Spectrogram-Based Spectrum Prediction for AI-managed Cognitive-Radio Edge Devices

This paper presents a Radio-Frequency (RF) spectrum prediction system based on a Convolutional Neural Network (CNN) architecture for image-based forecasting intended for Cognitive Radio (CR) terminals. Compared to previous approaches based on the use of time series, we propose a more efficient way for occupancy spectrum prediction based on channel availability tables (derived from spectrograms) to train a deep learning model. As a result, the neural engine is able to predict the best band available for Secondary User (SU) transmission in edge devices. As a proof of concept, a simple CR demonstrator combining a computational model for the neural engine – previously modeled and trained in Python using the Keras API – in MATLAB/SIMULINK with two Software-Defined Radio (SDR) boards has been developed. Three different table sizes were used as input for the predictor and a comparison is presented. The system performance is assessed using real over-the-air signals captured in the 2.412 GHz central frequency with 60 MHz bandwidth to validate the presented approach<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>.