A Transformer-based CTDNN Structure for Automatic Modulation Recognition

The application of deep learning enhances the ability of Automatic Modulation Recognition (AMR) to process a variety of complex types of signals, and improves the accuracy and speed of recognition. In order to solve the local dependency constraints in CNN and RNN, we propose a Transformer- based CTDNN structure for AMR to further improve the accuracy of recognition. First, the time-domain signal sequence is projected into a high-dimensional continuous space through embedding with a convolutional layer, and the local features of the signal are captured. What’s more, the encoder of Transformer is used to extract global features of the signal. Finally, the recognition result of the signal is output after the fully connected layer. In the simulation, the RML2016.10b dataset was used to analyze the structure and recognition results of CTDNN. And the comparison with existing methods shows that the structure has higher recognition accuracy, especially at low SNR.