An ATPSO-BP neural network modeling and its application in mechanical property prediction

Abstract

In order to improve the generalization ability and convergence rate of back propagation (BP) neural network and prevent it from falling into local optimum, an ATPSO-BP model, which is a hybrid of active target particle swarm optimization (ATPSO) and BP neural network, was proposed. In this model, the active target point and the inertia factor were introduced on the basis of the traditional particle swarm algorithm to obtain new particle velocity and position update, and then ATPSO algorithm was applied to optimize the weights and thresholds of BP model. ATPSO-BP and BP models were both employed to predict the stress of 6181H18 aluminum alloy, and the comparison and the analysis were carried out. The numerical results show that ATPSO algorithm enhances the optimization ability of the weights and thresholds of BP model. It is verified by the average relative error and the standard residual that ATPSO-BP model compared to BP model has higher prediction accuracy, and objectively reflects the influence law of thermodynamic parameters on the stress, which can provide significant guidelines in aspect of stress prediction methodology of aluminum alloys for its development.