桥面风场时程重构的机器学习方法  

Wind Time History Reconstruction Around Bridge Deck Based on Machine Learning

作　　者：战庆亮 刘鑫 张冠华 白春锦 葛耀君[3] ZHAN Qing-liang;LIU Xin;ZHANG Guan-hua;BAI Chun-jin;GE Yao-jun(College of Transportation and Engineering,Dalian Maritime University,Dalian 116026,Liaoning,China;Technology Research and Development Center,Liaoning Provincial Transportation Planning and Design Institute,Shenyang 110111,Liaoning,China;Key Laboratory of Transport Industry of Wind Resistant Technology for Bridge Structures,Tongji University,Shanghai 200092,China)

机构地区：[1]大连海事大学交通运输工程学院,辽宁大连116026 [2]辽宁省交通规划设计院有限责任公司技术研发中心,辽宁沈阳110111 [3]同济大学桥梁结构抗风技术交通行业重点实验室,上海200092

出　　处：《中国公路学报》2023年第8期22-31,共10页China Journal of Highway and Transport

基　　金：大连海事大学博联科研基金项目(3132023619);桥梁结构抗风技术交通行业重点实验室(上海)开放课题(KLWRTBMC21-02);国家自然科学基金项目(51978527);辽宁教育厅研究计划项目(批准号:LJKZ0052)。

摘　　要：获得桥面的高分辨率时变流场对研究桥梁风致问题尤为关键,然而受传感器布设与测量方法等因素制约,难以通过试验直接测得高空间分辨率的流场数据。机器学习是流场表征的有效手段,但是数据驱动的训练方法在已知样本较少时难以获得准确的模型。针对此问题,引入流场时程的人工神经网络方法,使用流体控制方程辅助模型训练,通过增加未知测点处的方程约束提高模型的精度,得到了考虑物理约束的桥面风场时程的机器学习重构模型。以低雷诺数桥面绕流为例,实现了基于稀疏已知测点时程数据的模型训练,得到了较好的效果。结果表明:通过引入未知测点处的控制方程约束,可在较少已知时程数据的情况下,获得更准确的桥面风场重构模型,为人工智能方法在风场实测时程数据中的应用提供了基础。Obtaining high-resolution,time-varying flow fields around bridge decks is essential for studying wind-induced problems.However,owing to factors such as the sensor layout and measurement methods,it is difficult to measure high-spatial-resolution bridge flow field data directly from experiments.Machine learning is an effective method for flow representation.However,it is difficult for data-driven methods to obtain accurate models when few training samples are available.An artificial neural network computational method for flow time history data was developed,and a machine-learning reconstruction model of the wind field time history considering physical constraints was obtained using governing equations to aid model training.Basic neural-network models were used to represent the time histories of the flow field at different locations.The error between the known time histories of the wind measurement points and the control equations of the unknown measurement points was used to train the models.The flow around a bridge deck at low Reynolds numbers was investigated as a case study,and good results were achieved with model training based on sparse known time history data.The results show that the method used introduces constraints on the flow control equations,enabling the model to obtain a more accurate reconstruction model of the bridge deck wind field with fewer known time series data and providing support for the application of artificial-intelligence methods to the measurement of bridge deck wind field time history data.

关 键 词：桥梁工程 桥面风场时程 人工智能 时程深度学习 物理方程约束 流场重构 

分 类 号：U441[建筑科学—桥梁与隧道工程]