Guided wavefield-based semi-supervised delamination imaging using convolutional autoencoder

Yitian Yan

doi:10.1016/j.ymssp.2025.113487

Guided wavefield-based semi-supervised delamination imaging using convolutional autoencoder

2025-10-11

0

PRE

AI

DOI：10.1016/j.ymssp.2025.113487

Original

Original request for help

Save

Abstract

En 中文

Guided wavefields encode rich information on wave–damage interactions, making them widely utilized for damage mapping. While deep learning-based supervised approaches have recently been developed for guided wavefield analysis, their practical application is limited by the high cost of data acquisition, potential labeling errors, and poor transferability—models trained for specific regions often require retraining when applied to new inspection areas. Semi-supervised approaches offer a promising alternative, relying solely on signals collected from healthy structures for model training. However, existing semi-supervised frameworks primarily utilize pitch-catch signals, enabling only qualitative detection or damage localization. To date, there’s no semi-supervised framework that leverages guided wavefields to achieve damage imaging. Therefore, this study proposes a novel semi-supervised framework based on Convolutional Autoencoder (CAE) and guided wavefields for delamination imaging. This approach employs a CAE to learn intrinsic patterns of healthy wavefields, enabling detection of wavefield anomalies caused by delamination. Unlike conventional semi-supervised methods that rely on pre-collected historical data, the training data in this framework are directly extracted from a healthy portion of the inspected structure. Furthermore, this framework does not rely on long-term wavefield sequences, but operates solely on a single-frame snapshot captured when the direct waves fully cover the inspection region. To facilitate efficient learning, the wavefield snapshot is segmented into uniformly sized patches. Additionally, symmetric flipping-based augmentation was applied to the training data to further enhance model performance. In this framework, each patch within the inspection area is individually reconstructed using the CAE, and a pixel-wise reconstruction error map is computed, wherein regions influenced by delamination exhibit elevated errors. Finally, this error map is thresholded to identify wavefield anomalies and generate delamination images. The effectiveness of flipping-based augmentation, the influence of patch size and sliding stride used for patch sampling, as well as the model’s noise robustness, are systematically investigated. To validate these findings, the proposed approach was experimentally evaluated using guided wavefield data acquired with a scanning laser Doppler vibrometer.

Guided wavefield-based semi-supervised delamination imaging using convolutional autoencoder

Abstract

Journal

Organization