非平稳流动系统降维表征与重构的时程深度学习方法

Reduced-order representation and reconstruction of non-stationary flow system using flow-time-history deep learning

本文提出了非平稳流动系统降维表征与重构的时程深度学习方法.与基于瞬态流场快照的深度学习方法不同,本文提出基于流场时程数据的研究方法,实现了基于一维卷积的非平稳流动系统的无监督训练模型.首先使用流场时程自编码模型建立了高维非平稳系统的降维表征模型,然后采用多层感知器构造流场测点的物理位置与编码隐变量空间的映射,最终实现复杂的非平稳流动时程数据的高分辨率重构.对低雷诺数圆柱绕流的非平稳流动发展过程进行了研究,采用有限测点的训练数据得到高空间分辨率的流场时程数据.本文所提出的方法是研究非平稳复杂流动系统的新方法,同时模型的无监督训练性质保证了该方法广泛适用于流场传感器所采集的流动时程信号处理.

A representation and reconstruction model for the time-variant flow features of non-stationary flow system is proposed.Different from snapshot-based methods,flow-time-history(FTH)data are used to extract and learn the hidden non-stationary features via self-supervised deep learning strategy.First,the high-dimensional non-stationary flow system is reduced to a low-dimensional latent representation using the flow-time-history autoencoder(FTH-AE).Second,a mapping from flow physical space to the latent code space is established using multi-layer perceptron neural network.Finally,FTH data for positions that have not been measured are generated using the FTH generator model.The proposed method is validated using data from non-stationary flow around a cylinder at Re=200.High-resolution results of the flow development process are simulated using sparse available FTH data.The results demonstrate that the proposed model is a new low-dimensional representation method for non-stationary flow systems.It is a self-supervised model that does not require paired or labeled datasets and is suitable for the abundant FTH data from simulations and experiments.