DeepRx: Deep Learning-Based OFDM Receiver

📌 Overview

This repository contains the implementation of DeepRx, a Deep Learning-based receiver for Orthogonal Frequency-Division Multiplexing (OFDM) systems built with PyTorch. The project demonstrates how deep neural networks can be utilized for channel estimation, equalization, and signal detection, providing a robust alternative to traditional receivers in physical layer communications.

The performance of DeepRx is thoroughly evaluated and compared against a traditional Linear Minimum Mean Square Error (LMMSE) baseline receiver under various channel conditions, including AWGN and multipath fading environments (TDL channel models).

✨ Key Features

• Custom Deep Learning Architecture: Utilizes Preactivation ResNet Blocks and Depthwise Separable Convolutions for efficient and accurate signal processing.
• Comprehensive OFDM Pipeline: Includes QAM Modulation (QPSK, 16QAM, 64QAM, 256QAM), OFDM transmission, and realistic channel modeling with Doppler shifts.
• Dynamic Data Generation: On-the-fly generation of training datasets with variable SNRs and channel profiles.
• Baseline Comparison: Full implementation of a traditional LS/LMMSE receiver for direct Bit Error Rate (BER) comparison.
• Publication-Quality Visualization: Automated scripts to generate BER vs. SNR, BER vs. Doppler shift curves, and per-bit analysis.

📊 Performance Results

The proposed DeepRx model was evaluated under challenging conditions (16-QAM, TDL-B channel, variable SNR, and varying mobility). The results demonstrate clear advantages over the traditional LMMSE approach.

1. BER vs. SNR Performance

DeepRx demonstrates significant performance gains (achieving up to 1.5x improvement) over the baseline LMMSE receiver across a wide range of Signal-to-Noise Ratios.

2. Robustness to Doppler Shift

Under high mobility scenarios (simulating Doppler shifts up to 500 Hz), DeepRx maintains superior detection accuracy, proving its high robustness to rapid channel variations compared to traditional equalizers.

3. Per-Bit BER Analysis

A detailed bit-level analysis reveals that DeepRx consistently lowers the Bit Error Rate across all bit positions (both Most Significant Bits and Least Significant Bits) for the I and Q components in higher-order modulations like 16-QAM.

📂 Project Structure

• deeprx_model.py: Core DeepRx neural network architecture and custom loss functions.
• ofdm_system.py: Implements the OFDM transmitter, multipath channel models, and QAM modulation/demodulation.
• traditional_receiver.py: Baseline LMMSE receiver implementation.
• data_generator.py: PyTorch Dataset class for dynamic generation of training/validation OFDM symbols.
• train.py: The complete training pipeline with gradient accumulation and learning rate scheduling.
• evaluate.py: Evaluation scripts to compute BER across different SNRs and channel models.
• plot_results.py: Generates the publication-quality comparative plots shown above.

🚀 Usage

1. Training the Model

To train the DeepRx model from scratch:

python train.py
2. Evaluation
To evaluate the trained model's performance against the LMMSE baseline:

Bash
python evaluate.py
3. Plotting Results
To generate the performance graphs locally (saves to figures/ directory):

Bash
python plot_results.py
🛠️ Requirements
Python 3.8+

PyTorch

NumPy

Matplotlib