基于遗传算法的身管-弹丸接触碰撞网络模型优化研究

GENETIC ALGORITHM-BASED OPTIMIZATION OF THE BARREL-PROJECTILE CONTACT/IMPACT NEURAL NETWORK MODEL

火炮作为一种复杂多体系统,是陆军核心火力支援装备.弹炮间隙引起的身管-弹丸接触碰撞现象,直接影响火炮系统射击精度.针对身管-弹丸复杂表面接触碰撞问题,传统的知识驱动建模策略难以较好地兼顾建模精度与预测效率.文章结合有限元数值模拟与人工神经网络方法,建立身管-弹丸接触碰撞神经网络模型,并开展模型综合性能评估:(1)建立身管内膛与弹丸前定心部接触碰撞有限元弹塑性动力学模型,分析不同初始碰撞速度下各物理量变化规律,构建网络模型训练所需多工况样本集;(2)引入多样性评价指标,为全面综合的接触碰撞模型性能评估奠定基础;(3)考虑神经网络超参数设置对所建模型预测精度及稳定性的重要影响,采用遗传算法开展合理超参数设置研究,重点关注遗传算法相关参数设置对所建模型预测性能的影响.仿真结果表明:随着超参数优化方案的改进,所建身管-弹丸复杂表面接触碰撞神经网络模型的预测精度和稳定性显著提升,进一步验证了研究策略的有效性.

As a complex multibody system,artillery serves as the core firepower support equipment of the army.The contact/impact phenomenon arising from the clearance between the barrel and projectile,directly impacts the firing accuracy of the artillery and holds significant research importance.Traditional knowledge-driven modeling methods struggle to balance the accuracy and efficiency when handling the complex surface contact/impact process between the barrel and projectile.This work proposes the integration of the artificial neural network technique with the finite element simulation to develop a neural network-based contact force model.A comprehensive performance evaluation of the established model used for the barrel-projectile contact/impact process will be conducted.Initially,a finite element elastic-plastic dynamics model will be established to analyze the interaction process between the barrel and projectile.Changes of physical quantities under different initial contacting velocities will be analyzed and employed to build a multi-condition sample set for subsequent network training process.After that,a range of evaluation indicators will be introduced to facilitate a thorough and unified assessment of the contact force model,including mean square error(MSE),relative square error(RSE),relative absolute error(RAE),symmetric mean absolute percentage error(sMAPE)and determination coefficient(R2).Considering the significant effects of network hyperparameter settings on the prediction accuracy and stability of the model,the genetic algorithm(GA)will be utilized to obtain optimal hyperparameter settings,focusing on the influence of relevant GA parameters on the prediction performance of the established contact force model.Simulation results demonstrate that with the improved hyperparameter optimization strategy,the prediction accuracy and stability of the proposed neural-network-based contact force model for the interaction process between the barrel and projectile are significantly improved,which further verifies the great necessity of research efforts towards refining hyperparameter settings to enhance the overall performance and generalization ability of the established model.