this paper employs autoregressive models to generate feature parameters and the knearest-neighbour method to classify modulation types.Simulation results indicate that this method can classify modulations in a very low SNR environment such as -1dB
and be robust to the frequency offset and the dynamic change of sample frequency.