A Machine Learning Method for Predicting Part Weight, Dimensions, and Residual Stress during Injection Molding

Injection molding is one of the main processes of polymer processing, which has the advantages of high efficiency and low manufacturing cost. Due to the high cost of time, labor, and equipment required for quality inspection in mass production, batch inspection is often used instead of full inspection, often resulting in difficult quality control. To achieve the goal of quality assurance, this study proposes a virtual measurement technique based on a real-time multi-quality prediction neural network combined with an autoencoder network (AE) and multilayer perceptron network (MLP). The main research content is that through sensing, quality indexing, and automated feature extraction technology, the captured data can be extracted, and the dimensionality reduction of the data is beneficial to the training of the MLP model. Experimental case studies show that the method can in-time predict the residual stress distribution, weight, and geometric dimensions of plastic parts, and the model prediction error (root mean squared error) is less than 5% of the total tolerance. In particular, the required prediction time is less than 0.24 s. The performance of the predicted residual stress distribution is highly similar to the actual picture. Further, the feature codes extracted from the AE model can be used to verify the residual stress quality of the molded part.