This project explores the application of deep learning techniques in physical-layer communication systems, focusing on the use of autoencoders to jointly optimize transmitter and receiver performance. Traditional communication systems are typically broken down into distinct blocks (source coding, channel coding, modulation), each optimized separately. In contrast, the channel autoencoder approach treats the communication system as an end-to-end learning problem, offering new ways to improve communication performance.
The project compares traditional communication methods with autoencoder-based approaches in both simulation and real-world environments, using SISO (Single Input Single Output) and SIMO (Single Input Multiple Output) systems. The experimental setup is based on GNU Radio and USRP B210 hardware.
-
Explore Factors Impacting Deep Learning in Communication: Investigating key factors that influence the performance of deep learning models in physical-layer communication systems.
-
Compare with Traditional Baselines: Benchmarking autoencoder-based methods against traditional communication systems in SISO and SIMO setups.
-
Test in Real-World Scenarios: Deploying the models in over-the-air transmissions using GNU Radio and USRP B210 hardware.
The autoencoder used in this project consists of:
- Encoder: Converts the one-hot encoded message into a complex IQ symbol for transmission.
- Channel: Simulated or real channel (AWGN, Rayleigh fading, etc.) over which the symbol is transmitted.
- Decoder: Receives the IQ samples and estimates the transmitted message.
The negative log-likelihood loss is used, and the model is optimized using the Adam optimizer.
- AWGN Channel: Additive White Gaussian Noise channel, used as the simplest case.
- Rayleigh Fading Channel: Modeled using filtered Gaussian noise and sum-of-sinusoids methods.
- Pulse Shaping: Implemented using a root-raised cosine filter to reduce the signal bandwidth and make the model more realistic for over-the-air transmission.
-
Phase 1: Simulated Channel Training: The autoencoder is trained using a simulated channel model, aiming to approximate real-world channel behavior.
-
Phase 2: Fine-tuning on Real Channel Data: After deployment, the receiver is fine-tuned using actual IQ samples collected from over-the-air transmissions to account for any mismatch between the simulated and real channels.
The hardware setup includes:
- USRP B210: SDR board used for transmitting and receiving over-the-air signals.
- GNU Radio: Software platform for signal processing.
- Zadoff-Chu Sequence: Used for frame synchronization, taking advantage of its low autocorrelation sidelobes.
Experiments are conducted in an office environment, with the following configuration:
- SISO System: Evaluated using 16-QAM modulation, compared with minimum distance decoding as a baseline.
- SIMO System: Evaluated with two receiver antennas, compared with Maximal Ratio Combining (MRC) as the baseline.
- AWGN Channel: The autoencoder outperforms traditional minimum distance decoding across all Signal-to-Noise Ratios (SNRs).
- Rayleigh Channel: The autoencoder performs better than traditional Zero Forcing techniques at higher SNRs but underperforms at lower SNRs.
- AWGN Channel: The autoencoder lags behind MRC at lower SNRs but surpasses it at higher SNRs.
- Rayleigh Channel: The autoencoder outperforms MRC across all SNRs.
The real-world experiments revealed higher Symbol Error Rates (SER) than simulations. This discrepancy is mainly due to the lack of symbol synchronization and unmodeled real-world factors like sampling time offsets and carrier frequency offsets.
- Symbol Synchronization: Implementing symbol synchronization to reduce over-the-air error rates.
- Improved Channel Models: Incorporating carrier frequency offset and sample time offset to bridge the gap between simulation and real-world performance.
- Advanced Architectures: Exploring the use of more sophisticated deep learning architectures, such as convolutional neural networks, recurrent neural networks, or transformer-based models for further improvement.