Attention-Guided Adaptive Compressed Sensing for Laser Ultrasonic Signal Reconstruction

Abstract

Laser ultrasonic (LU) testing is a high-resolution, non-contact nondestructive evaluation technique with considerable potential for surface defect inspection in additive manufacturing (AM) components. However, reliable LU imaging typically requires dense spatiotemporal sampling, resulting in excessive data volume and computational burden. Although compressed sensing (CS) provides an effective framework for sampling reduction, conventional random measurement strategies fail to exploit the intrinsic structural characteristics of LU signals. In particular, laser ultrasonic surface wave signals exhibit pronounced non-uniform energy distribution in the time domain, where defect-sensitive information is mainly concentrated in direct arrivals, defect-induced scattering, and interface reflections. Under low sampling rates, random measurements are therefore prone to losing critical information. To address this issue, an attention-guided adaptive CS method is proposed for LU signal reconstruction. By integrating prior energy distribution knowledge with a convolutional attention mechanism, a content-aware measurement matrix is constructed to preferentially sample information-rich signal segments. An end-to-end one-dimensional convolutional neural network is further developed to reconstruct high-quality ultrasonic signals from compressed measurements. Experimental results obtained from LU inspection of TC4 titanium alloy specimens with surface defects demonstrate that, at a compression ratio of 40%, the proposed method consistently outperforms conventional approaches in terms of reconstruction accuracy and imaging reliability, highlighting its effectiveness for efficient LU-based surface defect inspection.