A Deep Learning Based Signal Detection Framework for Non Orthogonal Multiple Access Systems

This paper presents GVDET-Net, an innovative signal detection framework designed to enhance the detection accuracy and operational efficiency of NOMA systems utilizing non-orthogonal time-frequency resources. The proposed model integrates VGG19-based CNN layers with GRU layers to jointly extract spatial and temporal dependencies from input data. By sequentially processing hierarchical features, GVDET-Net achieves superior NOMA channel signal detection compared to ML, LS, and MMSE approaches across SNRs from 4 dB to 28 dB. Simulation results demonstrate its effectiveness under realistic NOMA conditions, outperforming SIC-LS and SIC-MMSE under multiple test scenarios with 64 and 16 pilot configurations for dual-user cases. GVDET-Net achieves a minimum Symbol Error Rate (SER) of approximately 10⁻³ at high SNR levels, delivering significant performance gains. Additionally, the model attains 96.4% classification accuracy, 3.1 ms inference latency for standard packet sizes, and an AUC score of 0.968, validating its robustness and real-time applicability. This work introduces advanced detection techniques for NOMA systems, paving the way for optimized wireless networks and supporting next-generation communication standards.