Spectrum availability prediction based on RCS-GRU model

Abstract

Traditional spectrum sensing requires complex prior information from primary user (PU), and consumes a lot of energy and time in the sensing process. In order to improve the efficiency of spectrum sensing in cognitive radio networks, we propose a novel type of deep neural network that combines Residual network, Channel and Spatial attention modules (RCS), and Gated Recurrent Unit network (RCS-GRU). This network analyzes the occupancy correlation of the authorized spectrum on different channels by the PU, and predicts the spectrum occupancy of the next time slot, thereby reducing the collision probability between the secondary user and PU. In order to verify the effectiveness of the network in actual conditions, we adopt the proposed prediction model to experiment on the channel state simulated by M/M/N queuing theory and the real channel model constructed by GNU’s Not Unix (GNU) Radio. Experimental results show that the prediction accuracy of the RCS-GRU prediction model outperforms those of the GRU model and the Convolutional Long Short-Term Memory model. In addition, experiments with deep neural networks show that compared with the channel state date set simulated by queuing theory, the real date set constructed on GNU Radio has a faster convergence speed.