A Blind Identi cation and Separation Technique via Multi-layer Neural Networks

| This paper deals with the problem of blind identication and source separation which consists of estimation of the mixing matrix and/or the separation of a mixture of stochastically independent sources without a priori knowledge on the mixing matrix. The method we propose here estimates the mixture matrix by a recurrent Input-Output (IO) Identiication using as inputs a nonlinear transformation of the estimated sources. Herein, the nonlinear transformation (distortion) consists in constraining the modulus of the inputs of the IO-Identiication device to be a constant. In contrast to other existing approaches, the covariance of the additive noise do not need to be modeled and can be estimated as a regular parameter if needed. The proposed approach is implemented using multi-layer neural networks in order to improve performance of separation. New associated local on-line un-supervised learning rules are also developed. The eeectiveness of the proposed method is illustrated by some computer simulations.