Impact damage localization in woven CFRP laminates using laser-ultrasonic wavefields and a masked spatiotemporal reconstruction network

Barely visible impact damage (BVID) can compromise the structural integrity and service reliability of composite structures. Laser-ultrasonic full-wavefield analysis provides an effective nondestructive evaluation approach for such damage, as it can capture the spatiotemporal evolution of guided-wave propagation. However, existing data-driven wavefield analysis methods generally rely on costly damage annotations and are often limited to two-dimensional spatial image analysis, thereby insufficiently exploiting the spatiotemporal information contained in laser-ultrasonic wavefields. To address these limitations, this study proposes a self-supervised three-dimensional spatiotemporal anomaly localization method for impact damage detection in woven carbon fiber reinforced polymer (CFRP) laminates. The proposed method uses only undamaged wavefield data during training. Random spatiotemporal masking is applied to local wavefield clips, and a masked 3D reconstruction network is trained to learn the normal spatiotemporal propagation patterns of guided waves. During testing, anomaly heatmaps are generated through multiple random masked reconstructions of the inspected wavefield and spatiotemporal aggregation of the resulting reconstruction residuals, thereby characterizing local scattering that deviates from normal propagation patterns and enabling effective damage localization. Experimental results on woven CFRP specimens with single and multiple impact damage sites demonstrate that the proposed method can generate concentrated damage-related anomaly responses for most damage sites. Compared to conventional energy imaging techniques, the proposed method presents more distinct damage-related hotspots and does not produce obvious concentrated damage-like responses on the undamaged specimen. These results suggest that self-supervised spatiotemporal reconstruction based on normal wavefield learning provides a feasible framework for label-free laser-ultrasonic localization of impact damage in composite laminates.