Novel Active Structural Vibration Control Strategy Based on Deep Reinforcement Learning

Abstract

Numerous control systems have been devised for addressing structural responses under dynamic loading, especially when inherent structural resistance is inadequate. The Neural Network (NN) controller, an active control method, offers simple solutions for complex structural vibration problems in uncertain environments. However, prior NN-based control strategies often failed to achieve optimal control performance. This research introduces a cutting-edge deep reinforcement learning driven active vibration control strategy, utilizing NN controllers' learning potential while delivering control performance comparable to model-based optimal controllers. The learning algorithm forms control strategy through environmental interactions, without requiring dynamic system model knowledge. This model-free approach provides optimal control performance across various excitations. The strategy was tested on a 6-story shear-building model using full-state feedback and expanded to partially observed feedback scenarios. The results reveal that the control performance closely matches a traditional linear quadratic regulator. Additionally, the learned controller outperforms conventional output feedback controllers in partially observed systems.