UNDERWATER SHAKING TABLE TIME DELAY COMPENSATION AND CONTROL OPTIMIZATION BASED ON REINFORCEMENT LEARNING

The water dynamics and their effects on the different direction vibrations of facility are complicated and difficult to compensate during underwater shaking table tests. As a result, before conducting the tests, it is necessary to confirm the capacity and accuracy of facility. This paper investigated the effect of several factors on the coupling control performance of a hydrodynamic-shaking table, such as water depth, excitation frequency, and movement directions. The transfer function model for the water-shaking table system is firstly identified using measured data, and then a data-driven hybrid control strategy is proposed, combining model-based feedforward compensation and reinforcement learning (RL). The Actor-Critic networks in RL are trained offline using the error data of displacement commands according to the DDPG algorithm, and they are utilized to compensate the model-based commands in real-time. By comparing with the feedforward compensation, 50 test cases were conducted, considering different water depths, excitation frequency and shaking directions, to validate the method and to evaluate its performance. The results reveal that: the control performance decreases with the increase of water depth and excitation frequency; the water depth has a greater impact on the vertical motion of a shaking table. Under the most unfavorable condition of the vertical motion with a water depth of 2m, the proposed method enhanced the performance with 6.54% and 7.52% for indicators J₁ and J₂, respectively. The proposed method has an optimized time-delay compensation effect when considering the nonlinear dynamics of the water-shaking table interaction system, and it is also a broadband compensation technique.