Ultrasonic defect characterization using Bayesian inversion and scattering matrix denoising neural networks

Abstract

The use of scattering matrix and deep neural networks for ultrasonic characterization of inclined crack-like defects in noisy environments was explored. A distortion model was utilized to simulate coherent noises that could contaminate the experimental measurements in practice. Given a test scattering matrix, we first developed an approach for estimating the parameters of the distortion model based on the structural similarity index. Subsequently, a Bayesian inversion approach was adopted to determine the proportion of positive-angled cracks that should be included in the scattering matrix database. Based on this result, a deep neural network model was constructed and used in the denoising procedure, which can effectively reduce the characterization error induced by measurement noise. The simulation showed that the proposed approach can be reliably used for characterization of crack-like defects with large orientation angles relative to the array direction (e.g., 75∘) and small sizes (e.g., 0.8λ). In experiments, six crack-like defects with orientation angles of 60∘ and 75∘ were characterized with errors within 0.1λ (i.e., 0.25 mm) and 5∘ in size and angle, respectively. In addition, the characterization uncertainty measured by the root-mean-squared error was reduced by 44.7% in size compared with the conventional Bayesian approach.