Research on optimization design of vehicle explosion protection structure based on data drive

In order to balance the need for personnel protection and lightweight of modern combat vehicles, it is necessary to optimize its explosion protection structure. Due to the high cost of experiments, finite element simulation is usually used instead. However, vehicle explosion simulation requires a lot of computational resources and high computational costs, resulting in limited data sources for vehicle explosion protection structure optimization. Structural optimization requires sufficient data support. The larger the amount of data, the higher the accuracy of the approximate proxy model, the more accurate the final optimal solution, and the better the optimization effect. Therefore, a data-driven method is proposed to optimize the vehicle explosion protection structure. According to the data characteristics, the adversarial generating network (GAN) is improved, and the GDE-WGAN method is proposed, which is combined with semi-supervised support vector regression based on the self-training framework to expand the original data set. Meanwhile, the feasibility and superiority of this method are verified by comparing the performance improvement of different numerical data augmentation methods on the semi-supervised regression model. Finally, the optimal solutions of the data augmentation combined with semi-supervised regression model and the initial model were obtained by multi-objective optimization, and verified and compared by finite element simulation. The results show that GDE-WGAN method has a more significant effect on the performance improvement of semi-supervised regression model, and the generated data is more random and diverse through the network structure of GAN, which is beneficial to semi-supervised learning. When dealing with semi-supervised regression of high-dimensional nonlinear numerical data, not only the similarity of global data distribution is crucial, but also the similarity of local data, especially the distance between unlabeled samples and labeled samples. Through the finite element simulation, it is found that the improved model can predict the result more accurately and show better optimization effect than the original model.