Scholarlink

<p>Accurately inferring in-service loads and internal states is essential for life prediction, anomaly detection, and digital-twin updating of critical structures. Structural load and state inversion is hindered by non-stationary responses, multiple admissible solutions, and costly forward evaluations. We propose a sample-efficient inverse prediction method that combines Rotating Importance Acquisition (RIA) with Adaptive Interval Refinement (AIR). RIA enriches a Gaussian process (GP) surrogate by rotating a small leave-out subset to obtain out-of-subset likelihood exposure scores. These scores are fused with normalized predictive uncertainty so sampling concentrates on regions that are both misfit and uncertain, allocating more evaluations to locally non-stationary zones. AIR operates on the updated GP to achieve multi-solution inverse prediction: feasible parameter intervals are progressively refined, candidates are clustered into distinct solution basins to balance diversity, and representative solutions with dispersion measures are extracted until stability criteria are met. Benchmarks on nonlinear test functions show higher surrogate accuracy and better sample efficiency than standard GP, K-Fold ANN, and Interquartile Range (IQR) baselines. The inversion stage recovers multiple well-separated solution basins with low identification error. A turbine disk load reconstruction case using high-fidelity finite element analysis confirms robustness and practical applicability. The method provides a unified, data-efficient framework for non-stationary multi-solution structural load inversion and state 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

Non-stationary active learning for structural state inverse prediction

  • 期刊论文

  • 发表日期：

    2026年1月2日

查看全文
  • Jinfa Zhang,
  • Jian Gao,
  • Jiaqi Luo,
  • Diwang Ruan,
  • Shuyou Zhang,
  • Yiming Zhang

摘要

Accurately inferring in-service loads and internal states is essential for life prediction, anomaly detection, and digital-twin updating of critical structures. Structural load and state inversion is hindered by non-stationary responses, multiple admissible solutions, and costly forward evaluations. We propose a sample-efficient inverse prediction method that combines Rotating Importance Acquisition (RIA) with Adaptive Interval Refinement (AIR). RIA enriches a Gaussian process (GP) surrogate by rotating a small leave-out subset to obtain out-of-subset likelihood exposure scores. These scores are fused with normalized predictive uncertainty so sampling concentrates on regions that are both misfit and uncertain, allocating more evaluations to locally non-stationary zones. AIR operates on the updated GP to achieve multi-solution inverse prediction: feasible parameter intervals are progressively refined, candidates are clustered into distinct solution basins to balance diversity, and representative solutions with dispersion measures are extracted until stability criteria are met. Benchmarks on nonlinear test functions show higher surrogate accuracy and better sample efficiency than standard GP, K-Fold ANN, and Interquartile Range (IQR) baselines. The inversion stage recovers multiple well-separated solution basins with low identification error. A turbine disk load reconstruction case using high-fidelity finite element analysis confirms robustness and practical applicability. The method provides a unified, data-efficient framework for non-stationary multi-solution structural load inversion and state monitoring.