A Spectrum Sensing Method Based on CNN-LSTM Deep Neural Network

Abstract

Spectrum sensing can effectively improve the low utilization of spectrum resources and is one of the crucial components of cognitive radio networks. This paper proposes a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network cascaded spectrum sensing model. The model uses CNN to analyze the Short-Time Fourier transform spectrogram of the blind signal. Then the generated feature vector or feature map is passed to the LSTM according to the timestamp. Finally, it detects a signal in a specific spectrum and classifies the signal type to identify multiple signals accurately. The neural network model improves the detection probability by simultaneously acquiring the spatial and temporal characteristics of the blind signal. The experimental results show that the method in this paper can detect a variety of signals with higher detection probability within a wide range of SNR, especially under the condition of low SNR.