Exploring Spatio-Temporal Dynamics for Spectrum Sensing: A GCN-LSTM Approach

Due to the rapid development of artificial intelligence, many spectrum sensing methods based on deep learning have emerged to show impressive performance, especially the convolutional neural network (CNN) can be used to extract network features of the secondary users (SUs) in cognitive radio networks. However, the CNN-based methods have low performance under the small number of antennas due to limited received information and their restriction to local connections in the covariance matrix. To address this issue, we propose a new spectrum sensing method where the graph convolutional network (GCN) is adopted to extract spatial features and the long-short term memory network (LSTM) is adopted to extract temporal features associated with the states of the primary user (PU). The experiments demonstrate the superior performance of the proposed GCN-LSTM method under various antenna numbers and signal-to-noise ratio (SNR) conditions.