Modulation Classification in NOMA Systems using Lightweight Dual-Pooling CNN with Superposed Constellation Density Grids (SCDGs)

Automatic modulation classification (AMC) is essential in non-orthogonal multiple access (NOMA) systems, since it enables dynamic adaptation of modulation schemes to optimize spectral efficiency and reduce inter-user interference. To address the AMC of the interference users in NOMA downlink systems, a lightweight Dual-Pool convolutional neural network (Dual-Pool CNN) algorithm is proposed, which leverages the power of the pooling layers in CNNs for better classification accuracy. It synergistically combines max-pooling and average-pooling to extract both general as well as sharp features from the distorted, faded, and superposed constellation diagrams. Our extensive experimentation demonstrates that the proposed model can classify the signals of four different modulation schemes with an accuracy score of more than 90% at an SNR of 14 dB and above. Moreover, this higher accuracy is obtained at lower computational complexity using the superposed constellation density grid (SCDG) approach followed by lightweight Dual-Pool CNN architecture.