Abstract: To address the challenges of strong nonlinearity, parameter time-variation, and insufficient adaptive capability of traditional methods in hydraulic servo roll gap control during thin strip and ultra-thin strip rolling processes with multi-high mill, this paper proposes an intelligent control method integrating Bayesian estimation with deep reinforcement learning. A control-oriented mechanistic model of the hydraulic servo system for multi-high mills is established. Bayesian estimation is employed to identify key system parameters, achieving a mean square error of 1.788×10^-7 mm². On this basis, a deep reinforcement learning controller based on a multi-head attention mechanism and the TD3 algorithm is designed. The roll gap control problem is formulated as a Markov decision process through incremental action definition and state space construction incorporating historical deviations and process parameters. Simulation results demonstrate that the proposed method converges faster than the standard DDPG algorithm, reduces the tracking error fluctuation range to -3.4×10^-3~1.2×10^-3 mm, and significantly improves control accuracy. This provides intelligent control technical support for high-precision thin strip and ultra-thin strip rolling in multi-high mills.