Drawbead geometric parameters using an improved equivalent model and PSO-BP neural network

Abstract

To improve the simulation accuracy in sheet metal forming, an improved equivalent drawbead restraining force model is proposed considering the offset of the neutral layer and Bauschinger effect. It is validated with Nine's experimental data. BP neural network is optimized by the regularization and pruning theory to decrease neural network redundancy. The weights and threshold values of BP network are optimized by dynamic Particle Swarm Optimization (PSO) algorithm to reduce the possibility of local values. The surrogate mapping model of input–output variables is obtained from the improved PSO-BP model. The fender from NUMISHEET'93 is selected as case study. The main factors are sampled making use of Latin hypercube. The Latin hypercube sampling is optimized by simulated annealing algorithm. Based on the improved PSO-BP neural network, the metamodel between drawbead forces and forming objectives is established. Based on the multi-objective PSO optimization method, the mapping model is optimized to obtain the optimum drawbead restraining forces. The inverse model of drawbead is established based on the proposed drawbead force model to obtain its geometric parameters. The actual drawbeads based on the inversed parameters are simulated to verify the feasibility of the method. The results show that proposed method can significantly improve the forming quality.