Spectral-Temporal Model for Opportunistic Spectrum Access in Cognitive Radio Networks

Abstract

In the quest for overcoming spectrum scarcity, cognitive radios opt to dynamically exploit underutilized frequency bands in a primary wireless network to support communication among secondary users. However, the high fluctuation of channel usage requires continuous network monitoring through spectrum sensing, complicates medium access control, and forces secondary communications to be opportunistic with unbounded latency. To overcome these challenges, this paper advocates the use of machine learning (ML) to form a predictive model for when a channel becomes available. In particular, we propose a spectral-temporal spike neural network (ST-SNN). Unlike conventional ML-based approaches, our design tracks the variability of channel selection (spectral aspect) and infers access patterns of the channel over time (temporal aspect). The architecture fundamentally adopts both recurrent and feed-forward neural networks where channel usage tracking data is used for training and predicting the spectrum utilization for the next cycle; the error is fed back to improve the prediction accuracy. Our results using an Long Term Evolution (LTE) dataset show that the proposed predictive model achieves about 96.6% accuracy of channel availability assessment.