Learning Based Spectrum Hole Detection for Cognitive Radio Communication

This paper proposes a novel learning based (LB) solution for detection and quantification of spectrum holes in periodic communications of unmanned aerial vehicles (UAVs), Instead of hypothesis testing after implementation of spectrum sensing methods, the implemented LB solution based on spectral correlation function (SCF) uses region convolutional neural network (R-CNN) for extracting quantitative parameters of the spectrum holes. The proposed LB approach is implemented using GoogLeNet architecture for the wide band detection in the scenario of orthogonal frequency division multiplexing (OFDM) communication system with the additive white Gaussian noise (AWGN) channel model. The simulation of single input single output (SISO) communication system with spectrum holes is presented. Examples of wide band detection results for both SISO and multiple input multiple output (MIMO) systems are shown and the proposed LB detector is found to be fairly accurate in identification of spectrum holes. By analyzing the training performance, the GoogLeNet architecture, along with its hyperparameter configurations and training dataset is validated. We also demonstrated that our LB detector is resilient to the AWGN environment by analyzing the precision and recall curves, average precision and mean relative error (MRE) versus signal noise ratio (SNR).