Automatic modulation classification, feature extraction, digital modulation, higher order statistics, support vector machines

Automatic modulation classification (AMC) is a frequently required framework to determine the modulation type of an incoming modulated signal with an unknown modulation type. AMC applications are divided under two main titles in the literature as likelihood-based (LB) and feature-based (FB) methods. In this thesis, an AMC algorithm is developed with a FB approach. As classifier, Support Vector Machine (SVM) using linear, quadratic and cubic kernel is chosen and their performances are compared. Over-the-air collected modulated signals with the SNR values between 0 and 30 dB are used. Signals are modulated with 12 different digital modulation types containing M-ASK, M-PSK, M-APSK up to higher orders. Statistical features i.e. mean, variance, skewness and kurtosis of the instantaneous amplitude, phase and frequency of the signal are used in addition to higher-order moments and cumulants up to 8th order. SVM using quadratic kernel showed slightly higher performance. In addition, a hierarchical classification structure with less complexity compared to the literature has been proposed in order to improve performance especially in high order modulation types which show very poor performance when classified with using a single classifier. A significant improvement is observed in the accuracies of these modulations comparing with the traditional method. The overall performance is increased from 80% to 90%.