SDR-Based Deep Learning Spectrum Sensing for TV White Space Applications

This paper presents a real-time spectrum sensing system for TV white spaces (TVWS) applications based on offlinetrained deep learning (DL) models for online inference. The system employs an Ettus USRP B210 software-defined radio (SDR) and MATLAB to evaluate, over ultra high frequency (UHF) ISDB-TB signals, three DL-based sensing approaches: a dense fully connected network (DenseNet), a lightweight fully connected network (SimpleNet), and a convolutional neural network (CNN), with traditional energy detection (ED) as a baseline. Experimental evaluations were conducted under line-of-sight (LoS), non-line-of-sight (nLoS), and mobile propagation scenarios. The results show that DL-based sensing improves detection performance over ED in all evaluated cases, although the gain is significantly reduced under mobility. In addition, the execution-time analysis indicates that the trained models are suitable for real-time operation. These results demonstrate the practical feasibility of intelligent SDR-based spectrum sensing for future cognitive radio and 5G/6G networks.