandrangebetween[一o~T/r,+
o~T/r],thuswearguethat贾(t)canbesufficientlyap?
proximatedbythespectrumlinesinthemainlode, whichapp.roximatelyhasabandwidthof[(mo一4一
T/r)o9o,(mo+4+T/v)?o].Therefore,itissuffi?
cienttochooseAm=4+r.Notealsothattheaddi. tivenoiseeffectstheselectionofAm.Innoisycondi? tions.alargerAmresultsinmoreaccuratesinusoidal signals.However,insuchcases,morenoisesarealso included,whichwilldegradetheestimationaccuracy. SothechoiceofAmisatradeoffbetweentheenclosed usefulsignalandthenoise.Asdiscussedearlier,itis suggestedthatAminT/r?Am?4+T/rbeselected
inpractice.
Remark2:mocanbeselectedasthespectrum
linethathasthelargestamplitude.However,inlow
SNRcasestheremaybeafakepeakduetothenoise. Itisthusproposedthattheenergyinamovingwindow oflength2Am—lbedetectedinthefrequencydo? main.Thecenterofthewindowwiththelargestenergy isselectedasmo.
Nowwewouldliketoshowhowtoestimatea1,a2 from1(t).Denote
t
(7)
zeromeans,from
mo+?mA
:()=丽/'Imsin(,+咖m)
蔓cos()(8)
(9)
ThenextstepistousetheLS(LeastSquare)method toconstructanIIEbasedonthesamplesof1(t),(t) and"1(t),denotingas1(n),1(n),andx1(n), respectively.Let
e(n)=1(n)+a1(n)+a2x(n)(10)
(n=0,l,…,?一1)
whereNisthelengthofthesampledataset.Thepa? rametersa1,a2canbedeterminedbyusingtheLS metIlod.Let
?一1
E=?e2(n)n;0
(11)
BysettingOE/Oa1=0,OE/Oa2=0,wehavethefo' llowinglinearequationsystem,
f<1(n)x:(n)>+a1<:(n)x:(n)>+ j.zl(n):(n)=0fl2)I<1(n)x(n)>+01<:(n)x(n)>+,
【a2<Xl(n)(n)>=0
?一1
where<"(n)(n)>:?"(n)(n).FromEq.(12)n=o
we
caneasilyobtaintheestimatesofa1,a2.Witha1,a2 found.theIIEcanbedetermined,andthustheesti? matedfrequencycanbecalculatedbyEq.(4). Theproposedsingle?tonealgorithmcanbeeasily extendedtomultiple?tonecases.Inmultiple?toneca? ses,firstlyweneedtolocatemultiplepeaksinthe spectrum,thenfor"eachpeakweconstructanarrow? bandsignalshownasEq.(5),fromwhichwebuildan IIEandestimatetheassociateda1,a2.Inthiswayall thefrequenciescanbeestimatedseparatelybasedon Eq.(12).
2SimulationResuits
Theproposedalgorithmwasevaluatedandeom? paredwiththeCRB(Cramer?RaoBound).TheMSE (MeanSquareError)ofestimationisdefinedas MSE=E[(f一)](13)
wherefistheestimateandfistherealfrequency.It hasbeenshownthattheestimationerrorisbounded by[:
(14)
wehaveset
thesamplingintervalto1.Theproposedalgorithmis comparedwiththeLP(LinearPrediction)basedapp?
. roach[-]andtheESPRITalgorithm[引
whicharethe
representativealgorithmsofthedeterministicapproa?
chesandthestatistics?basedapproaches.Bothsingle? tonecasesandmultiple?tonescaseswereinvestigated. NotethattheLPapproachinRef.[3]isforsingle. tonecaseswhiletheapproachinRef.[4]isformulti. pie?tonecases.InRef.[9]aFFT?basedalgorithmfor single?tonefrequencyestimationisapplied,which
No.10WeiGangetal:SinusoidalFrequencyEstimationBasedonInversionIntegralEquatio
n151
WhencomparedwiththetraditionalFFTmethod, theproposedalgorithmrequirestheextracomputation ofEq.(8),(9)andEq.(12),whichisminorwhen comparedwiththecomputationofFFT.Therefore.itis concludedthatthenewalgorithmachievessatisfactory performancewithacceptablecomputation.
3Conclusions
Anewfrequencyestimationalgorithmbasedon IIEhasbeenproposed.Wemodeledthesinusoidbyan integralequationandalternativelyestimatetheequation coefficients.AfterperformingtheFourierTrans~rmon theavailablesamples,anarrow—bandsignalextracted
fromthespectrumisappliedtoestimatetheequation coefficientsfromwhichthefrequencyiscalculated.Its extensiontomultiple—tonecasesisalsoconsidered.
Simulationresulthasverifiedtheefficiencyofthepro—
posedalgorithm.
References:
[2]
[3]
[4]
[5]
[6]
[7]
[8]
KaySM,JrMarpleSL.Spectrumanalysis—Amodern
peIspective[J].Pr0ceedi"g.ftheIEEE,1981,69(11):[93 1380.1419.
StoicaP,MosesRL,FriedlanderB,eta1.Maximumlikeli—
hoodestimationoftheparametersofmultiplesinusoids fromnoisymeasurements[J].IEEETransAcoust, Speech,SignalProcessing,1989,37(3):378-392. SoHC.ChanKW.Reformulationofpisarenkoharmonic decompositionmethodforsingle—tonefrequencyestimation
[J].IEEETransSignalProcessing,2004,52(4):1128—
1135.
SoHC,ChanKitWing,ChanYT,eta1.Linearprediction approachforefficientfrequencyestimationofmultiplereal sinusoids:algorithmsandanalyses[J].IEEETranson SignalProcessing,2005,53(7):2290—2305.
RoyR,PaulrajA,KailathT.ESPRIT:Asubspacerotation approachtoestimationofparametersofcisoidinnoise [J].IEEETranAcoustics,SpeechandSignalProc, 1986,34(5):1340—1342.
MahataKaushik.Subspacefittingapproachesforfrequen- cyestimationusingreal—valueddata[J].IEEETranson
SignalProcessing,2005,53(8):3099-3110. RifeDC,BoorstynRR.Singletoneparameterestimation
fromdiscrete.timeobservations『J].IEEETransInform Theory,1974,20(5):591-598.
KaySteven,SahaSupratim.Meanlikelihoodfrequencyes—
timation[J].IEEETransonSignalProcessing,2000,48
(7):1937-1946.
AboutaniosElias,MemberIEEE,MulgrewBernard.hera.
tivefrequencyestimationbyinterpolationonFouriercoe—
fficients[J].IEEETransonSignalProcessing,2005,53
1242. (4):1237—
基于逆积分方程的正弦波频率估计
韦岗李道远陈芳炯
(华南理工大学电信学院,广东广州510640)
摘要:频率估计算法的普遍问题是计算量大并且在低信噪比时性能较差.文中提出一种基于逆积分方程
(InversionIntegralEquation,IIE)的频率估计新算法.首先利用快速傅立叶变换得到频率的粗估计,并从傅立
叶变换中提出一个窄带信号建立积分方程.然后通过对积分方程中参数和特征频率的估计得到最终的频率
估计.仿真结果显示文中算法以适中的计算量在低信噪比下达到了较好的性能. 关键词:频率估计;积分方程;特征频率
中图分类号:TN911文献标识码:A
文章编号:1000-565X(2007)10—0147—05
责任编辑:李嘉
收稿日期:2007.03.01
基金项目:国家自然科学基金资助项目(60402014,60625101) 作者简介:韦岗(1963一),男,教授,博士生导师,主要从事信息系统研究.E-mail:ecgwei@.t..d..