Using Neural Networks to Predict GPS Observations for RTK Positioning
John M. C. Jackson, Graduate Student, Aerospace Engineering and Mechanics, University of Minnesota, Twin Cities
Real-time kinematic (RTK) GPS methods can produce position solutions that have centimeter-level accuracy. RTK operation requires a roving receiver, a reference receiver to provide observations, and a reliable data connection between the two. When the data connection is lost, the rover can use old observations from the reference receiver for a period before it loses its RTK solution and falls back to a non-precise GPS solution.
This study explores using neural network architectures for the online training and subsequent prediction of GPS observations from a reference station. Such a network can be used to maintain an RTK position solution at a rover if the data connection between the rover and the receiver is lost. This could be useful were maintaining a centimeter-level accurate position is mission critical (e.g. close-proximity infrastructure inspection) or where internet connectivity could be intermittent (e.g. rural environments).
The reference receiver could be a physical receiver near the rover, or it could come in the form of a virtual reference station (VRS). Some continuously operating reference station (CORS) networks will interpolate measurements from multiple, stationery reference receivers to generate a VRS and observation data near a rover. Using a CORS network often requires an internet connection to access the observation data. The solutions proposed in this study could be used for either of these reference station setups.
Preliminary work has aimed to replicate and expand work done for differential-GPS (DGPS) methods, which predict pseudorange and carrier-phase corrections rather than the direct observations. An autoregressive, moving average (ARMA) neural-network model of order (8,7) was built and predicted the carrier-phase measurement of a single satellite vehicle for 100s with an error of +/- 0.2% before diverging. Near-term work will expand this to multivariate measurements and an arbitrary number of satellites. Of interest are how recurrent neural networks (RNNs) will compare to standard neural networks. RNNs have a fully-connected topology that is useful in predicting non-linear time-series, and this could be a benefit to predicting GPS measurements from multiple satellites.
A simple setup will be used to gauge the efficacy of different neural network architectures. An RTK-capable receiver will be connected to a computer that can interface with the Minnesota CORS (MnCORS) network. Data from the MnCORS network will be piped directly to the receiver and fed into a neural network. Whilst training, the neural network will be optimized using online training strategies. After a certain duration, the MnCORS data will be cutoff and the neural network will generate messages to feed to the receiver. The receiver will use these artificial data to generate an RTK position solution.
To gauge how the error in observation prediction will affect the final position output of the RTK receiver, tests will be performed with respect to a National Geodetic Survey (NGS) marker of centimeter-level accuracy or another precisely-surveyed point. This experimental error will be compared with the theoretical RTK position error that one would expect to see from diverging observation predictions.
This study will work towards the more difficult idea of creating a generalized observation model using neural networks. Observation data is available on a continuous basis from a service like MnCORS. For a given locality, a neural network that incorporates memory (e.g. LSTM) could be trained for a long duration using MnCORS observations. At a high level, such a network could learn the nuances of the sidereal orbit time differences and induced effects of the atmosphere on observations. A generalized model using neural networks could then use a much smaller dataset to refine itself when it is being used in the field.
The anticipated results of this study include a comparison of different neural network architectures and their respective abilities to predict multivariate GPS observations and the computational resources it takes to train them. This work is significant in that it explores a useful tool to increase the robustness of a navigation system that requires centimeter-level accurate positioning.