Bond strength of corroded reinforcement in concrete: Neural and tree based approaches

Reinforcement corrosion affects the existing concrete structures, particularly in the coastal regions. One of the effects of corrosion of reinforcement is degradation of the bond stress that can be developed between the reinforcement and the surrounding concrete and this in turn affects the capacity of the reinforced concrete member. Prediction of the bond stress applicable for the corroded reinforcement has been attempted using analytical, empirical and soft computing methods. This article presents the comparative performance of two data-driven tools, artificial neural network (ANN) and decision tree (DT) for the task of prediction of bond stress from the corrosion level, the compressive strength of concrete and the ratio of cover and diameter of reinforcement bar. From the extensive evaluation of performance with both quantitative and graphical methods, it was concluded that the ANN approach would be better suited for the application, with the available data. For development of the models 8-fold cross validation scheme was adopted due to the limitations of data. The ANN models trained with pull-out test data, when employed with ensemble approach in predictive mode for a different experiment setup and bond strength test (flexural) data, could produce results comparable to ANN models trained with flexural test data (reported in literature). The inclusion of the additional factors (compressive strength of concrete and the ratio of cover and diameter of reinforcement bar), 8-fold cross validation approach, and ensemble prediction could be the possible reasons for achieving such portability of pull-out test based model for prediction of flexural test data.