3D Reconstruction Of Bubble Flow Field Based On Improved Convolutional Neural Network

Gas-liquid two-phase flow,one of the common forms of multiphase flow.It is widely used in coal transportation,oil mining,crude oil transportation,industrial sewage and pneumatic transportation,which is closely related to human production and life and practical activities.Therefore,the study of liquid two-phase flow is of great significance.In the research work of gas-liquid two-phase flow,the measurement technology about it is the main research technology,this technology aims to measure the parameters of liquid or bubbles in the liquid phase.However,existing bubble parameter measurement methods have problems such as high measurement cost,poor real-time performance,and complex measurement systems and measurement algorithms.In response to these problems,this paper takes the bubbles in the liquid phase as the research object,and proposes a three-dimensional reconstruction method of the bubble flow field based on an improved convolutional neural network.The movement of the gas phase in the liquid phase is analyzed by reconstructing the threedimensional information of the bubble flow field.The relevant research carried out in this paper is shown below.Aiming at the problem of bubble flow field reconstruction in the study of gasliquid two-phase flow,this paper designs a bubble image acquisition system based on a light field camera to obtain light field images of sparse bubbles in water.And through the light field image calibration experiment,the query dictionary of the focused depth of the refocused image is obtained.In terms of image preprocessing,a bubble segmentation algorithm based on dynamic target positioning and a bubble image size standardization method based on edge growth are proposed to solve the difficulty of automatic segmentation of bubble images caused by the unfixed absolute position of the bubble in the refocusing image sequence.And the problem of label inconsistency and the problem of bubble feature changes caused by image size standardization.In order to realize the calculation of bubble depth information,this paper constructs a bubble depth prediction model based on a convolutional neural network,which is based on the LeNet5 classification model and the DIF-LeNet depth prediction model.The LeNet5 classification model mainly recognizes the "focused" image of the bubble,and then obtains the bubble depth according to the query dictionary.The DIFLeNet prediction model is an improved dual-information fusion convolutional neural network model,which realizes the feature fusion of different dimensional data and the regression prediction of the target quantity.The DIF-LeNet depth prediction model can predict the depth of the bubble only through a single refocused bubble image and the focused depth of the image,thereby improving the calculation speed and accuracy of bubble depth prediction.Combine the predicted depth with the center coordinates of two-dimensional and radius based on bubbles to reconstruct the bubble's three-dimensional model.The experimental results show that the DIF-LeNet depth prediction model based 3D reconstruction method can improve the accuracy by 0.3% and speed by 9.73 s/piece compared with the statistical reconstruction method;the DIF-LeNet depth prediction model based 3D reconstruction method can improve the accuracy by 5.4% and speed by 3.17 s/piece compared with the LeNet5 classification model.For the bubble without focused image,the DIF-LeNet depth prediction model can improve the 3D reconstruction accuracy and much above the other two methods.