Transfer learning to predict part quality for injection molding with recycled materials

Polymeric materials are inexpensive, lightweight, and easy to process. With injection molding technology, they can be used to mass-produce complex geometric products. Although virtual quality measurement systems can prevent product defects, they may not be effective for recycled polymers. Recycling is necessary to achieve sustainable production, but the recycling process can greatly affect material properties. Accordingly, we propose using transfer learning (TL) to apply quality models for virgin materials to recycled materials. Key features from cavity pressure sensor information were extracted as quality indices, and a neural network was trained to predict the part weight and geometric dimensions. TL was then used to fine-tune this pretrained model on data for recycled materials. This method requires little data collection and could greatly reduce the cost of improving quality when using recycled materials. For all quality metrics, the TL model achieved superior performance to a conventional model, even when applied to a 40% smaller dataset; it also converged more quickly and in fewer iterations. A comprehensive analysis indicated that the model had a favorable fit. These results confirm that TL is effective in predicting the quality of injection molding products made with recycled materials; it has superior predictive accuracy and requires less training time than conventional methods. The described process could reduce the cost of implementing process quality inspection systems, incentivizing the use of recycled materials.