Optimizing delamination imaging via full wavefield segmentation using augmented simulated wavefield data

To avoid catastrophic structural failures resulting from the hidden accumulation of delamination in carbon-fiber reinforced polymer, timely and accurate detection is crucial. Several deep learning-based methods have been developed for full wavefield segmentation to image delamination. However, the cost of experimentally acquiring extensive damage scenarios for dataset construction is prohibitively high. In addition, accurately segmenting delamination remains challenging due to the complex superposition of guided wave components in the full wavefield. This paper proposes a data preprocessing strategy that combines wavenumber filtering with a hybrid noise-flipping augmentation, to enhance the performance of deep learning models in full wavefield segmentation for delamination imaging. By isolating the derived guided wave modes introduced by delamination in the frequency domain, the deep learning models are guided to concentrate more effectively on delamination-relevant features. Noise augmentation and flipping augmentation are employed to improve the generalization of the delamination imaging models, enabling them to better handle real-world measurement conditions, which may include various external interference factors such as structural vibration and transducer noise. Seven distinct deep learning models were employed and evaluated in both simulated and experimental settings to examine the effectiveness of the data preprocessing strategy. The results demonstrate that models trained solely on simulated data can be effectively applied to experimental measurements, achieving a highest intersection over union score of 0.8634 and producing artifact-free delamination imaging.