Towards rapid and reliable GNSS ambiguity resolution using residual-based machine learning in challenging environments

Abstract High-precision Global Navigation Satellite System (GNSS) positioning relies on the successful resolution of carrier-phase ambiguities. However, in challenging environments characterized by severe Non-Line-of-Sight (NLOS) reception and multipath effects, such as urban canyons with dense vegetation and high-rise buildings, conventional model-driven validation methods often suffer from model–reality discrepancies, leading to degraded ambiguity fixing rate and reduced reliability of ambiguity resolution. Machine Learning (ML) has recently attracted increasing interest in GNSS because of its ability to capture complex nonlinear relationships. However, its application to ambiguity validation remains limited by training on benign datasets, reliance on conventional statistical indicators, and high computational and memory requirements. To overcome these limitations, this study proposes a residual-based ML validator for ambiguity resolution in challenging environments, achieving improved generalization while maintaining real-time applicability. First, a set of residual-based features with reduced sensitivity to environmental degradation is introduced. Subsequently, a compact multilayer perceptron is trained as the core classifier. Using two real-world GNSS datasets collected from distinct platforms, the proposed validator is comprehensively evaluated through both classification and Real-Time Kinematic (RTK) positioning performance. The results show that the proposed validator consistently achieves accuracy and precision above 90% on both datasets, with the residual-based features identified as the most influential. In Unmanned Ground Vehicle (UGV) scenarios characterized by dense vegetation and severe building occlusions, the proposed validator achieves an average correct ambiguity fixing rate of 85.21%, compared to 77.26% for Fixed Failure-Rate Ratio Test (FFRT), and a wrong fixing rate of only 1.42%, versus 16.37% for FFRT. In the CAR-S5 scenario on typical urban roads with varying environments, the proposed validator achieves a correct fixing rate of 85.15%, improving by 10.55% over FFRT while maintaining a wrong fixing rate of 0.92%. Furthermore, experiments on the publicly available SmartPNT-POS dataset further validate its effectiveness. In conclusion, the proposed method significantly improves the success rate and reliability of ambiguity resolution in challenging environments and can be readily extended to other real-time high-precision positioning applications.