SDR Implementation of Spectrum Sensing Using Deep Learning

Abstract

Software Defined Radio (SDR) is a technology that offers a high level of reconfigurability to address the issue of spectrum sparsity in wireless communication systems. This technology is widely used in Cognitive radio (CR), and researchers aim to develop new spectrum sensing methods that ensure a high signal detection performance and a low signal-to-noise ratio (SNR). In this context, deep learning (DL) based models can be an appropriate solution for building spectrum detection methods. This paper proposes a spectrum sensing architecture combining a convolutional neural network and long short-term memory (CNN-LSTM). This architecture takes advantage of the spatial modelling of CNN and the temporal modelling of LSTM to produce more separable features for detection. The paper aims to propose an SDR implementation of the CNN-LSTM model for real-time detection by using the Universal Software Radio Peripheral (USRP) board and GNU radio platform. Results and Discussion: Numerical Simulation results reveal that the proposed CNN-LSTM outperforms the CNN, the LSTM, and the energy detector (ED) in terms of higher detection probability Pd and lower false alarm probability Pfa, even at low SNR. The SDR implementation results show the robustness of the CNN-LSTM method under several real-time detection scenarios: FM, GSM, and OFDM. The CNN-LSTM model used for spectrum sensing provides a high detection performance in a low SNR environment compared to LSTM, CNN, and the ED detector.