Leveling Control of Multi-Cylinder Hydraulic Press: A Deep Reinforcement Learning Approach Based on Integral Compensation and Lyapunov Constraints

Abstract

Abstract. In the field of complex industrial control, the leveling task of multi-cylinder hydraulic presses imposes stringent requirements on control accuracy and system stability. Traditional control methods struggle to balance the performance and stability when facing unknown models due to their reliance on precise system modeling. In contrast, reinforcement learning optimizes policies through autonomous interaction, achieving both model-agnostic capability and multi-scenario adaptability. Based on the simplified dynamic model of the multi-cylinder hydraulic press, this study proposes a new control strategy based on reinforcement learning. The approach integrates the soft actor–critic (SAC) algorithm with Lyapunov constraints and state-error integral compensation for leveling control. Embedding Lyapunov constraints within SAC ensures system stability, while the integral compensation minimizes steady-state error and enhances precision. Experimental results demonstrate that—under simplified modeling assumptions—the proposed method retains SAC’s inherent advantages while significantly improving stability and leveling accuracy in specific complex scenarios (e.g., model-defined disturbances). By merging classical control theory with modern machine learning, this work offers new insights for designing reinforcement-learning controllers in complex settings and establishes a foundation for future validation on more realistic physical models. It aims to provide a reference for potential industrial deployment.