Machine learning to estimate the bond strength of the corroded steel bar‐concrete

Abstract

Steel bar corrosion is prone to serious degeneration of the bond strength (τBS) of reinforced concrete, which leads to structural failure. There are many factors affecting the τBS; thus, the assessment of the τBS of corroded steel bar‐concrete becomes very complicated. In this research, a τBS evolution database of 430 samples containing 14 important factors, such as concrete compressive strength, yield strength and corrosion rate of steel bar, was established. Six machine learning algorithms were employed for comparative study to achieve optimal prediction results. The results showed that random forest and extreme random trees models have the best performance with coefficient of determination (R2) of 0.83 and 0.82, mean square error (MSE) of 13.59 and 13.98, root mean square error of 3.68 and 3.74, mean absolute error values of 2.35 and 2.31, and a20 values of 0.62 and 0.70, respectively. In addition, both models demonstrated shorter training times compared to the others. The running times are 589.75 and 418.31 ms, respectively. The superior feasibility and reliability of the machine learning algorithm were validated by contrasting with the computation of the empirical formulas. The empirical formulas exhibited higher error values, with the majority of MSE indicators exceeding 50 and the lowest R2 value even presenting negative values. Furthermore, feature importance calculations were used to understand the ability of the machine learning models to perceive each influencing factor. Among them, the corrosion rate, the yield strength of the longitudinal bar, and the stirrup emerged as the three most important influential factors.