Feature extraction was studied. First, a new feature, compactness of power spectral, was brought forward. This feature extracted from frequency field was robust and not sensitive to noise. It needed no prior information. Then some difference processes of signal phase character were done and an improved parameter unwrapped difference phase normalized variance was given. This parameter did not rely on frequency and was more practical. The detection method of single frequency component was also re-defined, and the new extraction method after this re-definition was more robust and discriminable. At last signal classification system was designed, and the classification of AM, SSB, 2FSK, 4FSK, 8FSK, BPSK, QPSK and 8PSK was completed. The experiment simulation result showed that whenSNR of signal was more than 10 dB, the recognition accuracy is more than 92%.
