Prediction-Based Spectrum Sensing Framework for Cognitive Radio

This paper presents a hardware-software deep learning architecture for prediction-based spectrum sensing in Cognitive Radio (CR) applications. A convolutional neural network-based predictor for spectrum occupancy was trained using the band power from I/Q samples acquired by a softwaredefined radio (SDR). Additionally, a second neural engine was trained for radio frequency (RF) frame detection based on spectrograms. We implemented a transfer-learning solution using a You-Only-LookOnce version 8 nano model with a synthetic dataset comprising thousands of wireless signals, including Wi-Fi, Bluetooth, and collision frames. Once trained, the two neural networks were transferred to a Raspberry Pi 5 – an affordable single-board computer – connected to two (one for Rx, one for Tx) ADALM-PLUTO SDR systems for benchmarking using over-the-air signals in the 2.4 GHz band. Together with our methodology and experimental results, the paper also presents an overview of current spectrum prediction proposals and RF frame detectors. Remarkably, to the best of our knowledge, this proposed framework is the first approach towards an Internet of Things-suitable implementation of prediction-based spectrum sensing for CR applications.