This project explores the application of Deep Learning (DL) to Non-Orthogonal Multiple Access (NOMA) in wireless communications. Specifically, it implements a decentralized neural network architecture to replace the traditional Successive Interference Cancellation (SIC) algorithm for downlink signal detection.
In traditional NOMA (Power Domain), multiple users share the same frequency and time resources but are assigned different power levels.
- Weak User (Far): High transmission power.
- Strong User (Near): Low transmission power.
Conventionally, the Strong User must perform SIC (decode Weak User
-
NOMA Simulation: Generates synthetic NOMA datasets with ordered Rayleigh fading (
$|h_1| < |h_2|$ ). - Decentralized Architecture: Utilizes two independent PyTorch neural networks (one per user) to mimic a decentralized receiver setup.
- Dynamic SNR Training: Trains the model across a range of SNR levels (5dB - 20dB) to ensure robustness.
- Benchmarking: Compares ML performance (BER/SER) directly against the traditional SIC algorithm.
- Visualization: Includes scripts for Spectral Efficiency comparison, Constellation diagrams, and BER/Throughput plots.
The notebook includes visualizations demonstrating the spectral efficiency gain of NOMA over OMA (Orthogonal Multiple Access).
- OMA: Users are separated by time slots (inefficient usage of bandwidth).
- NOMA: Users transmit simultaneously.
- User 1 (Weak): Allocated 80% Power.
- User 2 (Strong): Allocated 20% Power.
The system uses a Decentralized Detector comprised of two sub-networks:
- Net U1: Decodes the high-power signal (Weak User).
- Net U2: Decodes the low-power signal (Strong User).
Input Features: [Real(y), Imag(y), Real(h), Imag(h)]
Output: Probability distribution over QPSK symbols.