APPARATUS AND METHOD FOR SIGNAL EXTRACTION AND SEPARATION

• US 20120259561A1
• Filed: 04/05/2012
• Published: 10/11/2012
• Est. Priority Date: 04/05/2011
• Status: Active Grant

• Current Assignee: Regents of the University of California (University of California)
• Original Assignee: Regents of the University of California (University of California)
• Inventors: Hoh, Carl Kalon
• Granted Patent: US 9,940,546 B2
• Time in Patent Office: Days
• US Class Current: 702/57
• CPC Class Codes: G06F 18/213 Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
• CPC Class Codes: G06F 18/2134 based on separation criteria, e.g. independent component analysis
• CPC Class Codes: G10L 21/0272 Voice signal separating

Abstract

Method and/or apparatus for independent component analysis of a mixed electronic signal. Candidate signal components are determined, and mutual information (MI) criteria between each candidate signal component is minimized. The candidate signal components are optimized using polar coordinates.

Claims

1. A method for independent component analysis of a mixed electronic signal, the method comprising:

determining candidate signal components;

minimizing mutual information (MI) criteria between each candidate signal component; and

optimizing candidate signal components using polar coordinates.

2. The method of claim 1, wherein said determining, minimizing, and optimizing provides an unmixing matrix;

wherein the method further comprises using the unmixing matrix to transform a signal mixture matrix based on the mixed electronic signal;

wherein said transforming provides signals comprising independent signal vectors.

3. The method of claim 1, wherein said minimizing mutual information criteria comprises:

providing an unmixing matrix;

transforming the signal mixture matrix using the unmixing matrix to provide a source matrix; and

determining total mutual information between rows of the source matrix.

4. The method of claim 3, wherein said optimizing comprises:

orthogonally rotating each row of the unmixing matrix by a plurality of angles to update the unmixing matrix.

5. The method of claim 4, wherein said optimizing further comprises:

transforming the signal mixture matrix using the updated unmixing matrix to provide an updated source matrix;

determining total mutual information between rows of the updated source matrix.

6. The method of claim 5, further comprising:

repeating said minimizing and optimizing until converge criteria is met.

7. The method of claim 3, wherein said providing an unmixing matrix comprises:

determining a number of column vectors in the signal mixture matrix that approximate apices in factor analysis, wherein the number of column vectors is equal to the number of determined candidate signal components;

performing a matrix inversion on the signal mixture matrix to provide an initial unmixing matrix;

orthogonally rotating each row of the initial unmixing matrix by a plurality of angles to provide the unmixing matrix.

8. The method of claim 7, further comprising:

dimensionally reducing mixed signal data to have a number of rows equal to the determined number of signal components; and

whitening the dimensionally reduced signal data to provide the signal mixture matrix.

9. The method of claim 1, further comprising:

testing behavior of total mutual information for the candidate signal components as affected by a factor taken from the group consisting of histogram bin size, potential effect of data interpolation, and analysis of numbers of data vectors.

10. The method of claim 1, wherein the mixed electronic signal comprises mixed temporal signals.

11. The method of claim 1, wherein the mixed electronic signal comprises mixed spatial signals.

12. A method for independent component analysis of a mixed electronic signal, the method comprising:

a) determining a number k of candidate signal components in the mixed electronic signal;

b) determining k column vectors in a signal mixture matrix X_w based on mixed electronic signal data, which column vectors approximate the apices in factor analysis, to provide a signal mixture matrix X_f;

c) performing matrix inversion on the signal mixture matrix X_f to obtain initial weight vectors w_o of an initial unmixing matrix W_o);

d) using Givens rotations based on angle θ

, orthogonally transforming each w_o separately to create an unmixing matrix W;

e) transforming the signal mixture matrix X_w with the unmixing matrix W to provide a source matrix;

f) calculating total mutual information MI_t between each row in S;

g) obtaining a new angle θ

;

h) generating Givens rotations R to reapply to each w_o to provide a new unmixing matrix W; and

repeating steps e)-h) until convergence criteria are met.

13. The method of claim 12, wherein the convergence criteria comprises the total mutual information MI_t and each parameter value in θ

changing by less than a predetermined amount.

14. The method of claim 12, further comprising:

dimensionally reducing mixed electronic signal data to have a number of rows equal to the determined number of signal components; and

whitening the dimensionally reduced signal data to provide the signal mixture matrix.

15. The method of claim 14, wherein said repeating steps e)-h) provides an optimized unmixing matrix;

further comprising;

transform the signal mixture matrix using the mixed electronic signal;

wherein said transforming provides signals comprising independent signal vectors.

16. The method of claim 12, further comprising:

testing behavior of said calculated total mutual information for the candidate signal components as affected by a factor taken from the group consisting of histogram bin size, potential effect of data interpolation, and analysis of numbers of data vectors.

17. The method of claim 12, wherein the mixed electronic signal comprises mixed temporal signals.

18. The method of claim 12, wherein the mixed electronic signal comprises mixed spatial signals.

19. An apparatus comprising computer readable instructions stored on a non-transitory medium, which when executed cause a computer to:

a) determine a number k of candidate signal components in a mixed electronic signal;

b) determine k column vectors in a signal mixture matrix X, based on mixed electronic signal data, which column vectors approximate the apices in factor analysis, to provide a signal mixture matrix X_f;

c) perform matrix inversion on the signal mixture matrix X_f to obtain initial weight vectors w_o of an initial unmixing matrix W_o);

d) using Givens rotations based on angle θ

, orthogonally transform each w_o separately to create an unmixing matrix W;

e) transform the signal mixture matrix X_w with the unmixing matrix W to provide a source matrix;

f) calculate total mutual information MI_t between each row in S;

g) obtain a new angle θ

;

h) generate Givens rotations R to reapply to each w_o to provide a new unmixing matrix W;

repeat steps e)-h) until convergence criteria are met to provide an optimized unmixing matrix; and

transform the signal mixture matrix with the optimizing unmixing matrix to provide signals comprising independent signal vectors.