Experimentation and a Data-Driven Machine Learning Approach for Strength Prediction of Non-Uniformly Corroded Rebar

Abstract Accurate prediction of the residual load-carrying capacity of corroded reinforced concrete (RC)structures is essential for assessing their service life and durability, particularly in aggressive environments. To address this, a series of accelerated electrolytic corrosion experiments, both with and without additional sustained loading, were conducted to investigate the degradation of corroded rebars. Results revealed that significant non-uniformity corrosion across cross-sections and along rebar length, accompanied by a decline in concrete compressive strength as corrosion increased. Monotonic tensile testing further demonstrated a marked reduction in elongation, yield and ultimate strengths, emphasizing the detrimental effects of corrosion on structural integrity. To expand the analysis, a dataset of 1,423 samples was utilized in four deep learning methods to assess the effects corrosion on corroded rebar strength. Feature importance analysis was conducted using the Shapley Additive Explanation (SHAP) approach, quantified each input parameter’s contribution to strength loss. Based on this analysis, a simplified predictive formula was developed to quantify strength loss based on two key parameters (i.e. corrosion ratio and non-uniform coefficient of diameter). This newly user-friendly expression serves as an effective tool for forecasting rebar strength degradation, providing valuable insights for the maintenance and assessment of aging infrastructure in harsh environmental conditions.