基于神经网络的既有站房结构位形反演法  

Configuration Inversion Algorithm for the Existing Long-Span Spatial Structures of Existing High-Speed Railway Station

作　　者：李浩 邹勇 LI Hao;ZOU Yong(Huizhou Residential Apartment Section,Guangdong Railway Co.,Ltd.,China Railway Guangzhou Group Co.,Ltd.,Huizhou Guangdong 516023,China)

机构地区：[1]中国铁路广州集团有限公司广东铁路有限公司惠州房建公寓段,广东惠州516023

出　　处：《中国铁道科学》2023年第6期25-33,共9页China Railway Science

基　　金：国家自然科学基金资助项目(51678431)。

摘　　要：针对传统位形检测受限于样本容量而无法建立完备的既有结构模型问题,提出一种可用于既有大跨度站房结构鉴定的基于神经网络的结构位形反演算法。依据大跨空间结构线弹性假定,以及位形偏差与结构体系及刚度差异的相关性,构建利用等效随机荷载叠加获得的结构位形基向量空间;以位形基向量和抽样节点实测位形作为输入,利用BP神经网络训练出结构实际整体位形,并以深圳坪山高铁站大跨空间结构作为实例进行验证及性能分析。结果表明:随机荷载数和训练样本数是影响反演算法误差和收敛速度的主要因素,对于与深圳坪山站站房屋面同类型的大跨空间结构,当基向量数大于200个时,训练样本数选取300组,能够达到误差小、收敛快的反演效果;基于神经网络的结构位形反演算法可依据少量节点位形测量结果较为理想地反演结构真实的整体位形。In response to the limitations of traditional structural configuration detection due to insufficient sample capacity in establishing comprehensive existing structural models,a neural network-based structural configuration inversion algorithm is proposed for identifying existing large-span station structures.Based on the linear elastic assumption of large-span spatial structures and the correlation between configuration deviations and differences in structural systems and stiffness,a structural configuration base vector space is constructed by superimposing equivalent random loads.Using configuration base vectors and measured configurations at sampled nodes as input,a backpropagation(BP)neural network is trained to obtain the actual overall structure configuration.The performance of the algorithm is validated and analyzed using the large-span spatial structure of the Pingshan highspeed railway station in Shenzhen as an example.The results indicate that the number of random loads and training samples are the primary factors influencing the accuracy and convergence speed of the inversion algorithm.For large-span spatial structures of the same type as the Pingshan station building,when the number of base vectors exceeds 200 and 300 sets of training samples are selected,the inversion effect with low error and fast convergence speed can be obtained.The proposed neural network-based structural configuration inversion algorithm can effectively infer the true overall configuration of the structure based on a small number of measured node configurations.

关 键 词：既有站房 结构位形 检测 反演算法 BP神经网络 随机荷载 基向量 

分 类 号：TU393.3[建筑科学—结构工程]