Explainable Machine Learning for LoRaWAN Link Budget Analysis and Modeling.

Area of Science:

• Artificial Intelligence
• Wireless Communication Networks
• Internet of Things

Background:

• Precise planning of LoRaWAN networks is challenging due to complex propagation environments.
• Existing propagation models often lack accuracy for large-scale and dense IoT deployments.
• Machine learning offers potential for developing more effective LoRaWAN propagation models.

Purpose of the Study:

• To develop an effective propagation model for LoRaWAN using machine learning and empirical data.
• To decouple feature extraction and regression analysis for reduced training data requirements.
• To improve the accuracy of signal strength estimation and understanding of LoRa propagation mechanisms.

Main Methods:

• Utilized machine learning algorithms, specifically decision-tree-based gradient boosting, with empirically collected data.
• Proposed a novel approach of decoupling feature extraction and regression analysis.
• Conducted a comparative analysis to evaluate model performance using root-mean-squared error (RMSE).

Main Results:

• Achieved the lowest RMSE of 5.53 dBm with the gradient boosting model.
• Demonstrated model interpretability for qualitative observation of propagation mechanisms.
• Identified a 1.5 dBm sensitivity improvement with a spreading factor change from 7 to 12.
• Revealed a non-linear impact of clutter on signal attenuation.

Conclusions:

• The developed machine learning model provides a more accurate estimation of LoRa propagation.
• This work enhances the understanding of signal strength dependencies on various environmental factors.
• The findings mitigate challenges in large-scale LoRaWAN deployments, improving link budget analysis, interference management, and overall network efficiency for IoT.