An Automatic Modulation Recognition Model with Multi-Scale Triplet Attention

Automatic modulation recognition (AMR) plays an important role in wireless communications. In recent years, AMR schemes based on deep learning (DL) have been proposed by researchers, where convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are adopted to extract discriminative features from the received signals. However, it is still a challenging task to achieve high AMR accuracy. In this paper, we propose a new DL-based AMR model, the backbone of which consists of a CNN module and a long short-term memory (LSTM) module. Particularly, we propose a multi-scale triplet attention (MSTA) module to refine the features extracted by the CNN module and thus improve the AMR accuracy. Experiment results on the RML2016.10a dataset reveal that the proposed AMR model achieves an average accuracy of $63.22 \%$ surpassing state-of-the-art AMR models considered.