Explainable Temporal Attribution for Time-Series-Based Quality Prediction in Injection Molding

Abstract

Time-series sensor signals collected during injection molding contain rich information about process dynamics and final part quality. However, quality prediction models based on high-dimensional temporal signals are often difficult to in terpret, and direct attribution in the original signal domain may produce unstable explanations. This paper proposes an explainable temporal attribution framework for time-series based quality prediction in injection molding. The framework first learns compact latent representations of multivariate pro cess signals using a one-dimensional convolutional autoen coder, and then trains a latent-space prediction model for defect classification or dimensional quality regression. To re cover process-level interpretability, SHapley Additive exPlana tions (SHAP) are computed in the latent space and propagated back to the original temporal signal domain through the de coder Jacobian. The proposed framework is evaluated on two injection molding datasets. The results show that the learned latent representations preserve dominant temporal process characteristics and provide useful information for downstream quality prediction. The temporal attribution results reveal that pressure-related signals and the filling-to-packing transition are especially important for quality prediction, while phase aware analysis indicates that packing and cooling stages play dominant roles in sink mark prediction. Robustness analysis further demonstrates that the decoder-based temporal attribu tion framework provides more stable and consistent explana tions than direct SHAP analysis in the original signal domain. The proposed method offers an interpretable way to connect latent predictive features with physically meaningful mold ing process dynamics, supporting process understanding and quality-related decision making.