Advanced Machine Learning Approaches for Quality Estimation in Polymer Industry 4.0

The evolving demands of modern polymer manufacturing require advancements in process optimization and quality monitoring. Laboratory-based systems, characterized by delays, infrequent measurements, and high costs, are insufficient for dynamic industrial environments. Industry 4.0 offers transformative potential, particularly through soft sensors, which exploit real-time process data to estimate quality parameters. First-principles models provide high accuracy but are costly and process-specific, while data-driven models are more flexible and accessible. Models used for polymer quality monitoring rely on very complex approaches, such as deep learning methodologies and non-linear mathematical models, resulting in accurate though less interpretable. In cases with limited training data and need for interpretability, models like Partial Least Squares (PLS) are more suitable. PLS model is robust, interpretable, and well-established but it has limitations, particularly its inability to adapt to changing process conditions. Recursive methodologies have been developed to enable model adaptation, by considering the sequential nature of the data or by focusing on the most relevant observations. Another limitation of PLS is its inability to account for autocorrelation, where a variable’s current value is influenced by its previous measurements. Methodologies incorporating autoregressive terms have been proposed to effectively model such temporal dependencies. Another general challenge is the need to identify optimal parameters for the models to ensure their effectiveness. Formal optimization methodologies offer high accuracy but they are complex, require expertise, and can be difficult to implement in practice. Heuristics and iterative approaches present more flexible alternatives, but they also pose challenges, such as accounting for the sequential nature of the data and the interactions between parameters. While established approaches provide reliable and robust solutions to many challenges, they fall short of fully addressing the complex scenarios examined in this work. This Dissertation addresses two key challenges in the development of adaptive soft sensors in polymer manufacturing: quality estimation in datasets with weak cross-correlation between process variables and product quality but high autocorrelation of the quality variable, in high-variability environments while maintaining model interpretability; joint parameter optimization in processes characterized by frequent campaign changes and the need for regular recalibration, accounting for temporal dependencies in the data. To address the first challenge, the Proportional-Integral-Derivative Partial Least Squares (PID-PLS) model is proposed. This methodology incorporates temporal variability through proportional, integral, and derivative components to model both abrupt changes (e.g., grade transitions) and more subtle cumulative effects (e.g., fouling or set point shifts). The model is validated using Versalis’ industrial case study on Ethylene-Propylene Diene Monomer (EPDM) production and a controlled numerical scenario. Results show superior performance compared to state-of-the-art methods, improving predictive accuracy under time-varying and autocorrelated conditions by effectively capturing both sharp transitions and gradual process changes. For the second challenge, the Recursive Adaptation (RA) algorithm is introduced to optimize model parameters jointly and adaptively. Validated on Versalis’ industrial case studies for EPDM and General-Purpose Polystyrene (GPPS) production, this algorithm outperforms conventional heuristics approaches. This algorithm facilitates rapid and efficient optimization, enabling real-time adaptable models for quality monitoring. A key improvement achieved through the soft sensors optimized by this methodology is the enhanced estimation accuracy during critical production periods, such as the initiation of a new production campaign.