This thesis addresses the blind deconvolution/identification problem in which the input signals to a multi-input multi-output system and the associated system transfer functions are to be identified from the output signals only, without any a priori knowledge of both. The necessary and sufficient conditions of exact deconvolution of the system using FIR filters are stated and proved. The required length of the deconvolution filters is also derived. We consider this problem for two different classes of input signals. First we investigate the problem where the input signals are non-Gaussian i.i.d. and mutually uncorrelated. A recently proposed algorithm based on optimization of a cumulant-based objective function is analyzed and modified. In the original algorithm, subsequently extracted...[ Read more ]

This thesis addresses the blind deconvolution/identification problem in which the input signals to a multi-input multi-output system and the associated system transfer functions are to be identified from the output signals only, without any a priori knowledge of both. The necessary and sufficient conditions of exact deconvolution of the system using FIR filters are stated and proved. The required length of the deconvolution filters is also derived. We consider this problem for two different classes of input signals. First we investigate the problem where the input signals are non-Gaussian i.i.d. and mutually uncorrelated. A recently proposed algorithm based on optimization of a cumulant-based objective function is analyzed and modified. In the original algorithm, subsequently extracted input signals suffer from degradation in quality due to accumulated errors in previously extracted signals. We propose a new reinitialization method which does not have such drawback. Next we investigate the problem where the input signals are mutually uncorrelated with different power spectral densities. A two-stage algorithm is proposed for blind identification of MIMO system. The first stage of signal separation is derived from a matrix pencil formed between output auto-correlation matrices at different delays. The second stage is based on a subspace method to identify and deconvolve the MIMO system. This algorithm employs only second-order statistics and does not involve any gradient search iteration. Computer simulation results are presented to demonstrate the validity of the proposed algorithms.