EXTRACTING SPATIAL-TEMPORAL FEATURES USING DEEP LEARNING IN COOPERATIVE SPECTRUM SENSING

Abstract

In cognitive radio systems, spectrum sensing (SS) plays a vital role in detecting the presence of the primary user (PU). In this work, a cooperative spectrum sensing (CSS) model based on a graph convolution network and bidirectional long short-term memory (GCNBiLSTM)
is proposed. Specifically, the GCN architecture is applied to extract the relationship between the secondary users (SUs). Besides, the BiLSTM architecture learns the temporal correlation of sensing information at SUs. The presence of PU is decided based on spatialtemporal
features, which are combined from the outputs of GCN and BiLSTM. The proposed model is evaluated in a scenario of a dynamic channel (i.e., fading channel). Experimental results show that the GCN-BiLSTM model obtains a detection probability (Pd) of 84.5% and
an accuracy of 87.25% at a Signal-to-Noise Ratio (SNR) of -14 dB, demonstrating superior performance compared to the baseline models.