A hybrid ICA-SVM approach to automatic modulation recognition

Automatic modulation recognition is a topic of interest in many fields including signal surveillance, multi-user detection and radio frequency spectrum monitoring. This thesis examines several issues related to automatic modulation recognition and presents an alternative low SNR recognition algorithm using elements of cyclostationary analysis, independent component analysis and support vector machines. Previous work by A.K. Nandi, E.E. Azzouz and A. Swami use instantaneous signal measurements or high order cummulants to identify modulations of interest. A major weakness of conventional modulation recognition algorithms is their sensitivity to SNR variations. In addition to this; the majority of research into this area does not consider continuous phase modulated signals or the effect of channel memory on algorithm performance. The hybrid algorithm developed in this thesis is applied to both digital and continuous phase modulated signals under a variety of channel impairments.;By deploying an independent component analysis algorithm on the cyclic feature sets, a pseudo-distance measure is shown to be increased. By increasing this pseudo-distance measure, the distance between modulation subspaces in the support vector machine hypothesis space is also increased. By increasing the distance between modulation subspaces the corresponding number of classification errors are decreased. Several empirical examples are provided to illustrate this connection. In addition to the modulation recognition problem the algorithm is applied to a different problem involving engine classification. Using a modified cepstral feature set; the hybrid algorithm is able to distinguish different engine types operating in different modes based upon vibrational and acoustic data.;The algorithm presented, based upon the novel combination of independent component analysis and support vector machines, is able to reliably classify a wide array of modulation types in much more unfavorable signal environments than previous approaches. Furthermore, the favorable results obtained by applying the algorithm to a separate recognition problem indicate a wider area of application than only modulation recognition.