Final Year Project — 28 GHz 5G Communication
Deep Neural Network for real-time phased array beam steering.
Enabling reliable ship-to-satellite communication at millimeter-wave.
Architecture
Simulate 4-element ULA with mutual coupling at 28 GHz using MATLAB Phased Array Toolbox. Generate 121 training samples across ±60°.
Train custom neural network (1→64→128→64→8) with sin/cos phase encoding. Backpropagation from scratch, no Deep Learning Toolbox needed.
Three-way comparison: Mathematical model vs DNN vs Optimal. DNN achieves 1.38° MAE with mutual coupling compensation.
Predict phase shifts, voltages, and power in <1ms. Hardware-ready outputs for 6-bit phase shifter with 0–5V control.
Neural Network
Interactive
Adjust the steering angle to see predicted phase shifts, voltages, and radiation pattern for the 4-element array.
Performance
Over fixed antenna (no beam steering)
vs 50% without steering during LEO pass
Improvement over no-steering baseline
Pointing error despite ship roll ±8°
| Metric | No Steering | Mathematical | AEP-DNN |
|---|---|---|---|
| Link Availability | ~50% | ~75% | ~90% |
| Avg Channel Capacity | ~100 Mbps | ~200 Mbps | ~250 Mbps |
| Pointing Error | Large | ~5° | <1° |
| Coupling Compensated | No | No | Yes |
Technical