Análise de técnicas de Machine Learning aplicadas para geolocalização LoRa

Abstract

LoRa is a low-power, long-range wireless communication system that supports native geolocation, through the analysis of network metadata, without the need for other geolocation technologies (e.g. GPS). A commercial solution for this functionality is offered by the proprietary LoRa Cloud Geolocation service, based on conventional multilateration algorithms, whose prerequisite is the reception of transmissions from each device by at least three gateways simultaneously. However, the low accuracy is the main limitation inherent to the native LoRa geolocation, which can vary between 20 m and 2000 m. The systematic mapping carried out in this work revealed that 40% of the studies used some type of machine learning technique with the general aim of improving the levels of accuracy, of which artificial neural networks stand out due to their affinity with non-linearities and other complexities of propagation of the LoRa signal. However, there is a scarcity of studies that validate the neural network approach with data from real LoRaWAN networks. With this in mind, a series of basic models of dense neural networks (DNN) are tested, based on the concept of geolocation by fingerprinting, using metadata from stationary devices of a professional-private LoRaWAN network, which covers the area. urban area of ​​a city of approximately 200 thousand inhabitants. The implementation is characterized by a series of typical adversities for native geolocation, such as a low number of gateways, a large share of uplinks with less than three receiving gateways, and the Adaptive Data Rate (ADR) parameter enabled. As a result, it appears that, despite this set of adversities and the basic architecture of the neural network models used, it was possible to estimate the geographic coordinates of the devices with an average accuracy equivalent to that of the proprietary LoRa Cloud Geolocation service, even for devices with less than three receiver gateways, which points to an advantage over conventional multilateration.