Prediction of Lateral Web Deviation in Transport Using MPC Framework with Deep Reinforcement Learning-Based Extended Kalman Filter
Year: 2026
Type: conference-paper
Abstract: This study proposes a framework for adaptive estimation and prediction of lateral web deviation in roll-to-roll (R2R) systems that transport and wind thin continuous materials such as paper and films. Lateral web deviation arises from time-varying material properties and external disturbances, leading to deformation of the wound web and degradation of product quality. Because these effects are difficult to model accurately and the web is transported at high speed, it is necessary to estimate and predict deviations in advance. To address this issue, we design a deep reinforcement learning–based extended Kalman filter (DRL-EKF) for adaptive state estimation under time-varying uncertainties. Based on the estimated states, we propose an integrated DRL-EKF–MPC framework to predict future lateral web deviation. Experimental validation using the developed R2R system demonstrates that the DRL-EKF improves estimation accuracy, and the proposed method enables reliable prediction of future web behavior under external disturbances and varying operating conditions.
Institutions: Shibaura Institute of Technology, Toshiba (Japan)
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204201415
