A Radio Frequency Region-of-Interest Convolutional Neural Network for Wideband Spectrum Sensing

Abstract

Wideband spectrum sensing plays a crucial role in various wireless communication applications. Traditional methods, such as energy detection with thresholding, have limitations like detecting signals with low signal-to-noise ratio (SNR). This article proposes a novel deep learning-based approach for RF signal detection in the wideband spectrum. The objective is to accurately estimate the noise distribution in a wideband radio spectrogram and improve the detection performance by substracting it. The proposed method utilizes convolutional neural networks to analyze radio spectrograms. Model evaluation demonstrates that the RFROI-CNN approach outperforms the traditional energy detection with thresholding method by achieving significantly better detection results, even up to 6 dB, and expanding the capabilities of wideband spectrum sensing systems. The proposed approach, with its precise estimation of noise distribution and consideration of neighboring signal power values, proves to be a promising solution for RF signal detection.

Clustering Method for Signals in the Wideband RF Spectrum Using Semi-Supervised Deep Contrastive Learning

“…In narrowband spectrum sensing solutions, various signal detection methods are typically employed, such as the energy detector, cyclostationary detector, matched filter, correlation detector, or wavelet detector. However, in the case of wideband analysis and unknown signal characteristics in the spectrum, the set of possible detectors narrows down to the energy detector (ED [12][13][14]) or its extensions (e.g., ED-ENP [14]), including convolutional neural network-based detectors such as RFROI-CNN [15]. These methods enable the identification of regions of interest in the wideband spectrum, even at low signal-to-noise ratio (SNR) values.…”

“…Evaluating the network based on wideband radio spectrogram requires generating subspectrograms containing detected radio signals, which will be processed by the neural network. Similar to the training process, the annotation files of the rfspec-db [22] database were used, enabling the extraction of signals from the wideband spectrum without the need for energy detectors such as ED [12][13][14] or RFROI-CNN [15]. The annotation files also provide information about the modulation used, which was used for validating the proposed simple hierarchical clustering method.…”