A Transformer Network for End-to-End Detection and Classification of Wideband RF Signals

Spectrum occupancy analysis is essential for enabling dynamic sharing of the scarce RF spectrum among coexisting wireless technologies. Recent works typically address wideband spectrum analysis by applying deep learning (DL) based object detection networks on spectrogram images of the monitored frequency bands to perform signal detection, localization, and classification. While effective, such approaches have certain limitations, including additional latency, information loss due to the transformation into image format, and reliance on pre- and post-processing steps. In this work, we propose an alternative DL framework that is able to detect, localize, and classify signals directly from frequency domain I/Q samples of wideband recordings. Following a direct set prediction problem formulation, our approach employs a transformer based network to generate a set of signal predictions in an end-to-end manner, eliminating the need for image generation and post-processing steps. We evaluate the proposed framework on a public dataset containing real-world wideband measurements covering five wireless technologies. Experimental results demonstrate that our approach achieves high signal detection performance, competitive spectrum segmentation results, and improved robustness to noise.