Feature Detection BasedonFilter Banksand Higher OrderCumulants
Qingfeng Meng,Youyun Zhang,Wuwei Feng
2006-01-01
Abstract:Basedonthat higher-order cumulants isinsensitive harmonic retrieval, time-delay estimation, adaptive filtering, toadd-Gaussian noise andsymmetric noise, a methodfor array processing, andblindequalization. Thesestatistics, detecting feature ofsignal embeddedinadditive noiseis knownascumulants, andtheir associated Fourier transforms, presented, whichcombines a filter banksandhigher-order knownaspolyspectra, notonlyreveal amplitude information cumulants, neither thesupposition oftheobserved signal is about a process, butalsoreveal phase information. Thisis Gaussian orstationary, nora priorinformation aboutthe waveform andarrive timeoftheobserved signal isnecessary.important, because, asiSwellknown, second-order statistics Theprimary principle oftheproposed method istoseparate the (i.e., correlation)are phaseblind. Cumulants, ontheother spectrum oftheobserved signal into narrow frequency bandby hand, areblind toanykindofa Gaussian process, whereas using abandpass filter banks, those subfilters whichhavelarge correlation isnot.Consequently, cumulant-based signal output energies arechosen togiveatime-frequency description processing methodsalwaysusedtodealwithseveral main oftheobserved signal approximately, eachchosen subfilter is problems, include tosuppress additive colored noisewith thenfollowed byashort-time estimation ofthird-order cumulantunknownpowerspectrum, toidentify nonminimumphase sothattosuppress colored noise anddetect thesignal feature.system, to extract information dueto deviations from Simulated andexperimental results showthatthemethodis Gaussianty, todetect andcharacterize nonlinear properties in effective andpractical, evenifthesignal feature issubmerged signals or systemsandtoanalyse a cyclostationary time intonoise signal completely. series, etc. Inthispaper, basedonthathigher-order cumulants is I.INTRODUCTION insensitive toadd-Gaussian noise andsymmetric noise, we Thefeature detection oftransient signal isaveryimportantproposed anewmethodology whichintegrate themeritof issue forcondition monitoring andfault diagnosis. Itiswell Higher-order cumulants andthefilter banktodetect the knownthat theoptimal detector todetect atransient signalobserved signal withunknownfeature andembedded in whosecharacteristic isknownisamatchfilter andsome additive noise, neither thesupposition oftheobserved signal is suboptimal detection methods, including theshort-time Gaussian orstationary, nora prior information aboutthe Fourier transform, theGabortransform, thewaveletwaveform andarrive timeoftheobserved signal isnecessary. transforms andnonlinear methods(such astheWigner-Ville Theprimary principle oftheproposed methodistoseparate distribution)are already available. Butthese approach areall thespectrum oftheobserved signal intonarrowfrequency needtosuppose thatthedetection signal isGaussion. In bandbyusing abandpass filter banks. Thesubband signals are practice, themeasured signals always arenon-stationary and treated independently of one another, withsubbands non-Gaussion andneither waveshape norarrival timeare containing lowsignal energies notbeing processed atall, but known, especially whentheequipment isinfault.(1), insuch those subfilters whichhavelarge output energies arechosen to case, abovethese methods cannotperform verywell. In givea time-frequency description oftheobserved signal general, theenergy detection methodusedtodetect those approximately, eachchosen subfilter isthenfollowed bya signal whosecharacteristic areunknown, but,thisdetectorshort-time estimation ofthird-order cumulant sothatto perform wellonlyfordetecting White-Gaussion signal from suppress colored noise anddetect thesignal feature. Simulated Gaussion background. So,whenthenoise background isnot andexperimental results showthat themethod iseffective and idealism white noise, thedetection performance ofaboveall practical, evenifthesignal feature issubmerged into noise methods will beaffected. Mostofdetection methods will be signal completely. notworkverywell whenSNRissmaller. Thispaperisorganized asfollows. Insection IIwe