[雙語(yǔ)翻譯]--（節(jié)選）外文翻譯--外文翻譯--一種新的自動(dòng)調(diào)制識(shí)別的方法（英文）

1、AppliedSoftComputing12(2012)453–461ContentslistsavailableatSciVerseScienceDirectAppliedSoftComputingjournalhomepage:www.elsevier.com/locate/asocAnovelmethodforautomaticmodulationrecognitionAtaollahEbrahimzadeShermeFacult

2、yofElectricalandComputerEngineering,BabolUniversityofTechnology,Babol,Irana r t i c l e i nf oArticlehistory:Received15September2009Receivedinrevisedform23December2010Accepted14August2011Availableonline30August2

3、011Keywords:ModulationrecognitionPatternrecognitionBeesAlgorithmHierarchicalsupportvectormachinebasedclassifierCombinationofthehigherordermoments(uptoeighth)andhigherordercumulants(uptoeighth)Spectralcharacteristicsab s

4、 t r a c tAutomaticrecognitionof the digitalmodulationplaysan importantrole in variousapplications.Thispaperinvestigatesthe designofan accuratesystemfor recognitionof digitalmodulations.First,itisintroducedan

5、 efficientpattern recognitionsystemthat includestwo mainmodules:the featureextractionmoduleand the classifiermodule.Feature extractionmoduleextractsa suitablecombinationof the higherordermomentsup to eighth,

6、higherorder cumulantsup to eighthand instantaneouscharacteristicsofdigitalmodulations.These combinationsofthe featuresare appliedfor the firsttimein this area.Intheclassifiermodule,two importantclassesof supe

7、rvisedclassifiers,i.e., multi-layerperceptron(MLP)neuralnetworkand hierarchicalmulti-classsupportvector machinebased classifierare investigated.Byexperimentalstudy, we choosethe best classifierfor recognitionof

8、 the consideredmodulations.Then,weproposea hybridheuristicrecognitionsystemthat an optimizationmodule is addedtoimprovethegeneralizationperformanceof the classifier.In thismodulewe have useda new optimizationa

9、lgorithmcalledBees Algorithm.This moduleoptimizesthe classifierdesignby searchingfor the best value of theparametersthat tuneits discriminantfunction,and upstreamby lookingfor the best subsetof featurest

10、hatfeed the classifier.Simulationresults show that the proposedhybridintelligenttechniquehasveryhighrecognitionaccuracyevenat low levels of SNR with a littlenumberof the features.©2011ElsevierB.V. All

11、 rightsreserved.1.IntroductionAutomaticmodulationrecognitionisatechniquethatrecog-nizesthetypeofthereceivedsignalatthereceiver.Itplaysanimportantroleinmilitaryandcivildomains.Forexample,inmil-itaryapplications,itcanbeemp

12、loyedforelectronicsurveillance,interferencerecognitionandmonitoring.Thewiderangeofcivil-ianapplicationsincludesspectrummanagement,networktrafficadministration,signalconfirmation,softwareradios,intelligentmodems,cognitive

13、radio,etc.Duetotheincreasingusageofdigi-talsignalsinnoveltechnologiessuchassoftwareradio,therecentresearcheshavebeenfocusedonidentifyingthesesignaltypes.Generally,digitalsignaltypeidentificationmethodsfallintotwomaincate

14、gories:decisiontheoretic(DT)methodsandpat-ternrecognition(PR)methods.DTmethodsuseprobabilisticandhypothesistestingargumentstoformulatetherecognitionprob-lem[1–3].ThemajordrawbacksofDTmethodsaretheirtoohighcomputationalco

15、mplexity,lackofrobustnesstothemodelmis-matchaswellascarefulanalysisthatarerequiredtosetthecorrectthresholdvalues[4]. PRmethods,however,donotneedsuchcarefultreatment.Theyareeasytoimplement.PRmethodscanbefurtherdividedint

16、womainsubsystems:thefeatureextractionsubsystemE-mailaddress:abrahamzadeh@gmail.comandtheclassifiersubsystem.Theformerextractsthefeatures(e.g.histograms,spectralcharacteristics,instantaneouscharacteristics,combinationofse

17、condandfourthordermoment,symmetry,etc.),andthelatterdeterminesthemembershipofsignal(e.g.neuralnetworks,K-nearestneighbor,fuzzylogicclassifier,etc.)[4–19].Fromthepublishedworks,itappearsclearthatinthedesignofasystemforaut

18、omaticrecognitionofdigitalsignaltype(modula-tion),therearesomeimportantissues,which,ifsuitablyaddressed,leadtothedevelopmentofmorerobustandefficientrecognizers.Oneoftheseissuesisrelatedtothechoiceoftheclassificationappro

19、achtobeadopted.Literaturereviewshowsthatdespiteitsgreatpotential,theapplicationofdifferentsupervisedclassifierhasnotreceivedtheattentionitdeservesinthemodulationclassifi-cation.Therefore,inthispaperweinvestigatedtheperfo

20、rmancesofmulti-layerperceptronneuralnetwork(MLP)[20],andsupportvectormachine(SVM)[21,22]. Inthispaper,we haveusedtheSVMsinthestructureoftheproposedhierarchicalclassifier.Choosingtherightfeaturesetisstillanotherissue.Int

21、hispaper,asuitablesetoftheinstantaneouscharacteristics,thehigherordermomentsuptoeighthandthehigherordercumulantsuptoeightharepro-posedastheeffectivefeatures.Turningbacktothedigitalsignalrecognitionsystems,itisfoundthat:(

22、1)featureselectionisnotperformedinacompletelyautomaticwayand(2)theselectionofthebestfreeparametersoftheadoptedclassifieraregenerallydoneempirically(modelselectionissue).Anotherissuethatisaddressedinthispaperisoptimizatio

23、n.Inthismodulewehaveusedanew1568-4946/$–seefrontmatter©2011ElsevierB.V.Allrightsreserved.doi:10.1016/j.asoc.2011.08.025A.E.Sherme/AppliedSoftComputing12(2012)453–461455wheremisthemeanoftherandomvariable.Thedefinitio

24、nfortheithmomentforafinitelengthdiscretesignalisgivenby:?i =N ?k=1(sk ?m)if(sk)(7)whereNisthedatalength.Inthisstudysignalsareassumedtobezeromean.Thus:?i =N ?k=1si kf(sk)(8)Next,theauto-momentoftherandomvariablemaybedefin

25、edasfollows:Mpq =E[sp?q(s?)q](9)wherepiscalledthemomentorderands* standsforcomplexconjugationofs.Assumeazero-meandiscretebased-bandsignalsequenceoftheformsk =ak +jbk.Usingthedefinitionoftheauto-moments,theexpressionsford

26、ifferentordersmaybeeasilyderived.Forexample:M83 =E[s5(s?)3]=E[(a+jb)5(a?jb)3]?M83 =E[(a5 +j5a4b+j210a3b2 +j310a2b3 +j45ab4 +j5b5)(a3 ?j3a2b+j23ab2 ?j3b3)]?M83 =E[a8 +j2a7b?j22a6b2 ?j36a5b3 +j560a3b5 +j62a2b6 ?j72ab7 ?j8b

27、8]?M83 =E[a8 +2a6b2 ?2a2b6 ?b8](10)Considerascalarzeromeanrandomvariableswithcharacter-isticfunction:? f(t)=E{ejts} (11)ExpandingthelogarithmofthecharacteristicfunctionasaTaylorseries,oneobtains:log ? f (t)=k1(jt)+·

28、··+ kr(jt)rr! +···(12)Theconstantskr in(12)arecalledthecumulants(ofthedistri-bution)ofs.Thesymbolismforpthorderofcumulantissimilartothatofthepthordermoment.Morespecially:Cpq =Cum[s,...,s ? ?? ?

29、(p?q)terms,s?,...,s? ? ?? ?(q)terms](13)Forexample:C81 =Cum(s,s,s,s,s,s,s,s?)(14)Thenthordercumulantisafunctionofthemomentsofordersupto(andincluding)n.Momentsmaybeexpressedintermsofcumulantsas:M[s1,..,sn]= ??vCum[{sj}j

30、∈ v1]...Cum[{sj}j ∈vq](15)wherethesummationindexisoverallpartitionsv=(v1,...,vq)forthesetofindexes(1,2,...,n),andqisthenumberofelementsinagivenpartition.Cumulantsmaybealsobederivedintermsofmoments:Cum[s1,...,sn]= ??v(?1

31、)q?1(q?1)!E???j∈v1sj?? ...E???j ∈vqsj??(16)wherethesummationisbeingperformedonallpartitionsv=(v1,...,vq)forthesetofindices(1,2,...,n).Wehavecomputedallofthehigherorderfeaturesforthedigitalcommunicationsignalsthatarecons

32、idered.Table1showsthesomeofthetheoreticalvaluesofthehigherorderstatisticsforanumberoftheconsidereddigitalsignaltypes.ThesevaluesarecomputedTable1Someofthehigherorderfeaturesforanumberoftheconsidereddigitalsignaltypes.PSK

33、2QAM16QAM64M41 100M61 1 ?1.32?1.3M84 13.133.9C61 162.081.797C80 ?244?13.99?11.5C84 ?24417.380undertheconstraintofunitvarianceinnoisefreeandnormalizedbytheoreticalsignalpower,i.e.,thesevaluesareobtainedassumingthesignalis

34、cleanandofinfinitelength.However,inpracticesig-nalsareusuallysubjecttosometypeofdistortion,eitherinsidethetransmitterorduringtransmission,andareoffinitelength.Fig.1showsoneofthehigherorderfeaturesforanumberoftheconsidere

35、ddigitalsignaltypes.3.ClassifierInthispaper,twoimportantsupervisedclassifiershaveused.Followingphrasesdescribesbrieflytheseclassifiers.3.1.MLPneuralnetworkAnMLPneuralnetworkconsistsofaninputlayer(ofsourcenodes),oneormore

36、hiddenlayers(ofcomputationnodes)andanoutputlayer[21].TheissueoflearningalgorithmanditsspeedisveryimportantforMLP.Oneofthemostpopularlearningalgo-rithmsisbackpropagation(BP)algorithm.Howeverundercertainconditions,theBPnet

37、workclassifiercanproducenon-robustclas-sificationresultsandeasilyconvergetoalocalminimum.Moreoveritistimeconsumingintrainingphase.Inrecentyears,newlearningalgorithmshavebeenproposedforneuralnetworktraining.Inthispaper,th

38、eresilientback-propagation(RPROP)algorithmisusedasthelearningalgorithmoftheMLPneuralnetwork[26].RPROPconsidersthesignofderivativesastheindicationforthedirectionoftheweightupdate.Indoingso,thesizeofthepar-tialderivativedo

39、esnotinfluencetheweightstep.Thefollowingequationshowstheadaptationoftheupdatevaluesof?ij (weightchanges)fortheRPROPalgorithm.Forinitialization,all?ij aresettosmallpositivevalues:?ij(t) =? ? ? ?? ? ??+ ??ij(t?1);if ?E?wij

40、 (t?1) ?E?wij (t)?0?? ??ij(t?1);if ?E?wij (t?1) ?E?wij (t)?0?0 ??ij(t?1);otherwise(17)where?0 =1,0??? ?1??+,??,0,+ areknownastheupdatefac-tors,wij representstheweightvaluefromneuronjtoneuroni,andErepresentstheerrorfuncti

41、on.Wheneverthederivativeofthecorrespondingweightchangesitssign,itimpliesthatthepreviousupdatevalueistoolargeandithasskippedaminimum.Therefore,theupdatevalueisthenreduced(??)asshownabove.However,ifthederivativeretainsitss

42、ign,theupdatevalueis(?+)increased.Thiswillhelptoaccelerateconvergenceinshallowareas.Toavoidover-acceleration,intheepochfollowingtheapplicationof?+,thenewupdatevalueisneitherincreasednordecreased(?0)fromthepreviousone.Not