QMIX Multiple Intelligences Reinforcement Learning Damping Control for Cylindrical Shell

Abstract

One of the fundamental mechanical constructions of ships and navigators is the cylindrical shell structure. Their damping control is difficult to predict and frequently depends on precise control models. For that reason, this work provides a data-driven multi-intelligence reinforcement learning damping control approach that is significance for damping control of massive structures. Firstly, the dynamics equations of cylindrical shell structure are established based on the hypothetical modal method, and modal variables are introduced to derive the state-space equations for damping control of cylindrical shell structure, and an interactive environment for multi-intelligent reinforcement learning is established. Secondly, the damping control strategy of cylindrical shell structure with multiple intelligences is designed based on the value decomposition QMIX algorithm. For a single smart body design vibration displacement, velocity, piezoelectric actuator voltage, smart body operation steps as the state space, quadratic performance indicators with saturation characteristics as the damping effect reward function, greedy strategy as damping action selection method for multi-intelligent body cooperative damping. The QMIX algorithm hybrid network performs fusion evaluation of the joint action value of each intelligence and updates the action value function of a single intelligence. Finally, five sets of hyperparameters are set based on the Grid Search approach for comparative simulation experiments for deep learning network hyperparameter selection. The result of the simulation demonstrate that the current tactic effectively suppresses the vibration of the cylindrical shell construction. Furthermore, the optimal hyperparameter is determined by comparing simulation trials with different values, proving that the approach described in this article has better damping performance under this parameter.