Modulation Classification for Overlapped Signals Using Deep Learning

Modulation recognition (MR) plays a pivotal role in tasks encompassing spectrum management, security enforcement, and interference mitigation. This research work introduces a deep learning-based MR method tailored for completely overlapped multi-user (MU) signals. A custom convolutional neural network (CNN) model has been developed to classify images of super-constellations. A super-constellation is defined as the mapping of superposed symbols from multiple users in the in-phase/quadrature (I/Q) plane. The proposed approach performs MR blindly in the presence of practical impairments, such as carrier frequency offsets (CFO), noise variance and timing offset. The proposed algorithm estimates the symbol rate from the received overlapped signal and downsamples it, enabling extraction of a super-constellation image. This image is subsequently processed by the custom CNN model for modulation classification. The study considers a modulation pool comprising four modulation schemes, namely {Binary-Phase Shift Keying (B-PSK), Quadrature-PSK (Q-PSK), 8-PSK, and 8-Quadrature Amplitude Modulation (8-QAM)} at each user, which leads to a twenty-class classification problem. Simulation results demonstrate that the developed method achieves an average classification accuracy of 90% at 10 dB signal-to-noise ratio (SNR) and radically outperforms the cumulants-based methods in the existing literature. The simulation results also show the model’s robustness against variations in the signal parameters.