A new machine learning based method for multi-GNSS data quality assurance and multipath detection

Global Navigation Satellite Systems (GNSS) based high-precision positioning techniques have been widely used in geodesy, attitude determination, engineering survey and agricultural applications. With the modernisation of GNSS, the number of visible satellites and new signals are increasing. Multi-constellation and multi-frequency data provide users with more observations, and hence increase redundancy in parameter estimation. However, increased number of satellites may increase the chance of multipath errors, especially in difficult environments. Therefore, this thesis aims at characterising the measurement signal quality of all available and new signals of multi-GNSS (GPS, GLONASS, Galileo, BDS, and QZSS) with real data. Also, a new multipath detection model based on machine learning methods is developed. The measurement noise levels in all currently available signals have been studied by investigating their double difference (DD) carrier phase residuals. The positioning precision, accuracy, and ambiguity success rate have been assessed using the selected individual GNSS constellations and their selected combinations with static and kinematic antennas in low multipath and severe multipath environments. The statistical results show the residuals vary from 0.5 mm to 2 mm with different signals and models of receivers. Short baseline tests show that in ideal conditions (i.e., a low multipath environment), using a single GNSS constellation (GPS, GLONASS, Galileo, or BDS) or their combinations can usually achieve millimetre-level precision and centimetre-level accuracy with almost 100% ambiguity success rates, regardless if the rover antenna is static or kinematic. In realistic condition (i.e. a severe multipath environment) the positioning precision and accuracy reduce to the centimetre level or even worse with a kinematic antenna. Multipath effect is one of the major error sources in GNSS positioning. Most of the currently available multipath detection and mitigation methods are based on stochastic modelling, advanced techniques in data processing, spatial geometry modelling, and special hardware designs. A new machine learning based multipath detection model is developed for undifferenced measurements (a single receiver approach). The approach is based on the fact that the multipath signature can be found in the multipath contaminated time series, and the features of multipath signature can be learned and identified by machine learning methods. The proposed model has been trained and validated with simulated data in an urban canyon environment with different satellite geometries. Moreover, the model has been trained and tested with real kinematic LoS and multipath data collected with a rotating arm rig in a multipath environment, and tested with multipath data collected near solar panels and near a building. The model has been tested using all available GNSS signals. The results show the model can achieve accuracy of 80%-90% with the simulated GNSS (GPS, Galileo, and BDS) data, and accuracy of 65%-70% with the real data collected using rotating arm rig on GPS L1 and GLONASS L1 signals. Real data collected near solar panels and near a building show that the well-trained model can achieve accuracy of about 60% in completely different multipath environments. The test results show the model was not well trained on GLONASS L2 and BDS data due to their carrier multipath errors are close to their carrier measurement error in ideal environment (low multipath environment)