Statistical Feature-Based SVM Wideband Sensing

Poor accuracy at low SNR makes fast autonomous wideband sensing in real-time difficult for cognitive radios. Support Vector Machines (SVM) can be an alternative in learning this sensing environment. But online training with conventional Eigen features has high complexity. This letter develops and validates on real-world data an efficient quick learning SVM-based blind sensing model using two new simple statistical features that can accurately detect signals at low SNR. Named as Smoothed Correlation of Reversed Spectrum Segments (SCRSS) and Variance of Multi-Scale Moving Averages (VMMA), they can speed up sensing to almost five times that of Eigen learning. With reduced computational cost, they appear to be promising in next-generation intelligent cognitive radio networks.