Prediction of the Postfire Flexural Capacity of RC Beam Using GA-BPNN Machine Learning

Abstract

To accurately predict the flexural capacity of postfire RC beams is imperative for fire safety design. In this paper, the residual flexural capacity of postfire RC beams is predicted based on a back-propagation (BP) neural network (NN) optimized by a genetic algorithm (GA). First, the temperature distribution of the beams was determined using the finite-element analysis software ABAQUS version 6.14-4, and the strength reduction factor of materials was determined. The flexural capacity of the RC beams after fire was calculated by the flexural strength reduction calculation model. The model was used to generate the training data for the NN. To enable machine learning, 480 data sets were produced, of which 360 were used to train the network; the remaining 120 were used to test the network. The predictive models were constructed using BPNN and GA-BPNN. The prediction accuracy was evaluated by comparing the predicted and target values. The comparison showed that the GA-BPNN has a faster convergence speed and higher stability and can reach the goal more times, reducing the possibility of BPNN falling into the local optimum and achieving the global optimum. The proposed GA-BPNN model for predicting the flexural capacity of postfire RC beams provides a new approach for design practice.