【摘要】 为推动高性能的智能化抗震试验平台建设，实现振动台控制算法的智能升级和优化，文中以地震模拟振动台的智能控制为研究对象，提出了一种基于LSTM（长短时记忆网络）的振动台深度学习控制器框架，通过LSTM对三参量控制器的输入-输出关系进行学习和模拟，验证了该控制器的可行性和有效性。针对LSTM依赖于完整且连续轨迹才能保存记忆的局限性，提出了一种基于LSTM闭环控制的实时反馈信号处理方法，即通过单独保存隐层状态“h”和长期记忆状态“c”的方式，避免了LSTM控制器在接收实时反馈信号时会因不断重置而丢失过往记忆。仿真结果表明，通过监督学习的训练方式，深度学习控制器能够有效逼近三参量算法的控制性能，较好地再现地震波加速度时程曲线，这说明深度学习控制器在处理非线性控制问题上具备足够的潜能。更多还原

【Abstract】 In order to promote the construction of high performance intelligent seismic test platform,realize the intelligent upgrade and optimization of shaking table control algorithm,a deep learning controller framework of shaking table based on LSTM(Long and Short-term Memory Network)was proposed in the paper. The feasibility and effectiveness of LSTM controller was verified by training and simulating the input-output relationship of threevariable controller. Considering the limitation that LSTM relies on complete and continuous trajectories to preserve memory,a method of processing real-time feedback signal based on LSTM closed-loop control was proposed,which helps the LSTM controller to avoid the loss of past memory when receiving real-time feedback signals by storing the hidden layer state“h”and the long-term memory state“c”separately. The simulation results shown that the deep learning controller can effectively imitate the control performance of the three-variable algorithm and reproduce the time-history curve of seismic acceleration wave through the training method of supervised learning,which indicated that the deep learning controller has enough potential in dealing with nonlinear control problems.更多还原