Previous Abstract Return to Session B3 Next Abstract

Session B3: AI and Machine Learning for GNSS Signal Processing, Error Modeling, and Threat Detection

Using Machine Learning for Enhanced Global Ionosphere Mapping in Trimble’s RTX Correction Service
Ulrich Weinbach, Markus Brandl, Andreas Goss, Michael Graml, Xiao Liu, and Moritz Rexer, Trimble Inc.
Location: Windsong 7-8
Date/Time: Thursday, Sep. 17, 11:26 a.m.

CenterPoint RTX is a state-of-the-art Precise Point Positioning (PPP) service operated by Trimble. The service is based on the estimation of precise satellite orbits, satellite clocks, satellite biases and atmosphere models using data from a global network of GNSS reference stations. The corrections are delivered to users via L-band geostationary satellite signals or the internet. Trimble RTX supports all major GNSS constellations: GPS, GLONASS, Galileo, BeiDou, and QZSS. The service achieves a 95% positioning accuracy of 2.5 cm, with convergence times of under 3 minutes worldwide for the RTX Standard service and less than 1 minute when using the RTX Fast services available in Europe and North America which make use of highly precise regional ionospheric and tropospheric models.
In order to speed up the convergence and to enhance the robustness of the positioning solution in challenging environments, the RTX Standard service features a global ionosphere model that utilizes spherical harmonic coefficients to describe the vertical electron content (VTEC) of the ionosphere. Despite the fact that the electron density distribution in the ionosphere is a complex and dynamic 3-dimensional structure it is common practice to assume that all electrons are concentrated in an infinitesimal thin shell at a fixed height (typically 300-500 km) above the Earth’s surface to simplify and stabilize the estimation process of the ionosphere model. Based on this, so called “single-layer”, assumption an isotropic mapping function (MF), solely relying on the elevation angle and the single-layer height, is typically used to relate the slant total electron content (STEC) of a given GNSS receiver-satellite link to the corresponding value of the VTEC provided by the ionosphere model. This approach is also adopted for the global RTX ionosphere model and generally works well during periods of moderate ionospheric activity in mid-latitudes but may introduce large errors in low latitudes and in the presence of strong ionospheric gradients.
In this contribution we present a new machine learning-based approach to improve the VTEC-to-STEC mapping based on historic data. Taking advantage of the large RTX data archive comprising 1 Hz GNSS measurements of hundreds of world-wide distributed GNSS reference stations and global RTX ionosphere models for the last 15 years, we have trained artificial neural networks with different architectures and input features that aim to provide a more realistic mapping of the VTEC by learning daily and seasonal patterns in different parts of the world. To achieve the best results, we incorporated a wide range of data available at the rover receiver. This information included the measurement direction, the geographic location of the ionospheric pierce point (where the ray path intersects the single-layer shell), the model Vertical Total Electron Content (VTEC), the time of day, and other relevant features. Extensive validation experiments indicate significant improvements of the ionospheric slant delay accuracy and consequently RTX rover convergence performance when using the neural network-derived mapping factor.
The specific improvement highly depends on the rover location (especially the latitude), the availability of training data for a particular location, and fluctuations in the solar activity. On average we observed an improvement of approximately 25 % in terms of the 1-sigma horizontal and vertical position convergence times for a set of 20 globally distributed test stations. The largest improvements can generally be observed at locations between 20° and 50° latitude in both the northern and southern hemisphere. Furthermore, this work will address the long-term stability of the predicted mapping factors and the necessity of implementing safeguard measures (guard rails) for the neural network-derived mapping factor, especially when operating in areas with limited or no training data.

Previous Abstract Return to Session B3 Next Abstract