CHENG Rongjian, FANG Yixiang, ZHAO Yi, ZHANG Tianzhu, LI Jun, WANG Junxiang
(School of Mechanical and Electronic Engineering, Jingdezhen Ceramic University, Jingdezhen 333403, Jiangxi, China)

Abstract: The modern ceramic tile production process can be taken as a complex system, where various regulatory behaviors could affect the quality of the final ceramic tile products. The process parameters of traditional ceramic tile production are usually determined through multiple experiments, which are based on the experience and behavior of engineers. However, it is difficult to accurately determine process parameters based on empirical behavior. Also, optimal process parameters would dynamically change with actual operating conditions (such as external atmosphere). Under the influence of variable working conditions in ceramic tile production lines, it is difficult to ensure the stability of product quality. In order to solve the manual control problem dominated by experience and achieve dynamic parameter updates in actual processes (working conditions), an intelligent control framework is proposed for the first time for ceramic tile production lines, based on deep reinforcement learning (DRL) algorithm. The framework includes an Environment module and an Agent module. The Environment module simulates and updates various working conditions in the ceramic tile production line, based on data mining technology, where the corresponding product quality could be timely predicted through random forest (RF) prediction model. The intelligent Agent control module can quickly adaptively adjust process parameters based on the predicted product quality, so that quality of the ceramic tile product meets the expectation. It is found that the accuracy of the prediction model constructed by using this method is higher than that of similar methods, with an average improvement rate of 2%. At the same time, after multiple iterations, the intelligent control algorithm for the ceramic tile production lines can increase the qualification of production process parameters to 95%, with desirable control effects.
Key words: deep reinforcement learning; ceramic tile production; production parameters; random forest