Effective system identification is critical to realize the high-performance control of shaking table, however, the common reduced-fidelity models can rarely capture the practical dynamics of the system over a wide frequency range. To address this issue, this study adopts the latest advances in deep learning to develop a physics-guided long short-term memory (PhyLSTM) network for system identification of the shaking table. The basic idea is to embed the physical laws describing the system into the network, which can prompt the model to converge quickly to the right solution and constrain the output to a rational solution space in the case of poor data quality. To give insight into the operating mechanism of the shaking table, a series of tests were designed to reveal the flow properties of servovalve and the compressibility of hydraulic oil. Based on the above cognition, a testing program including various signals and specimens was conducted to complete data collection. After detailed hyperparameters testing, PhyLSTM exhibited far superior performance to traditional transfer function model, which indicated that PhyLSTM is a promising approach for shaking table modeling.
Read full abstract