Modeling of a Roll-to-Roll Packaging Process Using System Identification and Deep Learning

Abstract

Abstract. Linear models, with parameters derived from physics-based formulations or learned through linear system identification techniques, are commonly applied to characterize disturbance propagation in roll-to-roll continuous manufacturing systems. However, such models inherently struggle to capture the nonlinear disturbance dynamics associated with these processes. This study introduces a hybrid framework of learned models, that combines linear transfer functions and nonlinear autoregressive paradigms to represent web dynamics in a large-scale roll-to-roll packaging system. The results demonstrate that the learned-model approach significantly improves the accuracy of modeling nonlinear disturbance propagation compared to physics-based formulations or strictly learned linear models. To that end, this modeling strategy establishes a foundation for real-time model-based design, simulation-based validation, and online implementation of advanced control strategies aimed at mitigating complex disturbances in industrial web handling applications.