振动台子结构试验具有试验测试与数值模拟的双重优势,可实现大尺度甚至足尺模型结构的地震作用复现;振动台子结构试验中数值子结构的实时求解是影响试验精度和系统稳定性、决定试验能否顺利实现的关键因素之一.为提升数值子结构的求解性能,将长短时记忆(long short-term memoryk,LSTM)网络引入到振动台子结构试验中,分别构建了用于模拟试验子结构和数值子结构的神经网络模型,并且在训练数据中引入时滞以模拟系统延迟带来的影响.选择5层钢框架模型结构对神经网络模型进行验证,结果显示,所构建的神经网络模型具有良好的精度、稳定性和时滞补偿能力,计算效率可满足实时控制要求;所提出的神经网络模型可用于振动台子结构试验的数值子结构实时求解.
Shaking table substructure test has dual advantages of real time testing and numerical simulation.It can realize the reappearance of seismic actions on large-scale or even full-scale model structures.Real-time solving of numerical substructures in shaking table substructure tests is one of key factors to affect test accuracy and system stability and determining whether tests can be successfully realized.Hence,to improve solving performance of numerical substructures,long short-term memory(LSTM)network was introduced into shaking table substructure tests.Neural network models were constructed to simulate test substructures and numerical substructures,respectively and time delay was introduced into training data to simulate effects of system time delay.A 5-story steel frame model structure was chosen to verify neural network models.The results show that the constructed neural network models have good accuracy,stability and time delay compensation ability,and the computational efficiency can meet real-time control requirements;the proposed neural network models can be used for real-time solving of numerical substructures tests.