Corrosion of reinforcing steel in concrete structures, especially in chloride-rich environments, remains a leading cause of structural degradation and presents significant challenges for maintenance. Traditional steel corrosion inspection techniques, both direct and indirect, often fall short in terms of accuracy, efficiency, and cost-effectiveness. In this paper, an interpretable machine learning (ML)-aided framework is developed for predicting steel corrosion degree in concrete, which integrates multi-scale feature selection (MSFS), optimal algorithm determination, and SHapley Additive exPlanations (SHAP) techniques, addressing key limitations of conventional ML models, such as limited feature selection, poor generalization, and black-box opacity. The learning capability of the computational model is verified through extensive comparisons with multiple baseline algorithms. A digital example is presented to demonstrate the accuracy and efficiency of the developed framework. From the example, it is found that the MSFS method can identify key features of steel corrosion, such as crack width (w), geometric ratios (cb/d, cl/d, cb/cl), and concrete properties (fc, W/C). This ensures an optimal balance between accuracy and generalization, as validated by 5-fold cross-validation and independent dataset testing, with the optimal model achieving a test set R2 of 0.94 and a 33.6% reduction in RMSE compared to the default model. It is also found that the SHAP technique can further vindicate w and cb/d as the most influential factors governing internal corrosion. This paper pioneers the ML computational models for the prediction of structural deterioration, e.g., steel corrosion in concrete, which can replace traditional mathematical model-based prediction. These innovations represent a significant step toward the future of digital-driven structural performance prediction.
Gui et al. (Fri,) studied this question.