Convolutive independent component analysis by density estimation and leave-one-out approximation

• Main Authors: Wan-Jhen Jhou, 周琬真
• Other Authors: Jiann-Ming Wu
• Format: Others
• Language: en_US
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/60627297990860247707

碩士 === 國立東華大學 === 應用數學系 === 93 === This work explores blind separation of convolutive mixtures of independent sources. The convolutive structure consists of multiple, e.g. τ, mixing matrices, each corresponding to a different time delay, through which a segment of consecutive source signals are convoluted to form an observation. Based on the convolutive structure, the observations are temporally correlated among their different components. As τ = 1, the convolutive structure reduces to a linear transformation that produces temporally uncorrelated observations among different components, simply to a case typically attacked by independent component analysis (ICA). For arbitrary τ, estimating the convolutive structure as well as source signals subject to given observations decomposes to τ simultaneous sub-tasks following the leave-one approximation, each corresponding to optimizing a mixing matrix subject to a set of intermediate observations, each measuring the convolutive result of source signals through the other τ-1 mixing matrices. By the decomposition, each of τ simultaneous sub-tasks translates to a typical ICA task and can be directly resolved by the density estimation based ICA algorithm here. Numerical simulations show the novel approach is effective for blind source separation of fetal ECG and event-related potentials (ERP) signals.