Machine Learning-Driven Channel Loss Modeling for Sigfox Technology

The Internet of Things (IoT) is a fundamental component of Industry 4.0, enabling seamless device interconnectivity. In this context, Low Power Wide Area Networks (LPWAN), such as Sigfox, play a critical role because of their low energy consumption and long-range capabilities. However, losses in wireless channels remain a significant challenge, particularly in urban environments, where network coverage and reliability are often compromised. Traditional propagation models struggle to account for the complexities of varying terrain, urban density, and environmental factors, resulting in suboptimal network planning and performance. To address these challenges, this study aimed to enhance the accuracy of channel loss models for low-power long-range communication systems, specifically for the Sigfox network. We evaluated a combination of traditional empirical models, curve-fitting regression models, and advanced machine learning (ML) techniques. Using RSSI measurements obtained from a Site Survey in Quito, Ecuador, we compared the performance of these models using Root Mean Square Error (RMSE) as the evaluation metric. The empirical Egli model provided a baseline with an RMSE of 12.5%, while the logarithmic regression-based curve fitting model improved this with an RMSE of 9.2%, reflecting a 26.1% reduction in error. Building upon these results, regression-based ML models demonstrated superior performance in predicting channel losses for the Sigfox 920 MHz frequency band. In particular, the XGBoostAdvanced model achieved the most significant improvement, with an optimized RMSE of 5.5%, representing a 37.2% reduction in error compared to the logarithmic model. These findings demonstrate that ML-based approaches significantly improve channel loss prediction.