Multi-Point Channel Charting for Long Range Wireless Networks: an Experimental Study

Long-range (LoRa) wireless networks have been widely proposed as efficient and scalable wireless access networks for battery-constrained internet of things (IoT) devices. In many practical IoT applications such as traffic monitoring and navigation, one of the main problems is finding the location of devices. Moreover, localization of LoRa-based IoT devices facilitates network functionalities such as spread factor (SF) management, power allocation, and grouping of LoRa nodes in multi-gateway networks. In this paper, the multi-point channel charting problem in a multi-gateway LoRa wireless networks is proposed in which multiple LoRa gateways map the multi-cell wireless characteristics to the spatial environment of network to assist the localization of IoT devices. In the formulated channel charting (CC) problem, the received power vector at the multiple LoRa gateways is used to provide a radio geometry chart of LoRa wireless networks. Then, the manifold learning techniques are applied to map the high-dimensional received power vectors to the two-dimensional spatial geometry of the user locations. In more detail, for visualizing the dissimilarity of high-dimensional received power values at different gateway antennas, the t-distributed stochastic neighbor embedding (t-SNE) method is applied over the received power vector at multiple gateways. Then, the unsupervised k-means algorithm is used to cluster the resulting low-dimensional data from t-SNE method. To analyze the performance of the proposed manifold learning-based mapping function on real dataset, a multi-gateway LoRa wireless network is designed and then deployed over the campus of Shahid Chamran University of Ahvaz. Extensive experimental measurements to collect spatiotemporal samples of received power over this LoRa wireless network with 12 gateways are performed. Our analysis on the experimental measurement shows that multi-point CC, which acts only on received power from mobile LoRa end nodes at 12 gateway antennas, can recover the spatial geometry of campus. Moreover, our practical results show that when the number of gateways increases from 6 to 12, the performance of clustering algorithm increases by 25%, on average. Moreover, decreasing the spread factor from 12 to 8 leads to more accurate mapping of campus areas.