APPLICATION OF TRANSFER LEARNING OF CAE TO THE TRAINING OF NEURAL NETWORKS OF DIFFERENT INJECTION PRODUCTS

摘要

A neural network has the advantages of high accuracy and fast speed in numerical prediction,and its disadvantage is that a large amount of training data is required for network training.There is a great variety of injection molding products,and the neural networks of various products cannot be shared directly.Therefore,each prediction module needs to take up a lot of time to make the training data,meaning that the prediction module cannot be applied to an actual injection molding mold trial.This research imported the concept of transfer learning to retrain the well-trained neutral network architecture and hyperparameter according to the training data of similar products and to explore the effect of form and structure of training data on the accuracy of transfer learning.This research used 2 models of a circle plate and square plate to transfer the well-trained circle plate model to be used by the square plate.The research results showedthat the Random Shuffle method for data pre-processing can improve the overfitting problem in addition to reducing the error rate of prediction.The prediction of complicated warpage is the most obvious.In the training of the circle plate,the error of gate warpage fell from 29.85% to 19.90%.When the Random Shuffle method is used in combination with the square plate model of transfer learning,the error rate of warpage also fell from 59.61 to 31.05.