Adaptive sampling for efficient Lamb wavefield reconstruction in composite laminates with Spatial-Temporal Masked AutoEncoder.

The increasing demand for high-accuracy damage quantification in carbon fiber reinforced plastics (CFRP) has led to the widespread adoption of ultrasonic Lamb wave testing (ULWT) for non-destructive testing (NDT) in various engineering applications. The non-contact Scanning Laser Doppler Vibrometer (SLDV) has emerged as a valuable tool for damage evaluation. However, despite significant research on Lamb wavefield analysis methods, the rapid and reliable acquisition of full wavefield data remains a substantial challenge, limiting SLDV's applicability in real-world engineering scenarios. This study presents a novel deep learning-based approach to reconstructing full wavefield data from highly under-sampled wavefield data using the Spatial-Temporal Masked AutoEncoder (STMAE). By leveraging time-series high-sparsity Lamb wavefield data, our method achieves remarkable reconstruction performance with a sampling ratio as low as 5%. Furthermore, we propose a novel scanning path optimization method based on Bayesian Optimization, which generates adaptive sparse spatial sampling patterns for wavefield reconstruction. The integration of this adaptive sampling pattern with the STMAE, termed as AdaSTMAE, yields lower precision wavefield prediction error around the damage areas. A comprehensive parametric study on the sampling ratio was conducted and validated through comparative experiments in both single-damage, multi-damage scenarios. The implementation of the adaptive sampling strategy resulted in a 2-16% reduction in reconstruction error for single-damage scenarios and a 0.7-5% reduction for multi-damage scenarios around damage areas, compared to scenarios without the adaptive strategy. Our experimental results demonstrate the outstanding performance of the proposed technique in wavefield reconstruction, achieving an 87-88% reduction in reconstruction error relative to the original masked autoencoder (MAE) across different sampling ratios (5%-25%). Additionally, cross-structural validation using composite blades with variable thickness confirmed the model's strong generalization capability, effectively reconstructing wavefront distortion and velocity variation without fine-tuning.