Blind Source Separation Based On Harmonic Extraction And Phase Coherence Criterion

Blind source separation(BSS)is to estimate the parameters of mixing system and recover the sources from the mixtures without the knowledge of mixing system.Two major categories of BSS are statistics based algorithm and sparse representation based algorithm,both of which have improvements in concurrently possessing high accuracy,high efficiency and the robust to small sample number.To solve the problem above,this paper proposes the phase coherence criterion based on harmonic extraction,and treats the criterion as the single-source component recognition criterion.Then,a novel BSS scheme can be formed through combining the criterion with spectrum correction,pattern screening,improved K-means clustering,data density based clustering,and subspace projection.The proposed scheme can improve the estimation accuracy of both the mixing system parameters and the source recoveries in spite of the overlapping frequencies.To be specific,the contribution of this work can be divided into 3 aspects:Firstly,a universal harmonic BSS model is proposed based on harmonic extraction.The proposed model integrates single-source pattern screening,improved K-means clustering and subspace projection technique together,which can not only deal with the basic harmonic BSS problem,but also solve the BSS problem of approximate periodic signals or harmonic-involved signals.Secondly,the proposed harmonic BSS model is applied into the BSS problem of rotating machinery signals with insufficient samples.In this case,the spectrum correction is introduced into parameter estimation to solve the spectrum leakage problem incurred by insufficient samples.Thirdly,the harmonic BSS model is applied into underdetermined non-stationary speech signals.Meanwhile,time-frequency analysis and spectrum correction are introduced in the scheme to accurately estimate the relevant harmonic parameters,and the improved K-means clustering is replaced by data density peak based clustering.In the scheme,the phase coherence criterion plays a key role in suppressing the serious frequency interference.In summary,the major innovations of this paper lie in: 1)proposing the idea of estimating the mixing matrix with harmonic parameters and introducing the spectrum correction technique into BSS problem to extract the harmonic parameters;2)proposing the phase coherence criterion to screen the single source components.Meanwhile,the proposed BSS method has application potential in other harmonic related BSS fields.