Development of Evolutionary Method for Optimizing a Roll Forming Process of Aluminum Parts

Abstract

This paper presents the development of the knowledge-based neural network (KBNN) and genetic algorithm (GA) in modeling and optimization of the roll forming (RF) process of aluminum parts. The idea of a KBNN using multifidelity finite element (FE) models was developed to model the mechanical behaviors of the aluminum sheet. Initially, the less costly but less accurate FE model was used to build the response surface functions for the knowledge path of the KBNN. After that, a small number of the more accurate but expensive finite element analysis (FEA) of the high fidelity FE model were utilized in a multilayer perceptron (MLP) neural network with the prior knowledge to produce the KBNN prediction results. Two powerful optimization algorithms, the Levenberg–Marquadrt (LM) and GA, were applied to train the KBNN. The trained KBNN was used to perform the parametric study for investigating the effects of process parameters on the part quality. After that, the optimization of the process parameters was carried out by employing the combination of the GA and KBNN. The optimization objective was minimizing the overall damage in the aluminum part while keeping the longitudinal strain and spring back angle less than allowable limits to prevent the existence of defects. The modeling and optimization results by using the KBNN and GA were compared with the results from other methods to prove the advantages of the developed one against others.