Primary User Activity Prediction in DSA Networks using Recurrent Structures

Abstract

For unlicensed (secondary) users to opportunistically access the shared radio spectrum on a non-interfering basis, it is important that they are able to sense the transmission activities of the licensed (primary) users. However, spectrum sensing expend a considerable amount of energy and time, which can be reduced by reliably predicting the primary user activities. In this paper, we present recurrent neural network models which are able to accurately predict the primary users’ activity in dynamic spectrum access (DSA) networks so that the secondary users can opportunistically access the unused spectrum. Using Universal Software Radio Peripheral (USRP) Software Defined Radios (SDRs), we collect over-the-air data from 8 primary users and train the learning models that we use in conjunction with a central spectrum sensor. We start by implementing two machine learning models: (i) traditional linear regression and (ii) neural network model using Long Short Term Memory (LSTM). These models are able to predict the primary users’ activity with 75% and 97% accuracy respectively. To further improve the prediction accuracy, we exploit the spatio-temporal correlation in the collected data by implementing a Convolutional LSTM model-which achieves 99% accuracy for predicting the long-term activity of primary users. The experimental results demonstrate that the proposed models are able to successfully predict the primary users’ activities, thereby reducing both the under-utilizations and interference violations in DSA networks.