Spectrogram-based systems for spectrum sensing using deep learning

This document describes and compares four systems for spectrum sensing (SS) using spectrograms. The proposed approaches aim to detect and predict vacant frequency bands for secondary users in cognitive radio networks. First, a review of deep learning (DL)-based SS approaches is presented, showing the necessity of prototyping these systems over Internet of Things (IoT) devices. Prior approaches to this problem only simulate their proposals but do not implement them on real devices. The spectrogram representation was chosen for these SS systems because it is a useful image that leverages image-processing techniques and DL networks to exploit features used for wireless signal detection. A first SS system was proposed to improve signal detection in high-noise environments by denoising spectrograms. A synthetic data set was generated using 4G long-term evolution (LTE) and 5G new radio (NR) signals. The bilateral Gaussian filter (BGF) was proposed as an image-processing technique for denoising images, which were later used to train a ResNet-18 architecture for semantic segmentation. The second system relies on another spectrogram data set of Wi-Fi and Bluetooth signals, which was used to train a YOLOv8 object detector for radio frequency frame detection. This system was implemented in a Raspberry Pi with a software defined radio (SDR) to test its performance on real over-the-air data. The third system addresses spectrum prediction by adding a spectrum hole algorithm to the second system. This proposal was implemented in SIMULINK for prototyping a secondary user transmission using predicted spectrum vacancy data based on 2D Convolutional Neural Networks (CNN). Finally, the fourth system deals with prediction-based SS through an IoT-suitable implementation in a Raspberry Pi. This framework improves the second system, including an occupancy prediction model based on 1D-CNN, BGF and another SDR for spectrum sharing. These proposals show a constant progress toward developing IoT-suitable systems for SS in cognitive radio applications.