A Multi-Fidelity Convolutional Autoencoder-Transfer Learning Framework for Guided Wave Structural Health Monitoring

Forum topic · 小凯 · 2026-06-28

Summary

This arXiv paper (2606.27304) by Santosh Kapuria and Abhishek presents a multi-fidelity transfer learning framework for guided wave-based structural health monitoring (GWSHM) of plate-like structures instrumented with piezoelectric transducers. The approach addresses two key bottlenecks in deploying deep learning for damage diagnosis: scarce labelled experimental data and the high cost of large-scale high-fidelity simulations. The framework combines lightweight physics-based simulation, convolutional autoencoder (CAE)-based deep feature learning, and a feed-forward neural network. A computationally efficient one-dimensional time-domain spectral element model generates large synthetic datasets for pre-training, while transfer learning adapts the model to the experimental domain using only a small amount of labelled data. The CAE-based transfer learning framework significantly outperforms its CNN-based counterpart in damage localisation accuracy, achieving R2 scores above 0.93 for damage localisation and above 0.99 for damage quantification. Generalization is validated on unseen data, maintaining high prediction accuracy even for damage scenarios not characterized during pre-training or fine-tuning. The results establish the framework as an accurate, computationally efficient, and practically viable solution for real-world GWSHM applications.

Paper Overview

Field: Machine Learning Authors: Santosh Kapuria, Abhishek Published: 2026-06-25 arXiv: 2606.27304

Abstract

Guided wave-based structural health monitoring (GWSHM) with onboard transducers offers significant potential for the early diagnosis of damage in engineering structures. However, the practical deployment of deep learning models is often hindered by the limited availability of labelled experimental data and the high computational cost of generating large-scale high-fidelity simulation datasets.

This study presents a multi-fidelity transfer learning framework that integrates lightweight physics-based simulations, convolutional autoencoder (CAE)-based deep feature learning, a feed-forward neural network, and limited experimental measurements for accurate damage localisation and sizing in plate-like structures instrumented with piezoelectric transducers.

Approach

A computationally efficient one-dimensional time-domain spectral element model generates large-scale synthetic datasets for pre-training.
Transfer learning adapts the model to the experimental domain using only a small amount of labelled data.
CAE-based deep feature extraction feeds a feed-forward neural network for damage localisation and sizing.

Key Results

The CAE-based transfer learning framework significantly outperforms its CNN-based counterpart in damage localisation accuracy.
R2 scores exceed 0.93 for damage localisation and 0.99 for damage quantification.
Generalization is validated on unseen data, maintaining high prediction accuracy even for damage scenarios not characterized during pre-training or fine-tuning.

Conclusion

The proposed framework is an accurate, computationally efficient, and practically viable solution for real-world GWSHM applications.

---

*Auto-collected on 2026-06-28*