A Modulation Classification Method for Communication Signals Using TFA and Two-Stream Convolutional Networks

Automatic Modulation Classification (AMC) is a key enabler of software radio systems for wireless communications and non-cooperative communications. In this paper, we propose a modulation classification method based on Time-Frequency Analysis (TFA) and Two-Stream Convolutional Networks. The time-dimension information of signals as well as the time-frequency distribution generated by TFA methods are fed into the network as the inputs. The Two-Stream Convolutional Networks utilizes a Two-Dimensional CNN and a One-Dimensional CNN to extract the local correlation information and track the In-phase and Quadrature component (IQ) changes of the time series, respectively. A feature-level fusion among the concatenation of extracted features is then carried out before the softmax layer performing modulation classification. Simulation results show that the proposed model offers a superior classification performance compared to the traditional algorithms and is significantly easier to train than that of the deeper networks.