【作者】 王俊；

【导师】 郭磊； 王昉；

【作者基本信息】 电子科技大学 ， 工程硕士（专业学位）， 2021， 硕士

【摘要】 随着计算流体力学(CFD)应用的精度需求不断提升,网格量越来越大,CFD产生的流场数据量达到了TB甚至PB量级。流场数据的时空复杂性提升,会导致时空特征难以辨认,也需要研究者耗费更多的时间人工抽取关键信息帮助认知流场中的复杂流动机理。如何自动抽取流场特征及关键时间步,将成为研究的热点,也是研究者面临的巨大挑战。近年来,深度学习的快速发展给各领域解决问题提供了新的思路。深度学习技术可以对海量数据进行特征发现和信息提取,极大提升了数据分析效率和准确性,已成为数据挖掘和特征提取的主流技术。本文依托于深度学习的方法对流场数据进行涡特征和关键瞬态提取,主要工作如下:(1)调研国内外在流场涡特征提取以及关键瞬态提取的研究进展,总结目前流场数据特征提取的主要研究方法,分析现有研究方法的优缺点,为本文所提出的方法做铺垫。(2)基于深度学习的涡特征识别算法研究,涡特征是理解流场潜在物理机制的重要手段。局部涡识别方法需结合人工选择合适的阈值判断是否为涡,且鲁棒性较差。全局涡识别方法计算复杂度高,耗时长。机器学习方法与流场的大小和形状有关,通用性和可扩展性较差。针对上述问题,本文提出了一种基于卷积极限学习机的涡识别方法。该方法能够从流场中快速、客观、鲁棒地检测出涡。通过大量的实验结果证明了该方法的有效性。(3)基于深度学习的关键瞬态提取算法研究。将自编解码器引入到计算流体力学流场关键时间步的选择中,提出了一种基于自编解码器的全局选择方法。与现有的单一的动态规划以及聚类思想的方法相比,该方法在选取的关键时间步流场,能更准确选取具有代表性的流场数据,并且选取的结果能准确表示流场数据的变化趋势。通过大量的实验结果证明了该方法在CFD数据集上良好的推广效果。实验证明,本文提出的两种算法框架在各自对应的流场数据特征与关键瞬态提取问题中都能取得较为理想的效果。更多还原

【Abstract】 With the continuous improvement of the accuracy requirements of Computational Fluid Dynamics(CFD)applications and the increasing number of grids,the flow field data volume generated by CFD reaches the magnitude of TB or even PB.As the temporal and spatial complexity of flow field data increases,it is difficult to identify the temporal and spatial features,and researchers need to spend more time manually extracting key information to help understand the complex flow mechanism in the flow field.How to automatically extract flow field features and key time steps will become a research hotspot and a huge challenge faced by researchers.In recent years,the rapid development of deep learning has provided new ideas for solving problems in various fields.Deep learning technology can find features and extract information from massive data,greatly improving the efficiency and accuracy of data analysis,and has become the mainstream technology of data mining and feature extraction.This paper relies on deep learning method to extract vortex features and key time-steps from flow field data,and the main work is as follows:1.To investigate the current research progress of vortex feature extraction and key time-steps extraction in flow field at home and abroad,summarize the current main research methods of flow field data feature extraction,analyze the advantages and disadvantages of various research methods,and lay a technical foundation for the method proposed in this paper.2.Research on Vortex Feature Recognition Algorithm Based on Deep Learning.Vortex Feature is an important means to understand the potential physical mechanism of flow field.The local vortex identification method needs to be combined with artificial selection of appropriate threshold value to judge whether it is a vortex,and its robustness is poor.Global vortex recognition method identification method is computationally complicated and time-consuming.The machine learning method is related to the size and shape of the flow field,which has poor universality and scalability.To solve these problems,a vortex recognition method based on convolutional extreme learning machine is proposed in this paper.This method can detect vortexes from the flow field quickly,objectively and robustly.The effectiveness of this method is proved by a large number of experimental results.3.Research on key time-steps extraction algorithms based on deep learning.The self-codec is introduced into the selection of the key time steps of the computational fluid dynamics flow field,and a global selection method based on the self-codec is proposed.Compared with the existing single dynamic programming and clustering methods,this method can more accurately select representative flow field data in the selection of key time-step flow field,and the selected results can accurately represent the change trend of flow field data.A large number of experimental results show that this method is effective in CFD data set.Experiments show that the two algorithm frameworks proposed in this paper can achieve ideal results in their respective flow field data characteristics and key time-steps extraction problems.更多还原