Optimization of tobacco drying process control based on reinforcement learning

Drying is an important procedure in tobacco production. The current PID based drying suffers from issues such as overheating or inconsistent control of the amount of moisture content. In order to boost quality assurance, reinforcement learning has been employed in this paper to facilitate dynamic configuration of dryer. A novel actor-critic based intelligent system is built on top of the current PID control. The new data-centric approach collects environment and machine states, incorporates historical production data and learns temperature adjustment strategies. Compared to automatic PID control and manual intervention, the introduced intelligence proves to be remarkably more effective to govern the drying and control the moisture content level with consistent performance. The proposed method provides new insights into precision achievement in industrial control process.