Blind Source Separation Using Two-Dimensional Nonnegative Matrix Factorization In Biomedical Field

• Main Author: Toh, Cheng Chuan
• Format: Thesis
• Language: English; English
• Published: 2018
• Subjects: T Technology (General); TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://eprints.utem.edu.my/id/eprint/23290/1/Blind%20Source%20Separation%20Using%20Two-Dimensional%20Nonnegative%20Matrix%20Factorization%20In%20Biomedical%20Field.pdf; http://eprints.utem.edu.my/id/eprint/23290/2/Blind%20Source%20Separation%20Using%20Two-Dimensional%20Nonnegative%20Matrix%20Factorization%20In%20Biomedical%20Field.pdf; http://eprints.utem.edu.my/id/eprint/23290/; https://plh.utem.edu.my/cgi-bin/koha/opac-detail.pl?biblionumber=112733

Blind Source Separation (BSS) refers to the statistical technique of separating a mixture of underlying source signals.BSS denotes as a phenomena and separation on mixed heart-lung sound is one of its example.The challenge of this research is to separate the separate lung sound and heart sound from mixed heart-lung sound.A clear lung sound for diagnosis purpose able to be obtained after separating the mixed heart-lung sound.In biomedical field,lung information is precious due to it has been provided for respiratory diagnosis.However,the interference of heart sound towards lung sound will generate ambiguity and it will lead to drop down the accuracy of diagnosis.Thus,a clean lung sound is needed to increases the accuracy of diagnosis.One of the ways for non-invasive respiratory diagnosis for obtaining lung information is through extracting lung sound from mixed heart-lung sound by using Two-Dimensional Nonnegative Matrix Factorization (NMF2D) algorithm.This method is based on cocktail party effect in which it refers to human brain able to selectively listen to target among a cacophony of conversations and background noise and this considered as a difficult task to machine.Therefore, duplication on cocktail party effect into machine is used to separate the mixed heart-lung sound.This research presents a novel approach NMF2D algorithm in which a suitable model for signal mixture that accommodated the reverberations and nonlinearity of the signals.The objectives of this research are focusing on investigating the useful signal analysis algorithms,defining a new technique of signal separability,designing and developing novel methods for BSS. In order to process estimation results,cost function such as β-divergence and α-divergence is integrated with NMF2D.Provisions of experiment are convolutive mixed signal is sampled and real recording using under single channel,Time-Frequency (TF) domain is computed by using Short Time Fourier Transform (STFT) respectively.Performance evaluation is done in term of Signal-to-Distortion Ratio (SDR). Theoretically,β and α is parameters that used to vary the NMF2D algorithm in order to yield high SDR value. Experimentally,for the simulation results,the highest SDR value for β-divergence NMF2D is SDR = 16.69dB at β = 0.8 and n = 100.For α-divergence NMF2D,the highest SDR value is SDR = 17.85dB at α = 1.5 and n = 100.Additional of sparseness constraints toward β-divergence NMF2D and α-divergence NMF2D lead to even higher SDR value.There are SDR = 17.06dB for sparse β-divergence NMF2D at λ = 2.5 and SDR = 17.99dB for sparse α-divergence NMF2D at λ = 5. This represents sparseness constraints yield to decrease ambiguity and provide uniqueness to the model.In comparison in between β-divergence,α-divergence,sparse β-divergence and sparse α-divergence NMF2D,it found that SDR value of sparse α-divergence NMF2D is the best decomposition method among all divergences.This can be concluded that sparse α-divergence NMF2D is more applicable in separating real data recording.