Abstract
<title>Abstract</title> <p>In complex advanced manufacturing processes, microstructural dynamic physical parameters (e.g., viscoelastic modulus and plasticizer migration rates) of porous fiber polymer matrix are extremely difficult to quantify online, creating a critical blind zone for real-time quality control. To address this, this study proposes a physics-informed data-driven modeling framework that translates macro-level rheological priors into statistical constraints for small-sample, high-noise time-series prediction. Utilizing a 14-day dynamic tracking dataset of fiber filter rods, a Gaussian Process Regression (GPR) framework integrating the Matern 2.5 kernel and the white noise kernel is constructed. Unlike traditional deterministic models, this composite-kernel architecture provides robust uncertainty quantification by filtering industrial background noise and delivering rigorous 95% confidence intervals for extrapolations within the 48-hour to 168-hour data void. Leave-one-out cross-validation demonstrates high predictive accuracy (RMSE = 4.15 Pa, R² = 0.9439). Crucially, from the perspective of process systems engineering, this framework bridges the gap between mathematical models and practical factory scheduling. It algorithmically pinpointed the operational degradation inflection point and delineated the optimal production handling window as 2 to 61.1 hours post-molding. These findings provide an intelligent decision-support tool for factory Manufacturing Execution Systems (MES) to optimize warehouse turnover cycles, eliminate quality consistency risks, and realize flexible production scheduling.</p>