2025 Volume 51 Issue 4 Pages 103-109
When applying AI control systems to industrial plants, there is a problem in that the control accuracy decreases if the operational data for learning the control method is insufficient. In this paper, we propose a method for collecting additional learning data using a simulator when the desired control performance cannot be achieved through normal learning using pre-collected operational data. First, we define the reward obtained when reaching the target state and calculate the state value, which is the expected value of the reward. Second, we calculate the standard deviation of the state value and identify the areas of high standard deviation as areas of insufficient learning data. Third, we collect the additional learning data by using trial-and-error control set values in these regions. Finally, we improve the control performance through additional learning using the collected data. The effectiveness of the proposed method was confirmed by using a vinyl acetate monomer plant simulator. Through analysis of the recovery operation for vaporizer pressure when abnormal ethylene feed pressure occurs, we confirmed that overshoots can be suppressed using additional learning, improving the settling time from 67 to 37 min.
