Abstract. Regarding blind signal separation based on eigenvalue decomposition, when the spatial dimension of signal is wrong in case of unknown source quantity, it will result in significant separation errors. By making use of kurtosis as the cost function, this paper attempts to construct a blind signal separation algorithm taking "space-kurtosis" spectrum as the basis, thus avoiding eigenvalue decomposition. Meanwhile, in the operation process, the source signal statistic independence and spatial distribution independence are fully utilized. It is proved by simulation experiments that the algorithm is with characteristics of high accuracy, fast operation and strong robustness.