Deep Learning-based Automatic Modulation Classification for Wireless OFDM Communications

This paper proposes a convolutional neural network (CNN) based method to blindly identify the modulations of multi-carrier signals in wireless orthogonal frequency-division multiplexing (OFDM) communications. In this work, we develop a novel data reformation scheme to convert a series of complex envelope samples to a high-dimensional array of amplitude and phase samples, enabling deep networks to calculate the correlations between samples within every OFDM symbol and among different symbols. To this end, we design a CNN with several processing blocks integrating residual connection and attention connection to improve learning performance. Based on simulations, the proposed method outperforms a baseline approach when evaluated on a synthetic signal dataset with the presence of frequency-selective multipath fading, additive noise, and Doppler shift.