Session B5a: Innovative Navigation Algorithms

Black-Boxing GNSS Signals Post-Processing Through Machine Learning for Multi-Agent Collaborative Positioning of IoT Devices
Alex Minetto, Department of Electronics and Telecommunications of Politecnico di Torino; Daniele Jahier-Pagliari, Control and Computer Engineering of Politecnico di Torino; Angela Rotunno, Department of Electronics and Telecommunications of Politecnico di Torino, Punch Torino S.p.A.
Location: Beacon A
Date/Time: Thursday, Jan. 25, 2:58 p.m.

Peer Reviewed

Nowadays, GNSS (GNSS) receivers are embedded in a variety of electronics devices, and a growing number of users rely on them to track their position, velocity, and time. The density of Global Navigation Satellite System (GNSS) receiver has especially increased in urban areas with the advent of small-scaled IoT devices. Due to the limited GNSS signal power and the variability of the environment, continuous GNSS signal tracking may represent a demanding task for receivers, which, in addition, have to perform demodulation of the navigation message to provide the user with meaningful information. Furthermore, when operated by low-power platforms, such as Internet of Things (IoT) devices, the aforementioned tasks may quickly drain the battery. On the other hand, the low-power-consumption network connectivity hosted by IoT electronics could represent an ideal environment to enable new patterns for their state estimation, based on collaborative, multi-agent Position Navigation and Time (PNT) methods that would not imply continuous operation of the embedded GNSS receiver. This preliminary study aims to understand whether machine learning techniques could support such paradigms for position estimation in IoT devices. We generated an artificial environment, where conventional GNSS share their multi-satellite delay-Doppler matrices and their positions. These data are meant to be used to estimate a “reference” IoT receiver position. We used two open-source libraries, XGBoost, to implement a gradient-boosted decision tree, and Keras, to implement a multi-layer perceptron. The estimation error on the IoT receiver position obtained using machine learning tools is lower (typically, 10 % to 20 %) than the estimation error returned by a simplistic reference model, based on the arithmetic average of the networked receivers’ positions. The results suggest that rudimentary ML algorithms can extract fundamental information from collaborative users, thus opening new frontiers to collaborative GNSS navigation.