Performance Analysis and Optimization of Novel High-Order Statistic Features in Modulation Classification

This paper proposes novel higher-order statistic amplitude features in high-order cyclic cumulant domains. More classification information is preserved to adapt to more modulation types. Three representative feature sets extracted from time/frequency, high-order cumulant and high-order cyclic cumulant domains. The combined feature sets could make full use of the information and advantage of the three by the majority logic rule. In addition, linear smoothing is used to preprocess the signal. Based on the modified algorithm of neural network recognizer, simulation results verify the improvement of average probability of correct classification by 20-30% at low SNR due to novel features, accurate parameter estimate, as well as the preprocessing.