GNSS Signal Anomaly Detection using DCB Estimates and Machine Learning Algorithms

Kurzfassung

The increasing number of Global Navigation Satellite System (GNSS) applications with reliability, high accuracy and integrity demands requires the implementation of appropriate monitoring systems. Monitoring can be used for various kinds of aspects, the assessment of the nominal performance and service quality of a system, failure detection or environmental analysis like radio frequency interference (RFI) detection, multipath analysis or others. The input data can originate from a single station for the analysis of local issues, but with limited timely observation capability according each space vehicle, up to globally distributed receiver networks to be able to observe every space vehicle of the GNSS continuously. Monitoring input data can range from raw samples, or multi-correlator data up to processed data like code and carrier phase pseudo-ranges or differential code bias (DCB) estimates. In this work we focus on methodologies for GNSS satellite signal anomaly detection. Previous studies have mainly employed received signal power respectively carrier-to-noise density ratio (C/N0) observations or multi-correlator approaches like evil waveform detectors to find GNSS errors in the satellite transmitter chains (Fenton and Jones 2005, Phelts and Walter 2022, Allende-Alba et al. 2022). Steigenberger et al. (2019) showed DCB variations as an effect of GPS flex power operation, where the power configuration between individual signal components have been changed with the result of slightly varying characteristic of the satellite transfer function and consequently signal deformation changes. Considering the availability of input data for a globally and continuously operation of an anomaly detection tool we have to exclude multi-correlator approaches, since no global data sources are available based on our knowledge. Globally and continuously available data are provided by the tracking network of the International GNSS Service (IGS, Johnston et al. 2017) in terms of real-time data streams or offline data stored in the Receiver INdepentent EXchange (RINEX) format. RINEX data of about 160 globally distributed stations provided the basis for the estimation of GPS DCBs for two years (2021 - 2022) with the approach of Montenbruck et al. (2014). This DCB time series with a sampling of two hours acts as input for our proposed anomaly detection methodology. The challenge is that we may not only have unintended GPS signal anomalies in the estimated DCBs. Since GPS flex power is still active and can be considered as an abrupt change in the satellite payload behavior, it shows up like anomalies. Therefore, we need to separate or ideally mitigate all intendedanomalies from the unwanted ones as well as taken external effects like multipath or interference into account. In order to tackle this challenge, we have applied a machine-learning-based methodology that includes the selection of relevant features and training data as well as the test of supervised and unsupervised algorithms. For validation purposes we used data sets that include known anomalies, like the ones analyzed by Phelts and Walter (2022).