An improved blind source separation algorithm based on generalized variance

In this paper, an improved blind source separation algorithm based on generalized variance is proposed. The observation signal is partitioned to nonoverlapping blocks and the covariance matrix is estimated at some time delay. An algorithm that converts the nonpositive definite covariance matrix into positive definite matrix is presented; thus, without searching for the positive definite covariance matrix, the objective function is defined by using the Hadamard inequation conveniently. The improved blind source separation algorithm belongs to nonorthogonal joint diagonalization and the error due to the whitening processing and the orthogonal transformation is eliminated. Simulations results demonstrate that the algorithm has a substantial improvement in the separating performance.