Torsional capacity evaluation of RC beams using an improved bird swarm algorithm optimised 2D convolutional neural network

Abstract

This study presents the application of deep learning technology in torsional capacity evaluation of reinforced concrete (RC) beams. A data-driven model based on 2D convolutional neural network (CNN) is established, where model inputs contain the beam width, beam height, stirrup width, stirrup height, concrete compressive strength, steel ratio of longitudinal reinforcement, yield strength of longitudinal reinforcement, steel ratio of transverse reinforcement, yield strength of transverse reinforcement and stirrup spacing. To enhance the prediction accuracy of the proposed model, an improved bird swarm algorithm (IBSA) is leveraged to optimise the hyperparameters of CNN in the training phase. A comprehensive dataset, comprising 268 groups of laboratory tests of RC beams collected from published articles, is used for model development and validation. The results show that the proposed 2D CNN with hyperparameter optimisation exhibits high performance in predicting torsional strength of RC beams, which outperforms other machine learning models, building codes and empirical formula in terms of a series of evaluation metrics.