Abstract: A classification algorithm of MFSK signal is analyzed and improved in this paper. Considering the pulse shaping impact in practical system, the constant envelope pretreatment is used. The Haar wavelet which can restrain noise well to extract signal character is selected. Aiming at the disadvantage of FCM algorithm which is sensitive to initial values and easily falls into local optimum solution, this paper gives a novel cluster method mixed subtractive and c-means cluster which not only avoids falling into local optimum solution, but also speeds up the rate of convergences.