Kernel-Based Generalized Discriminant Analysis for Signal Classification in Cognitive Radio

The reaction of detection and classification of signals in low Signal-to-Noise Ratio (SNR) regimes poses significant challenges in the physical layer design of cognitive radio networks. This paper considers a cognitive radio consisting of one Primary User (PU) and K Secondary Users (SUs), where the main objective for an arbitrary SU is to recognize the PU's signal from other secondary users' signals, in order to occupy the spectrum hole. Toward this goal, we present a Kernel-based Generalized Discriminant Analysis (KGDA) for the modulated signal classification where the scheme displays a simple model for each class of modulated signals in a feature space. We use both statistical and spectral features in the proposed scheme for some popular digital modulations. One advantage of the proposed scheme is the robustness of the approach against SNR variations. Simulation results show that our approach improves significantly the classification performance in the low SNR scenarios when compared to some traditional classification algorithms such as the Support Vector Machine (SVM) algorithm. The applied KGDA has the advantage of a very low computational complexity for both training and test phases which makes the proposed scheme can be deployed for real-time cognitive radio applications.