A new radar pulse signal clustering algorithm based on inner-pulse modulation characteristic is studied. The number of clusters in original data and the cluster that data vector belongs to is determined by clustering algorithm using the minimum description length (MDL)， thus obtaining the number of radar emitters in the electromagnetic environment. The radar pulse data collected by the interception system front-end is treated as vectors and is sorted after the necessary pre-processing. Pulse signals from different radars form different vector clusters after deinterleaving by using the algorithm based on MDL criterion， because of their different inner-pulse modulation characteristic. The largest advantage of the algorithm is that it is unnecessory for quantitative analysis, extraction and description of the inner-pulse modulation characteristic. Simulation proves its effective.