非定常流场时程重构的深度学习方法  

Unsteady flow time history reconstruction based on deep learning

作　　者：战庆亮 白春锦[1] 吴智虎 葛耀君 ZHAN Qing-liang;BAI Chun-jin;WU Zhi-hu;GE Yao-jun(College of Transportation and Engineering,Dalian Maritime University,Dalian 116026,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]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：《船舶力学》2024年第3期319-327,共9页Journal of Ship Mechanics

基　　金：大连海事大学博联科研基金项目(3132023619);国家自然科学基金项目(51978527);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);辽宁省教育厅研究计划资助(LJKZ0052);中央高校基本科研业务费专项资金资助(3132022189)。

摘　　要：高分辨率的流场数据对流动问题的研究具有重要意义。受测量方法、计算效率等多因素限制,高分辨率流场的直接获取仍有一定困难。本文基于流场时程数据的低维表征模型,提出非定常流动时程数据重构的深度学习方法。该方法直接面向样本时程数据,凭借一维卷积的特性提取出样本中包含的时程特征;然后,建立物理空间与表征模型编码空间之间的映射关系;最后,利用一维反卷积对低维表征进行解码,实现对流场中任意位置数据的重构。对Re_(D)=200的非定常圆柱层流绕流流场进行低维表征与验证,进而实现高分辨率流场时程数据的重构,并证明方法的准确性。本文方法是一种无监督方法,是一种时间维度上具有高精度的流场数据重构方法,适用于基于传感器的时程数据处理。High-resolution flow field data are of great significance to the study of fluid mechanics.Limited by measurement methods and calculation efficiency,it is still difficult to obtain high-resolution flow fields di⁃rectly in some circumstances.A low-dimensional representation model for flow time history data was poposed,and a deep learning method for reconstruction of unsteady flow time history data was developed.The proposed method extracted the time-history features contained in the samples using one-dimensional convolution directly;then,the mapping from the physical space and the encoding space was built;and finally,the decoder in the representation model was utilized to generate flow time history data at unknown positions.Unsteady laminar flow with Re_(D)=200 was studied,and the accuracy of the method was verified.The method proposed in this paper,a new flow field data reconstruction method in an unsupervised training manner in the time dimension,can be widely used in point-based sensor data analysis.

关 键 词：流场重构 流场时程 深度学习 特征提取 无监督模型 

分 类 号：O357.5[理学—流体力学]