Modulation Recognition Based on Hybrid Deep Learning Networks

Abstract A hybrid deep learning network is proposed for recognition of both analog modulation and digital modulation. This scheme consists of two different networks. First, parameter estimation and transformation are introduced to correct the phase shift of the signal, and convolutional neural network and long short-term memory are combined to extract the spatial and temporal features of the signal, respectively. Second, convolutional neural network and Gated Recurrent Unit are used to classify easily confused signals, and global pooling is used to reduce the use of fully connected layers and reduce the amount of network parameters. Finally, the output features of the two networks are fused through the fusion layer, and the recognition of the signal modulation method is realized through two fully connected layers. To reduce the complexity of network training, a two-stage training scheme is used. The simulation of the proposed network on the standard data set RML2016.10A verifies that the proposed network effectively improves the recognition accuracy for both analog modulation and digital modulation.