A deep transfer learning network for structural condition identification with limited real-world training data

Nengxin Bao, Tong Zhang, Ruizhi Huang, Suryakanta Biswal, Jingyong Su, Ying Wang

Abstract

Structural condition identification based on monitoring data is important for automatic civil infrastructure asset management. Nevertheless, the monitoring data is almost always insufficient, because the real-time monitoring data of a structure only reflects a limited number of structural conditions, while the number of possible structural conditions is infinite. With insufficient monitoring data, the identification performance may significantly degrade. This study aims to tackle this challenge by proposing a deep transfer learning (TL) approach for structural condition identification. It effectively integrates physics-based and data-driven methods, by generating various training data based on the calibrated finite element (FE) model, pretraining a deep learning (DL) network, and transferring its embedded knowledge to the real monitoring/testing domain. Its performance is demonstrated in a challenging case, vibration-based condition identification of steel frame structures with bolted connection damage. The results show that even though the training data are from a different domain and with different types of labels, intrinsic physics can be learned through the pretraining process, and the TL results can be clearly improved, with the identification accuracy increasing from 81.8% to 89.1%. The comparative studies show that SHMnet with three convolutional layers stands out as the pretraining DL architecture, with 21.8% and 25.5% higher identification accuracy values over the other two networks, VGGnet-16 and ResNet-18. The findings of this study advance the potential application of the proposed approach towards expert-level condition identification based on limited real-world training data.

Introduction

Structural Health Monitoring (SHM) serves as a critical diagnostic tool for civil infrastructure, yet its effectiveness is often hampered by a "data drought." While deep learning models excel at pattern recognition, they require vast amounts of labeled data to generalize effectively. In real-world civil engineering, obtaining labeled data for every potential failure mode is practically impossible; experiments are costly, and monitoring systems typically only record structures in their healthy states.

When deep learning models are trained on these limited datasets, their performance often collapses when encountering slightly different damage scenarios. This research addresses this bottleneck by proposing a framework that uses physics-based simulations to teach a model the fundamental "rules" of structural behavior before transferring that knowledge to real-world sensor data.

Methodology: The TL-SHMnet Framework

The proposed methodology integrates physics-based modeling with data-driven deep learning through a three-step transfer learning workflow.

The three-step transfer learning workflow for structural condition identification.
  1. Physics-Based Data Generation: A high-fidelity Finite Element (FE) model of a steel frame is constructed and calibrated. This model serves as a synthetic laboratory, generating thousands of vibration responses for various damage scenarios, such as bolt loosening at different joints.
  2. Pretraining (Source Domain): A custom 1D Convolutional Neural Network (CNN), named SHMnet, is trained on this massive synthetic dataset. Unlike standard image-based CNNs, SHMnet is designed to process raw acceleration time-history data directly, learning the intrinsic physics of how structural damage alters vibration signals.
  3. Knowledge Transfer (Target Domain): The pretrained model is transferred to the real-world domain and fine-tuned using a small amount of experimental data. The study explored various strategies, including freezing feature extraction layers and fine-tuning only the classification layers.

Network Architecture and Transfer Strategy

A key innovation of this study is the SHMnet architecture. While modern deep learning often favors extremely deep networks like ResNet-18 or VGG-16, the researchers found that a shallower, custom-built architecture was more effective for 1D vibration data.

Detailed architecture of SHMnet and the knowledge transfer process from pre-training to fine-tuning.

As shown in Figure 2, SHMnet utilizes three convolutional layers for feature extraction. The transfer process involves taking the weights learned from the source domain (synthetic data) and applying them to the target domain (real data). The most successful strategy involved freezing the convolutional layers, suggesting that the "low-level" features—such as wave decay patterns and frequency shifts—learned from simulation are universal enough to apply to real physical structures.

Key Results and Discussion

The performance of the TL-SHMnet framework was validated using a single-bay steel frame laboratory model. The experimental dataset included 10 damage scenarios involving loosened bolts, while the synthetic dataset covered 36 simulated scenarios.

  • Accuracy Improvement: The application of transfer learning increased identification accuracy from 81.8% (training from scratch on limited real data) to 89.1%.
  • Architecture Superiority: SHMnet significantly outperformed deeper industry-standard models. It achieved 21.8% and 25.5% higher accuracy than VGGnet-16 and ResNet-18, respectively. This suggests that for structural vibration data, overly complex models are prone to overfitting, whereas the shallower SHMnet captures essential physical features more robustly.
  • Generalization: The model demonstrated an ability to bridge the "domain gap," successfully identifying damage even when the training labels from the simulation did not perfectly match the real-world experimental labels.

Conclusion

This study demonstrates that deep transfer learning can effectively mitigate the data scarcity problem in structural health monitoring. By leveraging physics-based simulations for pretraining, models can learn intrinsic structural behaviors that translate to real-world applications. Future work will focus on utilizing Generative Adversarial Networks (GANs) to create even more realistic synthetic signals and scaling the framework for complex, large-scale infrastructure like urban bridges.

Summary updated

Abstract

Structural condition identification based on monitoring data is important for automatic civil infrastructure asset management. Nevertheless, the monitoring data is almost always insufficient, because the real-time monitoring data of a structure only reflects a limited number of structural conditions, while the number of possible structural conditions is infinite. With insufficient monitoring data, the identification performance may significantly degrade. This study aims to tackle this challenge by proposing a deep transfer learning (TL) approach for structural condition identification. It effectively integrates physics-based and data-driven methods, by generating various training data based on the calibrated finite element (FE) model, pretraining a deep learning (DL) network, and transferring its embedded knowledge to the real monitoring/testing domain. Its performance is demonstrated in a challenging case, vibration-based condition identification of steel frame structures with bolted connection damage. The results show that even though the training data are from a different domain and with different types of labels, intrinsic physics can be learned through the pretraining process, and the TL results can be clearly improved, with the identification accuracy increasing from 81.8% to 89.1%. The comparative studies show that SHMnet with three convolutional layers stands out as the pretraining DL architecture, with 21.8% and 25.5% higher identification accuracy values over the other two networks, VGGnet-16 and ResNet-18. The findings of this study advance the potential application of the proposed approach towards expert-level condition identification based on limited real-world training data.
BibTeX
@inproceedings{bao2023deep,
  author = {Nengxin Bao and Tong Zhang and Ruizhi Huang and Suryakanta Biswal and Jingyong Su and Ying Wang},
  title = {{A deep transfer learning network for structural condition identification with limited real-world training data}},
  booktitle = {Structural},
  year = {2023},
  doi = {10.1155/2023/8899806},
}
Figure