Deep learning enhanced quantitative debonding evaluation in tile panels using Lamb waves

This study introduces a novel approach for quantitatively evaluating debonding in exterior ceramic tile panels using Lamb wave signals and a data-driven deep learning methodology. Leveraging a two-dimensional convolutional neural network (2D-CNN), time-frequency representations of Lamb wave signals are utilised to predict debonding characteristics in ceramic tiles. The Lamb wave signals are transformed into time-frequency images using a continuous wavelet transform (CWT), which are subsequently processed by a multi-branch 2D-CNN to extract debonding information along sensing paths. The proposed methodology combines simulation data and experimental validation to create a comprehensive 2D damage index map that identifies the position and size of debonding defects. Results demonstrate the effectiveness of the model, particularly when trained with a hybrid dataset with experimental data, indicating its potential for reliable non-destructive evaluation of tiled structures.