PPO‐Based Reinforcement Learning for the Semi‐Active Vibration Control of MDOF Platform

Aiming at the coupled vibration problem of a multi‐degree‐of‐freedom (MDOF) vibration isolation platform under eccentric excitation, this paper proposes a semi‐active vibration control strategy based on Proximal Policy Optimization (PPO) ‐based reinforcement learning (PPO RL). First, the mechanical model of a 3‐DOF platform supported by four magnetorheological dampers (MRDs) and isolators is established. On this basis, the semi‐active vibration control problem is formulated as a Markov Decision Process (MDP). A 20‐dimensional state space is constructed, integrating the motion of the centroid, local support point motion, vibration excitation, and MRD internal hysteresis. Accordingly, a 4‐dimensional continuous action space corresponding to the input current of each MRD is designed. A multi‐objective reward function is developed, which accounts for vibration suppression, differential vibration penalty and control input efficiency. The actor‐critic dual network of the PPO algorithm is optimized with carefully tailored hyperparameters, and GPU acceleration is employed to enhance training efficiency. Simulation results demonstrate that the proposed strategy significantly suppresses both the vertical translation and the angular vibration of the platform's centroid, while effectively controlling the differential vibration of support points. More importantly, by suppressing the angular vibration of the centroid and ensuring the consistency of support point displacement, the strategy achieves high‐precision horizontal attitude maintenance of the platform under eccentric excitation. The maximum relative deviation of the support point displacement in z ‐direction is reduced from 28.5% to 6.2%. This research verifies the effectiveness and superiority of the PPO algorithm for semi‐active vibration control of MDOF vibration isolation platforms, and provides a new intelligent control method for the attitude and vibration control of isolation platforms under asymmetric excitation.