ConvLSTM-Based Spectrum Sensing at Very Low SNR
Qian Wang, Bo Su, Chenxi Wang, Liping Qian, Yuan Wu, Xiaoniu Yang
Abstract
Spectrum sensing can effectively improve the spectrum utilization. In practice, it is difficult to sense whether the spectrum is occupied or not due to the low signal energy at very low signal-to-noise ratio (SNR) (e.g., −20dB). To overcome this issue, this letter considers the correlation of the time-frequency domains, and proposes a ConvLSTM based spectrum sensing method. To be specific, we first apply the ConvLSTM network to extract the temporal and spatial features of the sensed IQ signals simultaneously, and then realize the low-SNR spectrum sensing according to the extracted features. Simulation results show that our proposed method can reduce the sensing error by about 25%, in comparison with other deep learning based spectrum sensing methods when the SNR is −20dB.