Abstract: In this paper, we study the problem of modulation classification in the presence of phase error and improve modulation classification algorithm which uses constellation shape as classification signature. First, dynamic clustering algorithm is utilized to recover unknown constellation based on similarity between samples and kernels. Then recovery constellation is matched with constellations of different modulated signals and classified by a proposed max-likelihood rule, which is simple and equivalent to minimum distance classification rule. This rule avoids a training phase which is necessary for spatial statistics of recovery constellation vertices in previous algorithms. Considering the effect of additive noise on phase estimation, simulations show that the correct rate is above 90 percent when SNR is 10db and 15db respectively under the condition of known and unknown number of modulation states.

Key words: modulation classification, constellation, clustering