Time Shift Parameter Setting of Temporal Decorrelation Source Separation for Periodic Gaussian Signals

Takeshi AMISHIMA  Kazufumi HIRATA  

Publication
IEICE TRANSACTIONS on Communications   Vol.E96-B   No.12   pp.3190-3198
Publication Date: 2013/12/01
Online ISSN: 1745-1345
DOI: 10.1587/transcom.E96.B.3190
Print ISSN: 0916-8516
Type of Manuscript: PAPER
Category: Sensing
Keyword: 
independent component analysis,  TDSEP,  blind separation,  parameter setting,  

Full Text: PDF(899.6KB)>>
Buy this Article




Summary: 
Temporal Decorrelation source SEParation (TDSEP) is a blind separation scheme that utilizes the time structure of the source signals, typically, their periodicities. The advantage of TDSEP over non-Gaussianity based methods is that it can separate Gaussian signals as long as they are periodic. However, its shortcoming is that separation performance (SEP) heavily depends upon the values of the time shift parameters (TSPs). This paper proposes a method to automatically and blindly estimate a set of TSPs that achieves optimal SEP against periodic Gaussian signals. It is also shown that, selecting the same number of TSPs as that of the source signals, is sufficient to obtain optimal SEP, and adding more TSPs does not improve SEP, but only increases the computational complexity. The simulation example showed that the SEP is higher by approximately 20dB, compared with the ordinary method. It is also shown that the proposed method successfully selects just the same number of TSPs as that of incoming signals.