Scholarlink

Purpose <p>Data-driven-based damage identification models rely heavily on the training data. However, collecting real damage data is often impractical, and data scarcity leads to poor extrapolation, reducing accuracy in identifying unseen damage events. A data-driven-model trained with finite element (FE) simulations of various damage events could help address the data scarcity issue. Yet, discrepancies between simulated and real data due to modelling errors may affect the damage identification performance.</p> Methods <p>This paper proposes a damage identification approach that integrates a FE model, machine learning (ML) model, and a transfer learning (TL) method to leverage the simulated training data for identifying actual damages in a plate-like structure through minimising the domain discrepancies. For the case study, stiffness-reduced single damage cases with low and high severity levels at plate supports were investigated. Support Vector Machine (SVM) was used as the ML model, with the first mode shape differences used as the damage-sensitive features. A semi-supervised domain adaptation approach was used as the TL method. Simulated data from the FE plate model was used as the source dataset, while limited experimental data was used as the few-shot labelled target data for domain adaptation. The relationship between the average discrepancies in dynamic characteristics and the unseen damage identification accuracy was investigated to study the effectiveness of the proposed method. The performance of Semi-Supervised Transfer Component Analysis (SSTCA) and Semi-supervised Maximum Independence Domain Adaptation (SMIDA) were compared.</p> Results <p>Results showed that SMIDA outperforms SSTCA, with the average improvement rates in damage identification accuracy of 50.39% for 1-shot SMIDA and 74.85% for 3-shot SMIDA when the average deviation in dynamic characteristics ranged from 3% to 32% due to varied boundary conditions. The trend where the accuracy reduces when the deviation increases is observed when SMIDA was used, suggesting the effectiveness of SMIDA in minimising the domain discrepancies due to deviation in dynamic characteristics.</p> Conclusion <p>The findings showed the potential of TL in leveraging FE models for data-driven-based damage identification through minimising the domain discrepancies due to deviation in dynamic characteristics. Future works include validating the proposed method on larger structures and in identifying multiple damages.</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

Leveraging finite element model with transfer learning for data-driven-based structural damage identification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月19日

查看全文
  • Pei Yi Siow,
  • Yuzhe Xia,
  • Zhi Chao Ong,
  • Shin Yee Khoo

摘要

Purpose

Data-driven-based damage identification models rely heavily on the training data. However, collecting real damage data is often impractical, and data scarcity leads to poor extrapolation, reducing accuracy in identifying unseen damage events. A data-driven-model trained with finite element (FE) simulations of various damage events could help address the data scarcity issue. Yet, discrepancies between simulated and real data due to modelling errors may affect the damage identification performance.

Methods

This paper proposes a damage identification approach that integrates a FE model, machine learning (ML) model, and a transfer learning (TL) method to leverage the simulated training data for identifying actual damages in a plate-like structure through minimising the domain discrepancies. For the case study, stiffness-reduced single damage cases with low and high severity levels at plate supports were investigated. Support Vector Machine (SVM) was used as the ML model, with the first mode shape differences used as the damage-sensitive features. A semi-supervised domain adaptation approach was used as the TL method. Simulated data from the FE plate model was used as the source dataset, while limited experimental data was used as the few-shot labelled target data for domain adaptation. The relationship between the average discrepancies in dynamic characteristics and the unseen damage identification accuracy was investigated to study the effectiveness of the proposed method. The performance of Semi-Supervised Transfer Component Analysis (SSTCA) and Semi-supervised Maximum Independence Domain Adaptation (SMIDA) were compared.

Results

Results showed that SMIDA outperforms SSTCA, with the average improvement rates in damage identification accuracy of 50.39% for 1-shot SMIDA and 74.85% for 3-shot SMIDA when the average deviation in dynamic characteristics ranged from 3% to 32% due to varied boundary conditions. The trend where the accuracy reduces when the deviation increases is observed when SMIDA was used, suggesting the effectiveness of SMIDA in minimising the domain discrepancies due to deviation in dynamic characteristics.

Conclusion

The findings showed the potential of TL in leveraging FE models for data-driven-based damage identification through minimising the domain discrepancies due to deviation in dynamic characteristics. Future works include validating the proposed method on larger structures and in identifying multiple damages.