tdsep2.m

SOBI算法

function [C,D]=tdsep2(x,sel);
%blind separation using approximate joint diagonalization of
%time delayed correlation-matrices
%
%version 2.01, 2/14/99 by AZ
%usage: [C,D]=tdsep2(x,sel);
% input   x     data matrix 
%         sel   array of integer time lag values, default sel=[0:1]
% output  C     estimated mixing matrix
%         D     set of aproximate diagonal matrices
%

%THIS IS UNPUBLISHED PROPRIETARY SOURCE CODE of GMD FIRST Berlin.
%
%  The purpose of this software is the dissemination of
%  scientific work for scientific use. The commercial
%  distribution or use of this source code is prohibited. 
%  (c) 1996-1999 GMD FIRST Berlin, Andreas Ziehe 
%              - All rights reserved -


if nargin<2,
  sel=[0 1];
end  

% whitening or sphering
M0=cor2(x',sel(1));

SPH=inv(sqrtm(M0));
spx=SPH*x;
[p,q]=size(M0);


N=length(sel);

% for two matrices, solve directly as general eigenvalue problem
if N==2,
     M1=cor2(spx',sel(2));
  [Q,D]=eig(M1);
      C=inv(SPH)*Q;
 else

t=1;          % compute correlation matrices
for tau=1:N,
  M(:,t*p+1:((t+1)*p))=cor2(spx',sel(tau));
  t=t+1;
end
% joint diagonalization
[Q,D]=jdiag(M,0.00000001);
% compute mixing matrix
C=inv(SPH)*Q;
end