Airborne GNSS PPP Based Pseudolite System

Abstract

Airborne-Pseudolite (A-PL) systems have been proposed to augment Global Navigation Satellite Systems (GNSSs) in difficult service areas. One of the challenges in realising such a system is to determine the precise positions of the A-PLs. This research focuses on improving the A-PL positioning performance based on GNSS Precise Point Positioning (PPP). The main contributions are: 1. A-PL distributed positioning based on real-time GNSS PPP combined with inter-PL range measurements was studied. The short-term predictions of precise orbit and satellite clock corrections were analysed. Simulation tests have demonstrated that the A-PL using GNSS PPP combined with inter-PL range measurements is able to achieve better positioning performance than using the GNSS PPP-only approach. The prediction models for short-term orbit and satellite clock correction predictions can effectively reduce the impact of a disruption of communications on GNSS PPP positioning. 2. To deal with the unmodelled measurement errors for GNSS PPP two model-learning based Kalman filter (KF) algorithms were studied: least-squares support vector machine (LS-SVM) and Gaussian process regression (GPR). These two algorithms were evaluated using both static and kinematic experiments. The results confirm that both algorithms can effectively reduce the effect of unmodelled measurement errors on the positioning performance of GNSS PPP. 3. To realise the optimal integration and stable positioning performance for real-time multi-GNSS PPP, two types of stochastic models were assessed by a static experiment: the a priori stochastic models and the real-time estimated variance methods. The experimental results indicate that the a priori stochastic models based on real-time signal-in-space ranging error (SISRE) and real-time estimated stochastic models could all achieve better performance than the stochastic model based on satellite elevation angle, as used in conventional multi-GNSS PPP. 4. To select the optimal subset of satellites (and therefore measurements) for multi-constellation GNSS, an end-to-end deep learning network for satellite selection was proposed. An experiment was conducted with training and validation data from 220 International GNSS Service (IGS) stations. It was shown that the trained models are capable of selecting most of the contributing satellites with less computational time compared with the brute force approach of satellite selection.