Short-time Modulation Classification of Complex Wireless Communication Signal Based on Deep Neural Network

Modulation classification of communication signal is one of the key technologies for realizing non-cooperative communication tasks, multi system communication interconnection and software radio. Therefore, when the decision process cannot wait for more data to increase certainty, how to effectively classify the modulation type in a short time is an unavoidable and challenging topic. In this paper, we make a performance comparison of traditional feature-based neural network and deep neural network (DNN) with complex digital modulation signal datasets. The results indicate that DNN has a stronger ability to extract classification features. Then we demonstrate two novel architectures based on DNN, which disentangle more meaning hidden features from the short-time signal and perform superiorly under limited signal length. Finally, we test the generalization ability of neural network models to signal-to-noise radio (SNR).