全球导航卫星系统应用
可解释性
计算机科学
稳健性(进化)
机器学习
人工智能
全球定位系统
贝叶斯推理
变量模型中的错误
异常检测
卡尔曼滤波器
数据挖掘
测量不确定度
重采样
不确定度量化
核密度估计
离群值
贝叶斯概率
可识别性
统计模型
噪音(视频)
均方误差
雷达
作者
Paul Dobre,Shichuang Nie,Hongzhou Yang
出处
期刊:Proceedings of the Satellite Division's International Technical Meeting
日期:2025-10-01
卷期号:: 2099-2109
被引量:1
摘要
GNSS positioning methods such as Kalman filters, factor graph optimization, and weighted least squares (WLS) have recently been complemented by machine learning (ML) models aimed at improving positioning accuracy and robustness. ML has been applied to GNSS for signal classification, anomaly detection, environmental inference, and position correction. However, in challenging conditions—such as urban canyons or unfamiliar scenarios—ML models face epistemic (model) and aleatoric (data) uncertainty, which can result in overconfident yet incorrect predictions that compromise system integrity. This work proposes an uncertainty-aware ML framework to enhance GNSS positioning error estimation by quantifying and incorporating both epistemic and aleatoric uncertainty into the model output. A spatial transformer is used to predict GNSS positioning error based on satellite-specific observation features. The model output includes both a positioning error estimate and its associated uncertainty, which enables more reliable integration with traditional GNSS solutions. Uncertainty is quantified through a model ensemble approach that aggregates predictions from multiple models to estimate uncertainty. The benefits of incorporating uncertainty include better anomaly handling, increased model interpretability, improved retraining strategies via active learning, and the ability to fall back to traditional methods in high-uncertainty situations. The preliminary experimental results demonstrate that incorporating uncertainty improves positioning interpretability and robustness compared to a standard ML-enhanced GNSS pipeline through effectively identifying out-of-distribution scenarios and high input noise and guiding system fallback decisions.
科研通智能强力驱动
Strongly Powered by AbleSci AI