AI-Enhanced Generalizable Scheme for Path Loss Prediction in LoRaWAN

• Published in: IEEE internet of things journal 2024-04, Vol.11 (8), p.14593-14606
• Main Authors: Chen, Mingyu, Zhang, Yan, Ji, Zijie, Briso-Rodriguez, Cesar, Zhang, Kaien
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Artificial intelligence; Artificial neural networks; Buildings; Computational modeling; Computer simulation; Data models; Feature extraction; Internet of Things; Internet of Things (IoT); long-range (LoRa); Machine learning; path loss (PL); Power consumption; Prediction algorithms; Predictive models; Root-mean-square errors; swin transformer; transfer learning; Wide area networks
• ISSN: 2327-4662; 2327-4662
• DOI: 10.1109/JIOT.2023.3342984

Long-range wide-area network (LoRaWAN) is a widely used technology in the Internet of Things (IoT), which provides long-range (LoRa) communication with low power consumption. In LoRaWAN, an accurate path loss (PL) model is essential to realize link budget and network coverage planning. In this article, we present an artificial intelligence (AI)-enhanced generalizable scheme for PL prediction in LoRaWAN. We propose a network that performs corrective adjustments to improve the PL estimates of empirical models. The network termed STransRadio benefits from the self-attention computation in Swin Transformer to model the LoRa correlation about propagation for enhancing the adjustment prediction accuracy. To generalize our scheme to new scenarios, an multiscenario deep transfer learning (MDTL) algorithm is proposed, which finetunes the pretrained STransRadio network with limited data. We conduct simulations and measurements in the 868-MHz bands to assess the performance of the scheme in terms of prediction accuracy and generalization ability. The effectiveness of the proposed scheme has been verified with both simulations and measurements. Moreover, the STransRadio network in the scheme outperforms the convolutional neural network (CNN) and deep vision transformer (DeepViT). With the MDTL algorithm, our scheme can achieve excellent prediction performances when it is applied in a new scenario with limited training data. Furthermore, we verify that the scheme utilized in the simulated scenario can be transferred to both the new simulated scenario and the realistic scenario. With only 100 samples, the scheme achieves root mean square error (RMSE) values of 7.27 and 5.96 dB between the predicted and actual PL, respectively.