LPWAN Localization via RSSI/SNR Fingerprinting and Lightweight Machine Learning

---

Résumé

Accurate geolocation for low-power wide-area (LPWAN) devices is desirable when GNSS is unavailable or too energy-expensive, yet RSSI-/TDoA-based approaches are often fragile under channel variability, collisions and cross-device heterogeneity. We address this gap with a reproducible, tabular pipeline that maps LoRa RSSI/SNR/ToA and PHY metadata to 2D positions, compares strong tabular baselines (k-NN, Random Forest, LightGBM, XGBoost), and crucially evaluates them under group-aware (device-wise) splits to avoid identity leakage. On an ns-3-generated LoRa dataset of about 3.3 × 104 labeled receptions, Random Forest attains the tightest distribution with p50 ≈ 0 m and p95 < 1 m, whereas k-NN, despite a low median, exhibits a much heavier tail (p95 ≈ 187 m), underscoring the need to report both central and tail metrics. These results indicate that simple, edge-feasible models can perform gateway-side inference with robust accuracy when fed cleaned features and evaluated with realistic splits, making the approach attractive for practical LPWAN/IoT deployments.

Mots-clés

LPWAN, LoRA, Localization, Fingerprinting, RSSI, SNR, IoT, Machine Learning