Combined improved U-Net network and electromagnetic ultrasonic circular array imaging methods for crack characterization

Electromagnetic acoustic transducer (EMAT) is widely used in defect characterization but face challenges such as low energy coupling efficiency and limited space, leading to low-resolution defect imaging. To overcome this limitation, an enhanced U-Net deep convolutional neural network denoising method, combined with an improved imaging method based on virtual element signal reproduction (VESR) for crack characterization, is proposed. This approach learns from two-dimensional time–frequency domain segments for denoising and utilizes crack scattering information along with VESR method for imaging, circumventing the limitations of the 2 λ criterion. The experimental results indicate that the minimum crack localization error is 1.12 mm, while the minimum angular error is 1.33°. These results highlight the enhanced clarity and reliability of the proposed method for crack characterization, even in noisy conditions. The proposed method successfully characterizes crack with sub-λ dimension. Furthermore, the denoising method provides an effective solution for various acoustic frequency band denoising tasks.