Multichannel Blind Identification from Noisy Sensor Array Observations: A Stochastic Realization Approach

Abstract

Subspace methods for blind multichannel identification can not be extended to the case of a non white noise. For an unknown temporally white but spatially correlated perturbation, we pr+ pose a method based on a stochastic realization approach. It relies on the fact that the observed signal spectral density matrix is the s u m of a rational rank 1 spectral density and of a constant positive definite matrix (the noise Covariance matrix). The generic unicity of this decomposition is shown. An identification method based on the parametrization of the (external) stochastic realizations of the observed signal whose innovation sequence has a prescribed dimension is developped. It results in a state-space realization of the multichannel transfer function and in the noise covariance matrix.