Abstract: Aiming at the problem of low separation accuracy at the initial stages after an abrupt change of the time-varying mixing matrix, this paper proposed an improved on-line blind separation algorithm. The mutation point of the time-varying mixing matrix was determined by detecting the changes of the error function, and then used better converged separation matrix to separate the observed data retrospectively. Simulation results show that with the mixing scene of intermittent mutation, the proposed retrospective on-line EASI blind source separation algorithm can improve the separation accuracy at the initial stages and effectively track the mutation of the mixing matrix.