Abstract
Thin and ultra-thin strip rolling is a high-technology segment of modern metallurgy supplying aerospace, automotive, electronics, and precision instrumentation industries. Below a strip thickness of 0․5 mm‚ the process stability decreases due to high interaction between rolling force‚ roll gap‚ friction and tension‚ delay of transport in the measurement and control circuit at high rolling speeds as well as thermal influences such as the formation of a time-varying roll crown‚ which increases gauge errors‚ shape defects and dynamic overloads exactly at operating points of maximum productivity․ The automation of multifunction rolling mills has been studied extensively in the past years due to the increasing demands regarding product quality and process stability; a four-layer classification model has been proposed: 1․ Sensing and estimation (L1); 2․ Process and disturbance models (L2); 3․ Control algorithms (L3); 4․ Cyber-physical integration (L4). The studies reviewed in this paper include classical AGC/HGC‚ model predictive control‚ coordinated control‚ and many contemporary data-driven approaches including deep learning and reinforcement learning․ These recent works show the relevance of sensing and estimation technologies for rolling mill process automation applications․ Other closed-loop and in-process control solutions include roll-force estimation using drive signal‚ online roll profile measurement and compensation‚ as well as vision-based online quality inspection․ Maintaining the desirable maintenance conditions for roll wear‚ lubrication evolution‚ vibration dynamics‚ and residual stress evolution over the entire service life of the roll is necessary․ Quantitative benchmarking over the seven schemes across three metrics of interest (gauge‚ tension tracking error‚ disturbance rejection settling time) reveals a clear progression from classical single-loop feedback‚ through coordinated and predictive learning to data-intensive and learning-driven architectures․ AGC/PID (Automatic Generation Control/ Proportional Integral Derivative control) is used as a baseline‚ with 8․0 µm gauge deviation and 0․42 s settling time․ With the lowest reported gauge error (2․1 µm) and the shortest settling time (0․15 s)‚ DDPG (Deep Deterministic Policy Gradient) reinforcement learning stands out as state of the art․ However‚ we identify three directions for future research: hybrid adaptive systems‚ reinforcement learning for representing narrow compensation problems under safety constraints‚ and closed loop digital twins that enable real-time optimization and energy driven mill operation․