SilentAegis 🎯

6G ISAC Passive Radar: Physics-Informed Neural Network + UKF + Real Hardware Validation

---

📋 Overview

SilentAegis is a real-time passive radar pipeline for 6G Integrated Sensing and Communication (ISAC) that combines:

Component	Description
① ISAC	Bistatic 6G OFDM channel exploitation
② PINN	Physics-Informed Neural Network localizer
③ UKF	Unscented Kalman Filter tracker
④ NLOS	Multi-path exploitation as Virtual Anchors

---

🏆 Results

Simulation (28 GHz Urban Canyon)

Method	RMSE	P95	Inference
UKF-only (baseline)	80.9 m	137.6 m	—
PINN-only	5.27 m	8.45 m	0.0016 ms
PINN+UKF (ours)	4.40 m	6.97 m	0.0016 ms

Real Hardware Validation (Oryx Dataset, 3.75 GHz)

Model	Features	RMSE	P95
Model 1 — 5G (LOS)	14	0.115 m	0.168 m
Model 2 — 6G ISAC (LOS+NLOS)	135	0.030 m	0.045 m

Key findings:

• 🎯 Physics loss reduces RMSE by 76.7% on real hardware
• 📡 NLOS features provide 3.8x improvement over LOS-only
• ⚡ Ablation: 304x gain from Baseline → Full 6G ISAC
• 🔧 ONNX export: 16.4 KB | 0.088 ms/sample

---

🗂️ Repository Structure

SilentAegis/ ├── models/ │ ├── SilentAegis_5G_Final.onnx # 5G model (14 features) │ └── SilentAegis_6G_Final.onnx # 6G ISAC model (135 features) ├── notebooks/ │ └── SilentAegis_Training.ipynb # Full training pipeline ├── paper/ │ └── IEEE_SilentAegis_Final.docx # IEEE paper ├── figures/ │ ├── SilentAegis_GraphicalAbstract.png │ └── SilentAegis_Ablation.png └── README.md

🚀 Quick Start

import onnxruntime as ort
import numpy as np

# Load 6G ISAC model
sess = ort.InferenceSession('models/SilentAegis_6G_Final.onnx')

# Input: 135 features (LOS + NLOS from 7 TX-RX links)
features = np.random.randn(1, 135).astype(np.float32)

# Inference
position_xy = sess.run(None, {'features': features})[0]
print(f"Position: X={position_xy[0,0]:.3f}m, Y={position_xy[0,1]:.3f}m")

---

📡 Dataset

Real hardware validation uses the Oryx ISAC Dataset (TU Ilmenau, 2026):

• FC: 3.75 GHz | BW: 48 MHz | Δr: 3.13 m
• Hardware: USRP X310 SDR
• Target: VTOL UAV | Samples: 625 synchronized
• DOI: 10.71758/refodat.60

---

🔬 Pipeline

6G ISAC Signal ↓ Hybrid Ray-Tracing Engine (1100x faster than Sionna RT) ↓ Sparse Sionna RT Calibration (N=10, -89% delay error) ↓ PINN Localizer (Physics Loss + NLOS features) ↓ UKF Tracker (simulation) / Direct PINN (real HW) ↓ UAV Position Output

📊 Ablation Study

Config	Features	RMSE	Gain
① Baseline (R_sum only)	4	1.722 m	1.0x
② + ISAC (SNR+Doppler)	12	0.032 m	53x
③ + NLOS (4RX)	76	0.006 m	304x
④ Full 6G (7 links)	135	0.006 m	293x

---

📝 Citation

@article{elamir2026silentaegis,
  title={SilentAegis: Hybrid Ray-Tracing Passive Radar for 6G ISAC:
         Physics-Informed Neural Network with Sparse Sionna Calibration
         and Unscented Kalman Filter Tracking},
  author={ELAMiR, Mahmoud A.},
  journal={IEEE Access},
  year={2026},
  institution={Faculty of Engineering, Mansoura University}
}

---

👥 Authors

Mahmoud A. ELAMiR Faculty of Engineering, Mansoura University, Egypt

Supervisors:

• Assoc. Prof. Doaa A. Altantawy
• Dr. Seham Abd-Elsamee

---

📜 License

Copyright (c) 2026 Mahmoud A. ELAMiR Faculty of Engineering, Mansoura University, Egypt

All Rights Reserved.

This software and associated documentation files (the "Software") are the intellectual property of Mahmoud A. ELAMiR.

Permission is hereby granted for:

• Academic review and evaluation purposes only
• Citation in academic papers with proper attribution

The following are strictly prohibited without written permission:

• Commercial use
• Redistribution or republication
• Modification for derivative works
• Patent filing based on this work

For licensing inquiries: mahmoudahmedelamir@std.mans.edu.eg