Induced network-based transfer learning in injection molding for process modelling and optimization with artificial neural networks

Abstract

The necessity of an abundance of training data commonly hinders the broad use of machine learning in the plastics processing industry. Induced network-based transfer learning is used to reduce the necessary amount of injection molding process data for the training of an artificial neural network in order to conduct a data-driven machine parameter optimization for injection molding processes. As base learners, source models for the injection molding process of 59 different parts are fitted to process data. A different process for another part is chosen as the target process on which transfer learning is applied. The models learn the relationship between 6 machine setting parameters and the part weight as quality parameter. The considered machine parameters are the injection flow rate, holding pressure time, holding pressure, cooling time, melt temperature, and cavity wall temperature. For the right source domain, only 4 sample points of the new process need to be generated to train a model of the injection molding process with a degree of determination R2 of 0.9 or and higher. Significant differences in the transferability of the source models can be seen between different part geometries: The source models of injection molding processes for similar parts to the part of the target process achieve the best results. The transfer learning technique has the potential to raise the relevance of AI methods for process optimization in the plastics processing industry significantly.

Application of Machine Learning Techniques in Injection Molding Quality Prediction: Implications on Sustainable Manufacturing Industry

“…The authors found that the deep reinforcement learning method outperformed the dispatching rules that are popularly used for minimizing the total weighted tardiness. Another recent study is transfer learning between different injection molding processes to reduce the amount of data needed for model training [17]. The authors used different approaches to ANN models; 16 training samples provided an average R2 value of 0.88 in this paper.…”

Novel Analysis Methodology of Cavity Pressure Profiles in Injection-Molding Processes Using Interpretation of Machine Learning Model

“…To build a neural network model, a multi-layer perceptron (MLP) with a single hidden layer was selected. Lockner and Hopmann suggested that one hidden layer having a single digit number of neurons would be suitable for high model quality for injection molding [ 38 ]. The time and pressure values of the PSPs were inputs to the input layer of the MLP, and the part quality was predicted from the output layer, as shown in Figure 7 .…”

Neural network feature and architecture optimization for injection molding surface defect prediction of model polypropylene

“…However, it is vital to understand that optimization methods that depend exclusively on CAE simulation data are not 100% accurate as they do not fully capture all the physics involved in polymer processing. For that reason, research has been conducted in predicting occurrences within injection molding by developing predictive frameworks [ 18, 19 ] and including real experimental data as training data. For example, Saad Mukras' optimization framework, based on the Kriging Model, predicted cycle time, warpage, and volumetric shrinkage with an error of 6.7%, 3.2%, and 8%, respectively, by analyzing samples from real injection molding trials.…”