Ultrasonic Signal Denoising Based on One-Dimensional Convolution and SWT-MRA Multi-Scale Representation

Ultrasonic measurement signals are susceptible to random noise, impulsive artifacts, and high-frequency interference, while conventional wavelet thresholding may attenuate useful signal components. To address the trade-off between noise suppression and waveform preservation, this study proposes an ultrasonic signal denoising method that integrates stationary wavelet transform-based multiresolution analysis (SWT-MRA) with a one-dimensional convolutional neural network (CNN). An eight-level SWT-MRA decomposition using the sym8 wavelet produces one approximation component and eight equal-length detail components. A dilated residual CNN then learns a component-domain denoising mapping under joint waveform-, frequency-, gradient-, and component-level constraints. The method was evaluated on three independent experimental batches, and the learning-based models were repeatedly trained using five random seeds. On the independent test set, the proposed method achieved a mean absolute error (MAE) of 0.003685, a root mean square error (RMSE) of 0.005024, and an output signal-to-noise ratio (SNR) of 22.198 dB. Compared with the time-domain CNN, it reduced MAE and RMSE by 3.10% and 3.39%, respectively, while increasing output SNR by 0.297 dB. Compared with 1D U-Net, MAE and RMSE were reduced by 9.22% and 10.19%, respectively. Ablation experiments showed that the main performance gain was attributable to the explicit SWT-MRA multiscale representation, whereas component-level auxiliary supervision provided limited additional benefit. These results demonstrate that explicit multiscale representation can improve ultrasonic waveform recovery with only a modest increase in network size.