On data-driven optimization design of protective structures for vehicles against explosion

In order to address the needs of modern combat vehicles for both personnel protection and lightweight design, optimizing their blast-resistant structures is necessary. Due to the high cost of physical experiments, finite element simulation has been commonly used instead. However, simulations of explosion and vehicle responses require extensive computational resources and incur high computational costs, leading to limited data availability for the optimization of explosion-proof structures. Since structural optimization demands sufficient data support, larger amount of valid data can improve the accuracy of the surrogate model and the precision of the optimal solution, yielding better optimization results. To overcome these challenges, a data-driven optimization method for vehicle’s explosion-proof structures was proposed, integrating data augmentation and semi-supervised regression. To address the limitations of generative adversarial networks (GANs) in handling numerical data, an improved model, a Gaussian density estimation-Wasserstein generative adversarial network (GDE-WGAN), was developed by modifying both the generator and discriminator of the WGAN model, a variant of the GANs. The feasibility of the proposed method was demonstrated based on the principle of information gain. The data generated by the GDE-WGAN were incorporated into a self-training framework, where an adaptive confidence assessment mechanism dynamically adjusted the way that the semi-supervised support vector regression model utilizes the generated data. The feasibility and superiority of the method were validated by comparing the enhanced performance of the semi-supervised regression model using different numerical data expansion techniques. Finally, multi-objective optimization was performed to obtain the optimal solutions of the data-augmented semi-supervised regression model and the initial model, followed by verification and comparison with finite element simulation results. It shows that the GDE-WGAN significantly enhances the performance of the semi-supervised regression model, and the generated data exhibit greater randomness and diversity through the network structure of the GANs, which benefits semi-supervised learning. When handling semi-supervised regression for high-dimensional nonlinear numerical data, both global and local data distribution similarities play a crucial role. Furthermore, finite element simulations indicate that the improved model predicts results more accurately than the initial model and achieves superior optimization outcomes.