This paper considers the blind source separation in under-determined case,when there are more sources than sensors.So many algorithms based on sparse in some signal representation domain,mostly in Time-Frequency (T-F) domain,are proposed in recent years.However,constrained by window effects and TF resolution,these algorithms cannot have good performance in many cases.Considering most of signals in real world are band-limited signals,a new method based on sub-band division is proposed in this paper.Sensing signals are divided into different sub-bands by complementary filter firstly.Then,classical Independent Component Analysis (ICA) algorithms are applied in each sub-band.Next,based on each sub-band s estimation of mixing matrix,the mixing matrix is estimated with cluster analysis algorithms.Mter that,the sub-band signals are recovered using the estimation mixing matrix,and then,the resource signals are reconstructed by combining the related sub-band signals together.This method can recover the source signals if active sources at any subband do not exceed that of sensors.This is also a well mixing matrix estimating algorithm.Finally,computer simulation confirms the validity and good separation performanne of this method.