Blind source separation based on cyclic spectra: Application to biomechanical signals

This paper introduces new frequency domain approaches for either blind source separation or MIMO system identification excited by cyclostationary inputs. The eigenvalue decomposition, the singular value decomposition, the diagonalization of a positive definite linear combination or the joint diagonalization, of the spectral correlation density matrices of the whitened measurements allows the identification of the mixing system at each frequency up to constant diagonal and frequency dependent permutation and phase ambiguity matrices. Two efficient algorithms to fix the permutation problem and to remove the phase ambiguity based on cyclostationarity are also presented. The new approaches exploit the fact that the inputs are cyclostationary with the same cyclic frequency. Simulation examples are presented to illustrate the effectiveness of this approaches. Furthermore, the AJD approach is applied to biomechanical signals for separation ends.