Three-Dimensional Super Resolution Reconstruction Of Turbulent Flame Based On Model Reduction And Deep Learning

The analysis of the three-dimensional structure of high turbulent flame is helpful to the understanding of the basic physical phenomena related to combustion,such as the three-dimensional propagation law of flame,the turbulence-combustion interaction,and the energy transfer of the vortex structure of reaction flow.One of the keys to the 3D structure analysis of turbulent combustion is to establish a tomography method from 2D projection to 3D structure.The traditional 3D reconstruction method based on iteration requires strict perspective information,the mapping relationship between 2D and 3D structures is complex,and the reconstruction requires a large amount of computing resources.At the same time,the spatial resolution of the tomographic reconstruction results is limited by the number of imaging angles.Although various spatial interpolation algorithms can be used to improve the(nominal)resolution,the mathematical nature makes the reconstruction results too smooth and the high frequency details are lost.To solve these problems,this paper proposes a Rereduced Order Tomographic Model without Angle calibration and a Super-resolution Network Model(3D-SRGAN)based on Generative Adversarial Network to realize the turbulent flame imaging method from 2D low resolution projection to 3D high resolution structure without Angle calibration.In order to verify the effectiveness of the proposed imaging method,two sets of high turbulence flame data(FlameB and FlameC)obtained from large eddy simulation were used to train and test the model method.Qualitatively and quantitatively analyze the reconstruction results through visual analysis,total reconstruction error ER and peak signal-to-noise ratio(PSNR),etc.The results show that the two-dimensional to three-dimensional imaging method proposed in this paper:(1)The rereduced order tomographic model without angle calibration greatly reduces the required computational resources,and the optical center restoration algorithm based on the flame timing symmetry can eliminate the errors caused by the optical lens center deviation,and obtain high-precision reconstruction results;(2)The 3D-SRGAN network can obtain convincing super-resolution reconstruction results under the condition that the single dimension is magnified by 2 times,that is,the total element is increased by 8 times.The overall error of the generated results is about 4%,and the PSNR is higher than 37dB.(3)A complete imaging method from low-resolution two-dimensional projection to three-dimensional high-resolution structure is established by using the rereduced order tomographic model and three-dimensional super-resolution neural network.In addition,the performance of the imaging method was also investigated under higher Reynolds number/noise data,and the results show that the imaging method has good applicability.Aiming at the difficulties of three-dimensional imaging of turbulent flame,this paper innovatively proposes an imaging method of projection from two-dimensional low resolution to high-resolution three-dimensional structure without calibration,which has a better promoting significance for high-resolution three-dimensional combustion diagnosis.