Research And Implementation Of Sparse Flow Field Processing Method Based On Deep Learning

Flow field data processing is an important research direction of Computational Fluid Dynamics.In recent years,with the rapid development of deep learning and high-performance computing,some neural network models have emerged for the research of complex feature extraction of flow field data,but there are still a large number of sparse flow field data in reality and experimental simulations which increases the difficulty of precise analysis.And how to realize efficient,accurate and fast numerical calculation and feature extraction process by using sparse flow data is a constraint to the development of deep learning in the direction of computational fluid dynamics.This paper proposes a method of data preprocessing and amplification based on numerical calculation of sparse fluid surface by designing a variety of unsupervised learning algorithms such as generative adversarial networks,auto-encoders and their variants,and optimizing loss function for sparse flow field parameters and grids,etc,in which the features are fitted in a high dimension,in order to provide more data for data flow classification and regression tasks.Specific content includes:1.Generation of sparse flow field data based on generative adversarial networkSince generative adversarial network performs well in the field of generative model,a continuous expression after combining of the grid structure and discrete hyperparameters are diversified by training and optimizing the network.During this experiment,the training stability and calculation efficiency are improved much by optimizing the target of loss function and operator on TensorFlow.2.Generation of sparse flow field data based on variational auto-encoderAs the experimental result shows that the accuracy is not high enough and the training period is still unstable,variational auto-encoder is proposed to replace the random noise with the real data which participates in the hidden variable satisfying the Gaussian distribution,which increases the accuracy of the generated data.And the combination of variational auto-encoder and generative adversarial network performs even better on multiple baselines.3.Design and implementation of sparse flow field data amplification prototype systemBased on the research and improvement of the sparse flow field data three-dimensional representation and data generation models,a sparse flow field data amplification prototype system was designed and implemented.The data preprocessing,three-dimensional representation and transformation as well as multiple generation models with evaluation and visualization were designed to form a complete data flow module,in which each sub-module provides an user interaction interface separately.