Robust statistical feature extraction and deep learning surrogates for highly accurate guided wave-based determination of global corrosion parameters in metallic plates

Abstract

Accurate evaluation of corrosion in metal plates is essential across many industries, particularly in shipbuilding. Among available techniques, non-destructive testing (NDT) is especially valuable, as it enables the assessment of material degradation without compromising structural integrity. Within this domain, guided ultrasonic waves have attracted increasing attention in recent years due to their potential for estimating statistical corrosion parameters, such as mean thickness loss and its standard deviation. However, practical application of these methods remains challenging because of the complex interaction between wave propagation and spatially varying thickness profiles. This study presents an advanced methodology for precise estimation of corrosion parameters using deep learning-based inverse models. The corrosion morphology is modeled as a stochastic field, while guided wave responses, obtained through numerical simulations and recorded by a sensor network, are postprocessed to extract statistical descriptors. In particular, the mean vector and covariance matrix of sensor signals, including wave packet location, width, and amplitude, are computed. To ensure robustness, feature scaling and elliptic envelope-based covariance estimation are employed, effectively mitigating the influence of outliers, which are especially pronounced at higher levels of thickness loss and surface roughness. The inverse model establishes a mapping between extracted feature vectors and the corresponding corrosion parameters, enabling their direct prediction. This framework addresses the intrinsic non-uniqueness of the problem, where identical statistical descriptors may arise from infinitely many realizations of the stochastic corrosion field. Extensive numerical investigations demonstrate that reliable feature extraction, particularly robust handling of outliers, combined with the ability of deep neural networks to capture nonlinear relationships, is critical for accurate parameter estimation. Prediction performance is further improved by aggregating responses from multiple independent realizations of the stochastic corrosion field. Overall, the results demonstrate that the proposed framework, combining robust statistical feature extraction, covariance-based feature representation and nonlinear inverse modeling, provides improved prediction accuracy compared with the previously developed feature-based linear regression model and the conventional time-of-flight approach under the considered numerical simulation conditions.