A Hybrid Neural Network for Predictive Model in A Plastic Injection Molding Process

Abstract

A reliable and sensitive technique for predicting quality of a plastic work-piece produced in injection molding process is essential help for practicing engineers. A system based on the process parameters that can estimate both two prime characteristics, %volume shrinkage and warpage of work-piece before it produced is significantly beneficial. In this paper, a fast feed forward network, Hybrid Neural Network (HNN), is proposed to construct the predictive model for those two quality characteristics. The unique algorithm of HNN based on the optimization of the weights of each layer is changed to a linear problem by linearization of the sigmoid functions. As iteration procedure used in Backpropagation algorithm is eliminated, the network training time is significant reduced. With this fast convergence of using HNN, the intelligent predictive model for injection molding process that can learn online is possible for further study. To entitle the network to cater for various process parameter conditions, a knowledge base as training and testing data have to be generated on the experimental data in a comprehensive working range of a plastic injection molding process. Consequently, the experiments were performed in 256 conditions based on the combination of nine basic process parameters. The neural networks were trained and the architecture of networks was appropriately selected by benchmarking the Root Mean Square error (RMS). The results of the novel network, HNN, have shown the ability to accurately predict the percentage of volume shrinkage with the 1.02% and 4.87% error at training and testing stages, respectively and for warpage with the 3.76% and 2.47% error at training and testing stages, respectively. These accuracy results are similar to those of backpropagation neural network (BPNN), but HNN has shown the superior fast converging about 38.5% and 66.7% over than those of BPNN.