The High-Sparsity Lamb Wavefield Reconstruction Framework Based on Masked Autoencoder

The Scanning Laser Doppler Vibrometer (SLDV) has emerged as a powerful tool for acquiring full wavefield data of propagating Lamb waves in carbon fiber reinforced plastics (CFRP). Conventional full wavefield scanning techniques, however, suffer from being time and energy-intensive, as they require a sufficient number of sampling points to maintain spatial resolution. To tackle this challenge, we proposed a novel autoencoder-based full wavefield reconstruction method. In our approach, the high-sparsity wavefield data were subjected to patch and positional embedding, converting the wavefield data into a sequence containing location information. This sequence was then fed into the encoder consisting of multiple Transformer blocks. Subsequently, the output was combined with the mask tokens representing the unsampled wavefield, which was sent into the decoder composed of multiple Transformer blocks. Furthermore, a fully connected layer projected the output to the original wavefield size and realized the wavefield reconstruction. Our analysis included determining the optimal sampling ratio and sampling strategy. Additionally, we evaluated the trained model on the validation dataset to prove the generalization capability of the proposed network. Finally, we compared the average prediction loss (APL) for damaged and pristine areas, confirming the model’s performance in reconstructing the wavefield on structural discontinuities. Our proposed approach exhibited a solid ability to reconstruct the high-sparsity wavefield accurately, making it a promising solution for more efficient Lamb wavefield data acquisition in CFRP and high-precision damage imaging.