Blind source separation (BSS) attempts to recover unknown independent sources from a given set of observed mixtures. Algorithm based on natural gradient is one of the main methods in BSS. An analysis has been done on the problem that the old algorithm goes to diverging under over-determined mode. A new improved algorithm based on orthogonal projected natural gradient is studied in the paper. The simulated result using crosstalk error proves the capability to perform the BSS under over-determined mode and the better convergence stability of the new algorithm. It is also effective with a dynamically changing source number.