A hybrid deep learning based approach for spectrum sensing in cognitive radio

The primary user (PU) transmission is sporadic in nature, which explains why the PU is inactive during some time slots, geographic directions or frequency bands. The frequency bands where the PU is not active are called "spectrum holes". Secondary users (SUs) periodically perform sensing to detect the spectrum holes and monitor primary spectrum. For the best possible spectrum utilization, PU signal detection is very crucial. For measuring the spectrum sensing performance, two main metrics are applied, like, probability of false alarm (PFA) and probability of detection (PD). Due to PFA and PD, the conventional sensing techniques have to face issues. These two constraints used to hinder spectrum utilization. Traditional sensing strategies are mostly based on feature extraction of received signal. Advancement of artificial intelligence (AI) has reduced the inaccuracy in detection of spectrum hole. Deep learning (DL) based approaches have shown a remarkable improvement in this aspect. Hence, the present research work was undertaken to address the problem of spectrum sensing in low SNR and improves accuracy. This research penetrates into the use of deep neural network (DNN) for sensing the vacant spectrum accurately. In this article, RadioML2016.10b dataset was used for the experiments. The results are also studied. The proposed approach shows betterment in sensing than other existing spectrum detection models. DeepSenseNet model was validated through simulation results and showed that it has achieved 98.84% prediction accuracy ($P_a$) with 97.53% precision and 97.62% recall.