Indoor localization of LoRa wireless modules based on RSSI fingerprint method using transfer learning

Fingerprint-based methods have gained popularity for indoor localization to their cost-effectiveness and flexible application. This paper presents an approach for indoor localization using LoRa technology using Received Signal Strength Indicator (RSSI) fingerprinting enhanced with transfer learning and COMSOL Multiphysics modeling. Traditional fingerprinting methods, although accurate, require extensive data collection, which is time-consuming and labor-intensive, especially in large or complex environments. The proposed solution addresses this issue using transfer learning, which enables generating a new radio map with minimal RSSI data, effectively reducing the measurement effort by up to 90% while maintaining sufficient localization accuracy. Five machine learning models were tested, with XGBoost achieving the best results (R² = 0.96172 without estimated distance (ED) and R² = 0.99933 with it). The transfer learning approach was tested using four scenarios, confirming that 20% or even 10% of the data is enough to achieve high prediction accuracy. COMSOL simulation further improved the localization accuracy, achieving an R² of 0.91566 and MAE of 0.33020 m under optimal conditions.