Machine Learning Techniques for Collaborative, Multi-agent GNSS Positioning in IoT devices

Nowadays, Global Navigation Satellite Systems (GNSS) receivers are embedded in a variety of electronic devices, and a growing number of users depend on them to navigate themselves to a destination. GNSS technology allows the user to estimate its position through multilateration, which is based on satellite detection, the estimation of the signal Time-of-Arrival (ToA), and the subsequent measurement of the receiver-to-satellite ranges (commonly referred to as “pseudoranges”). The low power of GNSS signals makes the continuous satellite’s signal tracking a challenging task for the receiver, especially considering the tight constraints on resource usage. Internet-of-Things (IoT) electronics host “by definition” low-power-consumption network connectivity, and they could enable new patterns for the Position Velocity and Time (PVT) estimation, based on collaborative, multi-agent Positioning, Navigation, and Timing (PNT) methods that would not imply a continuous operation of the embedded GNSS receiver. This study aims to understand whether Machine Learning techniques could support such paradigms for the PVT estimation in IoT devices, in the attempt to avoid the need for costly continuous signal tracking, demodulation of the navigation message, and pseudorange construction and correction, by exploiting the information made available by a set of networked collaborative users. In particular, we aim to share multi-satellite delay-Doppler matrices and their associated position estimates gathered by conventional GNSS receivers operating within a network. The work has been developed in two different stages. The first step was generating an experimental environment, building a dataset inclusive of the IoT receiver’s and its surrounding networked receivers’ position information. In detail, we considered 4 networked receivers, randomly distributed within a 200m radius of the IoT receiver. For each receiver, the pseudoranges and the delay-Doppler matrices were simulated, as well as the PVT solutions. The next step was to use machine learning techniques to estimate the IoT receiver’s position, based on the information shared by the networked receivers. Two different machine learning open-source libraries have been tested: XGBoost (eXtreme Gradient Boosting), and Keras, to implement a neural network (namely, a multi-layer perceptron). Finally, the estimation of the IoT position obtained using the two machine learning tools has been compared to a simplified “reference” model, where the IoT receiver’s position is approximated by the arithmetic average of the networked receivers’ positions. The estimation error on the IoT receiver position obtained using machine learning tools is lower (typically, 10 to 20%) than the estimation error shown by the simplified reference model. This observation is confirmed when doing a further validation of the models, over geographic regions different from the one used to generate the training dataset. Concerning, instead, the positioning error, the position estimates are in general showing a 50m offset compared to the “true” position. Although such a distance is not negligible, and considering also that the work has been developed in an experimental environment (with all the limitations that this entails), this preliminary study suggests that the PVT estimation via machine learning could work, and its use in support of PVT estimation might be further investigated.