Research On The Revers Optimization Of Drawbeads Parameters Based On Improved BP Neural Network

The forming qualities of vehicle panels are relevant with the impact factors like the mould structure、the production process、the material properities, et al. The vehicle panels have the characteristics of large outline dimensions、complicated curved shape and high precision, in which producted by the skilled workers. The traditional production technology will bring a great cost. During the stamping processs, the key and the difficult subjects are how to adjust the various parameters effectively.During the production process of complex vehicle panels, the drawbeads are always placed to restrain the material metals flow which to prevent the huge differential flow deformation. Thus, the restraining forces can be adjusted by dominating drawbeads distributions to remove the defects like wrinkles、fractures and buckles. The area where metals flow frequently needs to be set drawbeads which providing smaller resistance to balance the discrepancies of flowing velocity. During the production process in one type of new product, the moulds are always satisfied with the actual production in the trial-manufacture. During the process of trial-manufacture, the moulds are always scrapped, in which led to a higher cost. The numerical simulation techniques are used to produce the moulds that reducing the debug cycle significiantly. Based on the above methods, this paper is aimed to combine the finite element method with optimization reverse to obtain the geometrical parameters of drawbeads in the trial-manufacture.Firstly, an improved equivalent drawbeads restraining force model is proposed. The equivalent model is introduced with the offset of neutral layer and bauschinger effect. The equivalent model considering multy-factors effects improves the predicted accuracy significantly which validated with Nine experimental data. The geometrical parameters of equivalent drawbeads model are analyzed by grey relational analysis method. The output variables are the maximum incrassation value and the maximum reduction value. The improved BP neural network and PSO algorithm are used to build the mapping model. The nonlinearity mapping relationships between the input values and the output values are optimized by PSO algorithm to obtain the optimal drawbeads forces. Based on the optimal drawbeads forces, the geometrical parameters are reversed. The parameters are used to build a entity drawbeads model simulated by DYNAFORM. The forming limit diagram obtained from DYNAFORM is used to verify the feasibity of the method. The BP method optimized by PSO algorithm is improved by the following improvements:introduce the regularization coefficient and the pruning theory; amend the hidden nodes’redundancy; optimize the weights and threshold values of BP neural network by PSO algorithm.This fender obtained from Numisheet’93is used as an example in the paper. The impact factors of fender are analyzed to obtain a high quality part which used grey relational analysis. The simulated annealing algorithm is used to optimize the latin hypercube sampling. The sampling points are chosen from five drawbeads forces. After that, the simulation software DYNAFORM is operated for sheet metal forming simulation. The output values and input values obtained from DYNAFORM are generated to establish the mapping model. The improved PSO algorithm is used to optimize the mapping model established by BP neural network. Based on crowding distance multi-objective PSO optimization method, the optimal non-inferior solutions are obtained. The non-inferior solutions are the optimizal drawbeads forces. The drawbeads geometric parameters are obtained from reversing the drawbeads equation. The entity drawbeads based on the equivalent drawbeads model are generated in die mould and the binder. DYNAFORM is generated to simulate the fender forming process which used the entity drawbeads. The simulation forming graphs show that the drawbeads geometric parameters set reasonably.