Reliable identification of global corrosion parameters in metallic plates through Data-Driven inverse Modelling of...

Assessing corrosion in metallic plates using guided waves remains challenging due to the complex interactions between wave propagation and spatially varying thickness distributions. While conventional approaches typically assume uniform thickness reduction, real corrosion processes lead to stochastic, spatially heterogeneous degradation patterns. This introduces significant uncertainty in the interpretation of measured signals and limits the applicability of classical physics-based identification methods. In this study, a behavioral inverse modeling framework is proposed for the identification of global corrosion parameters, namely the mean thickness reduction and the standard deviation of thickness, in plates with randomly varying corrosion geometries. The corrosion morphology is represented as a stochastic field, and guided wave responses are generated through numerical simulations. Signal processing is performed to extract descriptive features, which are subsequently used as inputs to an inverse regression model that maps signal characteristics to corrosion parameters. The performance of the proposed approach is compared with that of a classical time-of-flight-based identification method derived from dispersion curve analysis. The results indicate that the time-of-flight approach provides mediocre estimates of the mean thickness. Furthermore, this approach relies on assumptions derived for plates of uniform thickness, in which the wave velocity can be related to thickness through dispersion characteristics of guided waves. In the presence of spatially varying thickness, this relationship becomes more complex, and using an effective average velocity alone leads to inaccuracies. The proposed inverse model addresses these limitations by learning the mapping directly from signal characteristics and their extracted multiple features, enabling accurate and robust identification of both the mean thickness and its variability across broad ranges thereof. Its efficacy is corroborated through extensive verification studies and supported by rigorous metrics such as relative prediction error and prediction repeatability for stochastically varying corrosion morphologies. The effects of feature selection on the model's predictive power are also investigated.