基于多层感知深度学习的大跨度斜拉桥索力调整  

Cable Force Adjustment for Long-Span Cable-Stayed Bridge Based on Multilayer Perceptron Deep Learning

作　　者：单德山[1] 张潇 顾晓宇 李乔[1] SHAN De-shan;ZHANG Xiao;GU Xiao-yu;LI Qiao(School of Civil Engineering,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学土木工程学院,四川成都610031

出　　处：《桥梁建设》2021年第1期14-20,共7页Bridge Construction

基　　金：国家重点研发计划项目(2016YFC0802202);国家自然科学基金项目(51678489,51978577);四川省科技计划项目(2016JY0130);云南省交通运输厅科技项目(2017(A)03)。

摘　　要：为提高大跨度斜拉桥施工过程中索力调整的速度及准确性,基于多层感知深度学习,构建了索力调整的深度网络架构。将索力调整实质定义为:在斜拉索无应力长度允许调整的范围内,拟合目标响应与索力调整量之间的映射关系,转化为机器学习和统计学中的回归问题。结合深度学习的二阶梯度下降和深度网络正则化策略,采用多层感知器索力调整的4层深度神经网络,以某混合梁斜拉桥为工程背景,验算数学和结构响应两方面下索力调整量的预测误差,及调索后的结合梁线形及索力误差。结果表明:预测误差均在工程允许范围内,成桥阶段调索后的结合梁线形误差在40mm以内,索力误差均在5%以内;索力调整的多层感知深度网络能快速、精确地预测索力调整量,可用于大跨度斜拉桥的索力调整。Based on multilayer perceptron deep learning,a novel deep learning neural network for cable force adjustment is proposed,aiming to improve the speed and accuracy of cable force adjustment in the construction process of long-span cable-stayed bridges.The essence of cable force adjustment is to map the relationship between the target responses and the cable force adjustment amount within the allowable adjustment ranges for the unstressed length of the stay cables,and then the cable force adjustment problem is transformed into regression problem in machine learning and statistics.Combined with the second-order gradient descent algorithm and deep network regularization strategy,a specific 4-layer neural network for cable force adjustment is constructed,and a hybrid girder cable-stayed bridge is taken as the engineering background,the prediction error for cable force adjustment from the perspectives of both mathematics and structural response is evaluated,and the alignment error of the composite girder and cable force error after cable force adjustment are also verified.It is shown that the prediction errors are all within the allowable ranges,and girder alignment error and cable force error are less than 40mm and 5%,respectively,in the completed bridge state.The proposed multilayer perceptron deep learning network can rapidly and accurately predict the cable force adjustment amounts,which can be adopted to adjust the cable forces of long-span cable-stayed bridges.

关 键 词：大跨度斜拉桥 斜拉索 目标响应 索力调整 多层感知器 深度学习 线形 

分 类 号：U448.27[建筑科学—桥梁与隧道工程] U443.38[交通运输工程—道路与铁道工程]