Artificial neural network model for steel-concrete bond prediction
Résumé
In this paper, an Artificial Neural Network (ANN) is proposed for modelling the bond between conventional ribbed steel bars and concrete. The purpose is to predict the ultimate pull-out load from the concrete mix constituents (first ANN model) or the compressive strength (second ANN model) and from the steel bar diameter according to the RILEM test configuration [RILEM. Essai portant sur l’adhérence des armatures du béton: essai par traction. Materials and Structures 1970; 3 (3) 175–78]. The ANN models were implemented using an experimental database of 112 pull-out test results performed with ribbed bars 10 mm or 12 mm in diameter and three concrete mixes with different constituent proportions. A Multi-Layer-Perceptron was trained according to a back-propagation algorithm. The first model has six inputs (ANN-6): the diameter of the ribbed bar, the water to cement ratio, the gravel to sand ratio, the crushed to rolled gravel ratio, the type of cement and the concrete maturity. The second model has two inputs (ANN-2): the diameter of the bar and the concrete compressive strength. The ultimate pull-out load was the output data for both models. The results show that the implemented models have good prediction and generalisation capacity with low errors. The ANN-6 model is more accurate, regarding the generalisation capacity, than the ANN-2 model. Concrete mix constituents as input parameter, instead of the compressive strength, are more representative of the local phenomenon at the steel-ribs-to-concrete interface.