Model-free active vibration control based on imitation learning and deep reinforcement learning: experimental implementation in an SDOF structure

Abstract As a model-free method, neural network (NN) control learns the control policies directly from measurement data and eliminates the need for precise structural models and parameters. However, how to achieve optimal control performance remains a challenge. The development of deep reinforcement learning (DRL) algorithms provides a solution to achieve model-free optimal control theoretically. However, deploying DRL in real-world structural vibration control problems faces practical obstacles, considering its training requires many interactions with a real-world structure, which is impractical due to safety risks, time constraints, and operational costs. A novel imitation learning (IL) plus DRL framework is proposed to address these issues for the first time. The NN controller is pre-trained by IL using data from a reference controller to reduce the risk from random initial exploration phases. This pre-trained controller is subsequently fine-tuned through DRL to achieve optimal control performance. This framework was experimentally validated on a single-degree-of-freedom (SDOF) structure equipped with an active mass damper (AMD). Results demonstrated that the proposed approach can successfully train an optimal controller without any structural model information of the tested SDOF structure-AMD system. These findings show that combining IL and DRL can significantly improve the sampling efficiency and performance of NN controllers, making them suitable for real-world structural vibration control applications.