Scholarlink

<p>Global navigation satellite system–real-time kinematic (GNSS–RTK) is increasingly adopted in structural health monitoring (SHM) due to its centimeter-level precision and real-time capabilities. However, the high dimensionality, noise, and non-stationarity of GNSS time-series data pose critical challenges for reliable pattern recognition and anomaly detection. This study introduces an unsupervised learning framework tailored specifically for GNSS-based SHM, which integrates robust preprocessing, nonlinear dimensionality reduction, and adaptive time-series modeling. Key innovations include the application of t-distributed stochastic neighbor embedding (t-SNE) for enhanced latent structure discovery, and the introduction of an adaptive dynamic time warping barycenter averaging (ADBA) method to generate interpretable cluster-level signal prototypes. Experimental validation on real-world GNSS–RTK data collected from the Tran Thi Ly cable-stayed bridge demonstrates that the proposed framework outperforms conventional multidimensional scaling (MDS)-based approaches in terms of cluster separability and behavior interpretability. The approach effectively uncovers recurring temporal patterns&#xa0;such as daily peak displacements&#xa0;and isolates abnormal structural responses without the need for labeled data. This work highlights a promising direction toward scalable, low-cost, and data-driven SHM systems, offering a practical foundation for real-time anomaly detection and decision support in civil infrastructure monitoring.</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

Latent pattern recognition in GNSS-based SHM using t-SNE and adaptive time-series modeling

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Ho Thi Lan Huong,
  • Tran Duc Cong,
  • Le Van Vu,
  • Le Khanh Giang

摘要

Global navigation satellite system–real-time kinematic (GNSS–RTK) is increasingly adopted in structural health monitoring (SHM) due to its centimeter-level precision and real-time capabilities. However, the high dimensionality, noise, and non-stationarity of GNSS time-series data pose critical challenges for reliable pattern recognition and anomaly detection. This study introduces an unsupervised learning framework tailored specifically for GNSS-based SHM, which integrates robust preprocessing, nonlinear dimensionality reduction, and adaptive time-series modeling. Key innovations include the application of t-distributed stochastic neighbor embedding (t-SNE) for enhanced latent structure discovery, and the introduction of an adaptive dynamic time warping barycenter averaging (ADBA) method to generate interpretable cluster-level signal prototypes. Experimental validation on real-world GNSS–RTK data collected from the Tran Thi Ly cable-stayed bridge demonstrates that the proposed framework outperforms conventional multidimensional scaling (MDS)-based approaches in terms of cluster separability and behavior interpretability. The approach effectively uncovers recurring temporal patterns such as daily peak displacements and isolates abnormal structural responses without the need for labeled data. This work highlights a promising direction toward scalable, low-cost, and data-driven SHM systems, offering a practical foundation for real-time anomaly detection and decision support in civil infrastructure monitoring.