Prediction of limiting dome height using neural network and finite element method

Abstract

Using neural network to predict limiting dome height (LDH) based on the result of finite element analysis is a high efficiency work in spite of little error. Finite element results are presented with different working condition parameters, such as material thickness, punch speed, friction coefficient between punch, die and sheet metal, and blank holder force. Then, a two-layer back propagation network is developed to best fit this discrete engineering problem. Different number of neurons in the hidden layer, three commonly used training algorithms, and two performance functions are adopted and compared to choose the suitable one to minimize the error between the predictive value and the simulation results (one with ideal output). After comparison, the neuron number in the hidden layer is determined to be 12 and the appropriate learning algorithm is Levenberg–Marquardt algorithm. The difference between two performance algorithms is small. The mean square error between the predicted value and targeted one is less than 0.2%. Finally, five sheet metal forming processes under various working conditions are predicted and compared with the finite element method (FEM) result to verify the validity of this neural network model. The small difference indicates that this neural network can predict the LDH in a certain range of working conditions.