Comparing Machine Learning Algorithms for RSS-Based Localization in LPWAN

Abstract

In smart cities, a myriad of devices is connected via Low Power Wide Area Networks (LPWAN) such as LoRaWAN. There is a growing need for location information about these devices, especially to facilitate managing and retrieving them. Since most devices are battery-powered, we investigate energy-efficient solutions such a Received Signal Strength (RSS)-based fingerprinting localization. For this research, we use a publicly available dataset of 130 426 LoRaWAN fingerprint messages. We evaluate ten different Machine Learning algorithms in terms of location accuracy, R2 score, and evaluation time. By changing the representation of the RSS data in the most optimal way, we achieve a mean location estimation error of 340 m when using the Random Forest regression method. Although the k Nearest Neighbor (kNN) method leads to a similar location accuracy, the computational performance decreases compared to the Random Forest regressor.