Spectrum Sensing Based on CNN-LSTM with Attention for Cognitive Radio Networks

Spectrum sensing has considerable application value in cognitive radio networks. However, spectrum detectors mostly depend on the signal-noise model in aspect of accuracy. To address this problem, most studies focus on the solutions based on deep learning such as convolutional neural networks (CNN) that are good at image processing and object detection, and long and short-term memory networks (LSTM) that do well in establishing time series prediction. These solutions are not affected by the signal-noise model. In this paper, we combine the attention mechanism in our spectrum sensing model and propose a spectrum sensing approach based on CNN-LSTM. The input set of the approach is derived from preprocessing and labeling the received signal. Firstly, we adopt the CNN to extract the energy-related features of the covariance matrix, which are fed into the LSTM network for channel state feature extraction. Next, in order to greatly promote the attention of the CNN-LSTM network to the feature of channel states, a convolutional block attention module (CBAM) combined with CNN is presented and employed to capture the important information in a larger receptive field, and CBAM can facilitate LSTM to capture channel states more effectively. Simulation results show that our proposed approach has an advantage over the methods based on CNN and LSTM in the accuracy performance, respectively, and is more robust than the CNN-LSTM method without attention mechanism under the low signal-to-noise ratios.