基于时程深度学习的复杂流场流动特性表征方法

Deep learning representation of flow time history for complex flow field

流场的特征分析与表征研究对流动机理的明确具有重要意义.然而湍流流场具有复杂的非定常时空演化特征,对其流场数据的低维表征有一定困难.针对此问题,本文提出了基于流场时程数据深度学习方法的湍流低维表征模型,实现了复杂流动数据的降维表征.分别建立了基于一维线性卷积、非线性全连接和非线性卷积的自动编码方法,对非定常时程数据进行降维并得到了低维空间到时域的解码映射关系,实现了特征提取与压缩.通过Re=2.2×10^(4)的方柱绕流场进行了研究与验证,结果表明:时程深度学习方法可以有效地实现流场的低维表征,适用于复杂湍流问题;非线性一维卷积自编码器对复杂流场的表征准确性优于全连接和线性卷积方法.本文方法是无监督训练方法,可应用于基于一点的传感器数据处理中,是研究复杂流场特征的新方法.

Flow analysis and low-dimensional representation model is of great significance in studying the complex flow mechanism. However, the turbulent flow field has complex and unstable spatiotemporal evolution feature,and it is difficult to establish the low-dimensional representation model for the flow big data. A low-dimensional representation model of complex flow is proposed and verified based on the flow time-history deep learning method. One-dimensional linear convolution, nonlinear full connection and nonlinear convolution autoencoding methods are established to reduce the dimension of unsteady flow time history data. The decoding mapping from low-dimensional space to time domain is obtained to build the representation model for turbulence. The proposed method is verified by using flow around the square clyinder with Re = 2.2×10^(4). The results show that the flow time-history deep learning method can be used to effectively realize the low-dimensional representation of the flow and is suitable for solving the complex turbulent flow problems;the nonlinear one-dimensional convolutional autoencoder is superior to the full connection and linear convolution methods in representing the complex flow features. The method in this work is an unsupervised training method, which can be widely used in single-point-based sensor data processing, and is a new method to study the characteristics of turbulence and complex flow problems.