Gain parameters optimization strategy of cross-coupled controller based on deep reinforcement learning. Issue 5 (4th May 2022)

• Record Type: Journal Article
• Authors: Zhang, Tie; Wu, Caicheng; He, Yingwu; Zou, Yanbiao; Liao, Cailei
• Is Part Of: Engineering optimization. Volume 54:Issue 5(2022)
• Page Start: 727
• Page End: 742
• Publication Date: 2022-05-04
• Subjects: Parameter optimization -- cross-coupled controller -- contour error -- deep reinforcement learning; Engineering design -- Periodicals; Mathematical optimization -- Periodicals; 620.0042
• DOI: 10.1080/0305215X.2021.1897801
• Languages: English
• ISSNs: 0305-215X

Abstract

Abstract : In this article, a deep reinforcement learning-based strategy for the optimization of gain parameters of a cross-coupled controller is proposed. First, to compensate the contour error caused by servo lag, a proportional–integral (PI)-type cross-coupled controller is designed, with its gain parameters being determined using the cut-off frequency of the contour error transfer function. Next, on the basis of a deep reinforcement learning algorithm, a neural network structure suitable for contour error compensation is established, through which the optimal gain parameters are obtained. Finally, some experiments are carried out, in which the optimal gain parameters sought through off-line learning schemes are applied to the biaxial motion control. The results indicate that the proposed gain parameters optimization strategy can effectively converge the gain parameters with the optimal intervals, and that the optimal gain parameters obtained by the proposed strategy can significantly improve the contour control accuracy in biaxial contour tracking tasks.