RadioDL: Deep Learning-Based Signal Intelligence of IQ Captures

Spectrum monitoring has become essential to ensure the reliability of wireless communications. Spectrograms provide a visual time-frequency representation of the radio signal, which analyzed with deep learning tools of the visual world, enables the identification of coexistence problems, interference and critical propagation conditions. In this paper, we propose a spectrum analysis system able to detect 5G control signals. Our methodology is based on the acquisition of 16 -bit IQ signals on extended time scales, transformed and analyzed as images. The paper discusses issues related to dataset creation for model training, as well as the possibility of using pre-trained models through transfer learning and few-shot learning to detect signals in real environments. To this end, a proof of concept of the proposed methodology was conducted, based on a dataset created from real-world acquisitions used to train a YOLOv8x model. The results show that the methodology is very promising, but highlight the need for significantly more examples to train the model.