Real-time Prediction Of Vehicle Aerodynamic Performance Based On 3d Deep Learning

Vehicle aerodynamic layout design is an important part of vehicle design.The aerodynamic performance of the car can be improved by reducing the drag coefficient,which is of great significance for Energy conservation and emissions reduction of the fuel car and improvement of the endurance mileage of the electric car.At present,the calculation of vehicle aerodynamic shape key parameters still depends on the traditional fluid calculation method.When the number of grid and DOF increases,the calculation time will increase exponentially,which cannot meet the requirements of modern fast car design.The purpose of this paper is to develop a real-time vehicle prediction method for aerodynamic parameters based on 3d deep learning.This method is simple in calculation,does not require complicated calculations,and has good real-time performance.Combining deep learning with CFD can solve the problem of slow calculation time in simulation.The main work of this article is as follows:(1)Construct a vehicle aerodynamic simulation data set.Based on the XFlow software,the car model in the ShapeNet 3D model data set was placed in five 5 virtual wind tunnels with different wind speeds for simulation,and analyze the simulation results.A date set of vehicle aerodynamic simulation labeled vehicle aerodynamic drag is constructed using the simulation data.The data set contains 400 vehicle aerodynamic simulation calculation results.(2)Real-time aerodynamic parameter prediction based on O-CNN.The O-CNN neural network is used to store 3D model of the data set as the octree structure.Train three different levels of octree structure,modify the network model and parameters,and train the aerodynamic simulation data set to enable O-CNN to predict the aerodynamic drag coefficient in real time.(3)Training a 3D neural network model which can predict the key parameters of vehicle aerodynamic performance in real time.According to the trained model for verification and testing,the results show,the performance of 5 layers octree structure is the best,the R-Square coefficient is 0.8016,and the 4 layers octree structure has the fastest neural network running time.It can achieve real-time output of aerodynamic parameters,and the efficiency is much higher than the traditional CFD numerical simulation method.In summary,we have developed a 3D deep learning real-time vehicle aerodynamic parameter prediction method,which can predict key aerodynamic parameters such as vehicle drag coefficient in real time.This method can be used not only for the prediction of key parameters of automotive aerodynamic shape,but also for other related fields.It has important application value for rapid vehicle design and shortening the vehicle development cycle.