• Synthetic pipeline generates realistic clean–noisy signals for guided wave testing. • Wavelet-informed neural network separates echoes from noise with attention design. • Validated on real-world applications, surpassing classical and learning approaches. • Enables fast synthetic data generation and real-time model deployment. • Provides a scalable denoising solution for reliable non-destructive evaluation. Guided wave ultrasonic testing (GWUT) in industrial environments is often limited by low signal-to-noise ratio (SNR), which reduces defect detectability. This study proposes a knowledge-guided framework that combines synthetic data generation with a tailored denoising network. From a single reference acquisition, paired clean and noisy signals are constructed using dual-Gaussian echo modeling and composite noise synthesis based on measured spectra. A Wavelet-Initialized Attention U-Net is developed with wavelet-informed kernels, a dual-decoder structure, and an attention bottleneck for efficient temporal integration. Experiments on two representative GWUT systems, a railway switch rail monitoring setup and a storage tank wall inspection robot, show that the proposed framework achieves up to 29.7 dB ROI-based SNR improvement on synthetic data, and substantial CNR improvement on real signals accompanied by a marked reduction of false detections (FP/FN), outperforming classical and deep learning baselines. The method also achieves real-time inference and efficient data generation with moderate computational cost. These results indicate that physics-guided synthesis combined with a tailored network provides a practical solution for GWUT denoising and supports reliable defect detection in industrial applications.
Qian et al. (Sun,) studied this question.