No abstract

No abstract

Spectrum sensing is an efficient technology for addressing the shortage of spectrum resources. Widely used methods usually employ model-based features as the test statistics, such as energies and eigenvalues, ignoring the temporal correlation aspect. Deep learning based methods have the potential to focus on various aspects, including temporal correlation. However, the existing ones are not good at capturing the temporal correlation features from spectrum data as traditional convolutional neural network (CNN) and long short-term memory network (LSTM) are used for feature extraction. Traditional CNNs were not designed to capture the global temporal correlations from time series data. Standard LSTM captures the temporal correlations based on the data collected from previous time slots only and cannot emphasize some important parts of a time series. This article proposes a data-driven deep learning based model to classify the received raw signals automatically, where the received signal data is considered time-series data. The proposed deep neural network (DNN) model is mainly featured with 1-dimensional convolutional neural network (1D CNN), bidirectional long short-term memory network (BiLSTM), and self-attention (SA). The 1D CNN and BiLSTM are responsible for extracting the local features and global correlations from the time series data, and BiLSTM could extract sufficient features in opposite directions. The SA layer enables the classifier network to emphasize those important features obtained by BiLSTM. The sim-ulation results demonstrate that our model performs better than a number of existing DNN models in terms of the probabilities of missed detection and false alarm, especially when the signal to noise ratio is low. Moreover, the impacts of the modulation scheme and sample length on the detection performance are studied.

No abstract

No abstract