A Robust On-Line Blind Separation Algorithm with Dynamic Source Number Based on Neural Network

Most blind source separation (BSS) algorithms deal with cases in which the number of sources is assumed known. This paper proposes a robust on-line blind separation algorithm with unknown and dynamic number of sources. Based on existing over-determined (more sensors than sources) architecture, after estimating the source number using SVD, we add the momentum term to improve the classical Cichocki-Unbenauen algorithm, which can not only keep the independence of the outputs but also avoid separation results from dropping into the local minimization. Moreover, we set changeable learning step and moving step to further enhance the separation performance. Compared with the newly proposed ANA algorithm, computer simulations validate our algorithm's efficiency which has both a higher convergence speed and a lower steady-state error.