user_alg3.m

Matlab盲信号分离很有用的工具

%function [W,He,Y] = evd( X, n, p, alpha );
function [W] = evd( X, n, p, alpha );
% EVD method for source separation
% written by Andrzej Cichocki and Pando Georgiev
% m - number of sensors
% n - number of sources 
% p - number of time-delated covariance matrices

[m,N]=size(X);

if nargin == 1,
	n = m;  			% 
 	p = 200;  			% the number of time delays can be defined by user
 	alpha = ones(1, p);
 
elseif nargin == 2,
 	p = 5; 			% number of the covriance matrices
	alpha = ones(1, p);
 
elseif nargin == 3,
 	alpha = ones(1, p);
    
end; 

%X	= X - mean(X')' * ones(1,N);

%%% Standard whitening
Rx0 = (X*X')/N;

% Rx0=(X(:,1:N-1)*X(:,2:N)')/(N-1); %Estimation of sample covariance matrix 
%for the time delay p=1 in order to reduce influence of a white noise.
%
% Q=inv(sqrtm(Rx0));

[Ux,Dx,Vx]=svd(Rx0);
 Dx=diag(Dx);
% n=11;
 if n<m, % under assumption of additive white noise and
        %when the number of sources are known or can a priori estimated
  Dx=Dx-real((mean(Dx(n+1:m))));
  Q= diag(real(sqrt(1./Dx(1:n))))*Ux(:,1:n)';
else    % under assumption of no additive noise and when the 
        % number of sources is unknown
   n=max(find(Dx>1e-14)); %Detection the number of sources
   Q= diag(real(sqrt(1./Dx(1:n))))*Ux(:,1:n)';
end;
Xb=Q*X; % prewhitened data
%
%%%correlation matrices estimation
%alpha=rand(1,p)
%
Rxb=zeros(n);
% Set of paramters alpha is now random
% The user can define them, optional all should one
k=1;
for u=1:p, 
	k=k+1; 
   Rxb=Rxb+alpha(u)*Xb(:,k:N)*Xb(:,1:N-k+1)'/(N-k+1); 
end;

[V,D] = eig(Rxb+Rxb');
[D1 perm]=sort(diag(D));
D1=flipud(D1);
V=V(:,flipud(perm));

%Y=V'*Xb;
%He=inv(Q)*V;
 W=V'*Q;