Seleção de parâmetros de transmissão de redes LoRa através de técnicas de aprendizado de máquinas

Abstract

The increase in LoRa networks density lead the LoRaWAN protocol’s Adaptive Data Rate (ADR) mechanism to show a drop in performance, in terms of packet delivery rate and energy consumed, since internal interference due to concurrent transmissions increases. Such interference increases the dynamics of the environment, requiring a proactive mechanism to adjust transmission parameters to guarantee the capacity of the links. Furthermore, the ADR mechanism is unable to adjust the parameters when the radio signal is degraded, which can occur, for example, if the network is subjected to external interference. This dissertation proposes the use of supervised learning and reinforcement learning techniques to adjust the transmission parameters. The decision tree and support vector machine supervised learning techniques are used to select the transmission parameters and, at the same time, reduce the power consumption. The reinforcement learning technique Q-learning is used to adjust the transmission parameters online, mapping the probability of collision occurrence for each selected parameter and optimizing the packet delivery rate and the energy consumed in the transmission. Two case studies are evaluated. The first represents a dense network with the gateway located in the center of the topology, while the second represents a network with devices located among 5 and 10 km subjected to an external source of interference. In the first case, the results show that, considering a number up to 100 devices, the performance of the Q-learning technique outperforms the ADR mechanism. In the second case, the ADR mechanism is unable to maintain the capacity of the links and the performance of the Q-learning technique indicates that this can be an effective alternative to dynamically adjust the transmission parameters.