Data-driven representation learning and stability validation for multi-quality prediction in smart injection molding

This study proposes a data-driven artificial intelligence framework for multi-quality prediction in injection molding by integrating unsupervised representation learning and supervised regression modeling. In-mold pressure signals collected from five sensing locations were encoded using autoencoders to extract compact latent representations that capture local flow dynamics. To evaluate the reproducibility of the learned representations, multiple independent autoencoder training trials were conducted under identical input data and hyperparameter settings, and cosine similarity was employed as a quantitative indicator of representation consistency. The results showed stable similarity values predominantly in the range of 0.6–0.8, indicating consistent latent space structures despite the stochastic nature of deep learning optimization. The extracted latent features were subsequently used as inputs to a multilayer perceptron (MLP) model for simultaneous prediction of six key quality indicators, including three widths, two lengths, and part weight. Prediction results demonstrated consistent accuracy and low variability across different sensing locations and independent encoder instances, confirming that stable representations can support reliable downstream multi-quality prediction without retraining under identical process conditions. Rather than introducing a new prediction architecture, this work emphasizes the validation of representation stability and its relationship to predictive reliability within a controlled injection molding process window. The findings demonstrate the potential of stable, reusable latent representations for data-driven quality prediction and provide a systematic foundation for future extensions toward noise-aware validation, cross-condition generalization, and industrial deployment in smart manufacturing environments. Hybrid AI integrates unsupervised and supervised learning. Autoencoders extract stable pressure features. Latent features show high reproducibility. AE-MLP provides consistent multi-quality prediction. Features demonstrate reusability under controlled process conditions.