Scholarlink

<p>This manuscript presents a scalable computational engine that integrates Finite Element-based Digital Twin (FE-DT) model with a feature-transfer-dependent damage-detection framework to assess structural dynamics at non-sensor locations and enable cross-structure damage identification. Unlike traditional numerical models, FE-DT models capture not only the geometric characteristics, but also the dynamic behavior of physical structures in complex and massive parallel computation. However, accurate observation at non-sensor locations remains challenging under sparse sensing specifically in high-performance computing scenario, leading to degraded prediction quality and limited availability of labeled data for training across unseen structures. To address this, the proposed framework integrates a computational efficient FE-DT model with unsupervised domain adaptation-based transfer learning. A local-global-local strategy is employed, where local sensor measurements (e.g., acceleration and displacement) are used to reconstruct global structural behavior within the digital twin and subsequently estimate responses at non-sensor locations. The FE-DT model is calibrated using Bayesian sampling optimization to handle non-convex parameter spaces. Furthermore, a Generative Adversarial Network (GAN)-based detector is trained on FE-DT-generated synthetic responses as the source domain, and adapted to target structures using unlabeled healthy data, enabling reliable damage identification without requiring labeled target-domain damage samples. The framework is designed as a computationally efficient and scalable workflow, where FE-based response generation and transfer learning stages can be executed in a high-throughput manner across multiple scenarios. In addition, an experimental validation on a beam structure demonstrates accurate reconstruction of structural responses under sparse sensing. Also, evaluation on some benchmark datasets shows effective discrimination between healthy and damaged states, achieving AUC values up to 0.99 with consistently high precision, recall, and F1 scores.</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

Integrated finite element-based digital twin model and scalable computational feature transfer framework for structural dynamics and damage detection

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Nidhi Agrawal,
  • Rehan Nath,
  • Aashritha Lakshmi Mallampati,
  • Keshav Goel,
  • Shashank Gupta,
  • Shuvendu Narayan Patel,
  • Devendra Gokul Patil

摘要

This manuscript presents a scalable computational engine that integrates Finite Element-based Digital Twin (FE-DT) model with a feature-transfer-dependent damage-detection framework to assess structural dynamics at non-sensor locations and enable cross-structure damage identification. Unlike traditional numerical models, FE-DT models capture not only the geometric characteristics, but also the dynamic behavior of physical structures in complex and massive parallel computation. However, accurate observation at non-sensor locations remains challenging under sparse sensing specifically in high-performance computing scenario, leading to degraded prediction quality and limited availability of labeled data for training across unseen structures. To address this, the proposed framework integrates a computational efficient FE-DT model with unsupervised domain adaptation-based transfer learning. A local-global-local strategy is employed, where local sensor measurements (e.g., acceleration and displacement) are used to reconstruct global structural behavior within the digital twin and subsequently estimate responses at non-sensor locations. The FE-DT model is calibrated using Bayesian sampling optimization to handle non-convex parameter spaces. Furthermore, a Generative Adversarial Network (GAN)-based detector is trained on FE-DT-generated synthetic responses as the source domain, and adapted to target structures using unlabeled healthy data, enabling reliable damage identification without requiring labeled target-domain damage samples. The framework is designed as a computationally efficient and scalable workflow, where FE-based response generation and transfer learning stages can be executed in a high-throughput manner across multiple scenarios. In addition, an experimental validation on a beam structure demonstrates accurate reconstruction of structural responses under sparse sensing. Also, evaluation on some benchmark datasets shows effective discrimination between healthy and damaged states, achieving AUC values up to 0.99 with consistently high precision, recall, and F1 scores.