Wind noise performance prediction based on deep learning approach with vehicle exterior snapshot images

Abstract

Wind noise is one of the most important factor in vehicle development, and it is significantly influenced by the exterior design. The exterior design is changed many times throughout the development process, therefore it's very inefficient to make prototypes or to perform CFD simulations. Our CFD simulation accuracy has been improved over many years through validation studies and it's now the most trustful source however it has limitation on a long simulation time. In this research, a method to efficiently create a training data set to develop a CNN Deep learning model based on exterior images is proposed. First, CFD simulation has been performed several times with changing wind noise influence factors, and a meta-model is created based on these initial simulations. This meta-model creates various vehicle shapes, and calculates wind noise simulation results. After that, CNN DL model is created based on the images that has been created by the meta-model which best express the wind noise influence factor. This model promptly predicts the wind noise performances and verified through CFD simulation. Through this research, we were able to predict wind noise only with images, hence validated the possibility of general use that can be applied to various vehicles.