Scholarlink

<p>Reliable real-time analysis of sensor data is essential for structural health monitoring (SHM) of high-value assets, where accurate prediction of the target field and uncertainty quantification are required for trustworthy decision-making. Such capabilities can help prevent structural fracture while reducing unnecessary maintenance and resource consumption. To reconstruct the target full-field from sparse sensor measurements, the high-dimensional full-field response is compressed into leading modes using principal component analysis (PCA). A Bayesian neural network (BNN) is then employed to predict these leading PCA modes from the sensor data. The parameters of the BNN are learned through pre-training followed by Hamiltonian Monte Carlo (HMC) sampling, allowing reconstruction of the target full-field with uncertainty quantification. Data-inherent aleatoric uncertainty is quantified through mode-wise variances during the pre-training stage, while epistemic uncertainty is quantified during HMC sampling. The framework was validated through cyclic four-point bending tests on carbon fiber reinforced polymer (CFRP) specimens with varying crack lengths. The experimental data comprised sparse strain gauge measurements and high-dimensional strain fields. Based on these data, the framework achieved accurate strain field reconstruction (R² &gt; 0.9) while simultaneously producing real-time uncertainty fields, remaining robust under sparse and noisy measurements. In addition, the full-field aleatoric and epistemic uncertainty estimates enable local diagnosis of whether low-confidence regions are driven by data-inherent noise or model-related limitations. Collectively, the results demonstrate that the proposed framework advances SHM toward trustworthy digital twin deployment and risk-aware structural diagnostics.</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

Real-Time Structural Health Monitoring with Bayesian Neural Networks: Distinguishing Aleatoric and Epistemic Uncertainty for Digital Twin Frameworks

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Hanbin Cho,
  • Jecheon Yu,
  • Hyeonbin Moon,
  • Jiyoung Yoon,
  • Junhyeong Lee,
  • Jinhyoung Park,
  • Seunghwa Ryu

摘要

Reliable real-time analysis of sensor data is essential for structural health monitoring (SHM) of high-value assets, where accurate prediction of the target field and uncertainty quantification are required for trustworthy decision-making. Such capabilities can help prevent structural fracture while reducing unnecessary maintenance and resource consumption. To reconstruct the target full-field from sparse sensor measurements, the high-dimensional full-field response is compressed into leading modes using principal component analysis (PCA). A Bayesian neural network (BNN) is then employed to predict these leading PCA modes from the sensor data. The parameters of the BNN are learned through pre-training followed by Hamiltonian Monte Carlo (HMC) sampling, allowing reconstruction of the target full-field with uncertainty quantification. Data-inherent aleatoric uncertainty is quantified through mode-wise variances during the pre-training stage, while epistemic uncertainty is quantified during HMC sampling. The framework was validated through cyclic four-point bending tests on carbon fiber reinforced polymer (CFRP) specimens with varying crack lengths. The experimental data comprised sparse strain gauge measurements and high-dimensional strain fields. Based on these data, the framework achieved accurate strain field reconstruction (R² > 0.9) while simultaneously producing real-time uncertainty fields, remaining robust under sparse and noisy measurements. In addition, the full-field aleatoric and epistemic uncertainty estimates enable local diagnosis of whether low-confidence regions are driven by data-inherent noise or model-related limitations. Collectively, the results demonstrate that the proposed framework advances SHM toward trustworthy digital twin deployment and risk-aware structural diagnostics.