Scholarlink

<p>Abnormal observations caused by strong multipath effects, diffraction, and non-line-of-sight (NLOS) signals in constrained observation environments can severely degrade positioning accuracy and reliability. Therefore, effective anomaly detection is essential for ensuring high-precision GNSS positioning performance. While machine and deep learning-based AI methods are widely used for anomaly detection, their limited interpretability remains a challenge. To address this limitation, we propose a novel interpretability analysis framework based on eXplainable Artificial Intelligence (XAI). The framework performs post-hoc interpretability analysis on clustering or classification-based anomaly detection results, enabling validation of feature selection rationality and model reliability. The framework leverages feature marginal effects to characterize the local decision boundaries of anomaly detection models and integrates SHAP values to ensure globally consistent feature attribution. To validate the proposed framework, we conduct comprehensive case studies using GNSS datasets collected from two highway slope monitoring stations characterized by distinct and systematic occlusion conditions. In such environments, satellites affected by similar occlusion geometries are subject to consistent NLOS effects, leading to spatially coherent anomaly patterns. By leveraging this physical consistency, our framework validates the physical meaningfulness and interpretability of clustering-based pseudo-labels. The results demonstrate the complete workflow of the XAI-based analysis, from feature attribution to decision validation, thereby enhancing transparency, traceability, and reliability in AI-driven GNSS anomaly detection. Empirical results further show that after excluding the abnormal observations, the ambiguity fix rate improves significantly, with both stations achieving a fix rate exceeding 90%. This performance boost confirms the practical effectiveness of our XAI-validated framework.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An XAI-based framework for GNSS observation data anomaly detection in constrained observation environments

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Dongxu Li,
  • Yuan Du,
  • Guanwen Huang,
  • Qin Zhang

摘要

Abnormal observations caused by strong multipath effects, diffraction, and non-line-of-sight (NLOS) signals in constrained observation environments can severely degrade positioning accuracy and reliability. Therefore, effective anomaly detection is essential for ensuring high-precision GNSS positioning performance. While machine and deep learning-based AI methods are widely used for anomaly detection, their limited interpretability remains a challenge. To address this limitation, we propose a novel interpretability analysis framework based on eXplainable Artificial Intelligence (XAI). The framework performs post-hoc interpretability analysis on clustering or classification-based anomaly detection results, enabling validation of feature selection rationality and model reliability. The framework leverages feature marginal effects to characterize the local decision boundaries of anomaly detection models and integrates SHAP values to ensure globally consistent feature attribution. To validate the proposed framework, we conduct comprehensive case studies using GNSS datasets collected from two highway slope monitoring stations characterized by distinct and systematic occlusion conditions. In such environments, satellites affected by similar occlusion geometries are subject to consistent NLOS effects, leading to spatially coherent anomaly patterns. By leveraging this physical consistency, our framework validates the physical meaningfulness and interpretability of clustering-based pseudo-labels. The results demonstrate the complete workflow of the XAI-based analysis, from feature attribution to decision validation, thereby enhancing transparency, traceability, and reliability in AI-driven GNSS anomaly detection. Empirical results further show that after excluding the abnormal observations, the ambiguity fix rate improves significantly, with both stations achieving a fix rate exceeding 90%. This performance boost confirms the practical effectiveness of our XAI-validated framework.