Abstract: This paper solved a problem that any MFSK signal modulation index couldn’t be recognized if there wasn’t any prior information. It proposed a novel signal modulation recognition method which was based on the clustering characteristics of MFSK. First, extracted the short-term peak frequency of the signal by using the pole of AR model; and then, clusterde the peak sequence by using the improved clustering algorithm; finally, according to the FSK signals in the different characteristic parameters in different clustering centers, classify by supporting vector machine to complete the modulation recognition of 2、4、8、16FSK. This algorithm doesn’t need any priori knowledge and fits the FSK signals of different modulation indexs. The simulation experiment proves the algorithm effectively and correctly.