基于卷积神经网络-长短时记忆(convolutional neural network and long short-term memory, CNN-LSTM)模型开展了曲线变量的假人伤害预测应用研究. 分别使用标定后1D简化模型和约束系统CAE模型开展变量和响应采集,搭建训练和测试数据库,使用Pytorch搭建假人头部加速度和胸部压缩量CNN-LSTM预测模型,研究了样本数量对训练后模型精度的影响. 结果显示,当训练样本数量达到一定规模后,继续增加训练样本数量对模型泛化能力提升有限,然而,在本文应用场景中,当样本数量到50时,测试样本预测精度R均值超过0.85,满足工程开发的预测精度要求.
Abstract
The convolutional neural network and long short-term memory(CNN-LSTM) model were utilized to carry out the application study of the dummy injury prediction with the curve variable input. To construct the training and testing dataset, the correlated 1D simplified car model and restraint system CAE model were introduced herein to collect the input and output responses, separately. Besides, Pytorch was used to construct both the dummy head acceleration and chest deflection CNN-LSTM prediction models, and the effect of sample number on the precision of the training model was further studied. The results indicate that the increase of training sample number could not further enhance the model performance once the training datasets reach a certain size. However, the R value of the predicted results could be over 0.85 with the training sample number coming to 50, which can satisfy the predictive accuracy requirements of engineering development.
随着人工智能技术的快速发展,汽车主机厂研发环节均开启了人工智能化进程,在团队、软件、算力、流程等方面增加了投入,以期借助人工智能技术提升产品研发效率. 近些年,人工智能技术在汽车安全领域中的应用研究也逐渐增多[2-4],尤其是在汽车约束系统性能领域,如张绍伟等[5]将长短时记忆网络模型应用于鞭打性能预测中,取得了较好的预测效果. 高伟钊等[6]将时域降阶模型算法应用于约束系统鲁棒性分析,使用降阶模型(reduced order model, ROM)代替CAE模型,实现对假人伤害曲线的有效预测. 黄泽辉等[7]研究了3种深度学习算法在乘员伤害评估中的预测精度,为该领域的深度学习算法选型提供了参考.
针对车型研发中的约束系统假人伤害预测问题,提出一种基于1D简化模型、约束系统CAE模型和深度学习算法相结合的全新方法,基于卷积神经网络-长短时记忆(convolutional neural network and long short-term memory, CNN-LSTM)算法,研究假人头部加速度和胸部压缩量曲线预测中的工程应用.
WEIG, YUC W, YANGW L,et al .Rapid development of facelift vehicle model based on virtual fatigue analysis[J]. Automobile Technology,2017(12): 13-17.(in Chinese)
[3]
HASIJAV, TAKHOUNTSE G .Deep learning methodology for predicting time history of head angular kinematics from simulated crash videos[J].Scientific Reports,2022,12(1): 6526.
[4]
TRILLINGJ, SCHUMACHERA, ZHOUM .Reinforcement learning based agents for improving layouts of automotive crash structures[J].Applied Intelligence,2024,54(2):1751-1769.
[5]
GUC W, XUJ L, LIS Q,et al. Injury risk assessment and interpretation for roadway crashes based on pre-crash indicators and machine learning methods[J]. Applied Sciences, 2023, 13(12): 6983.
ZHANGS W, ZHUD W, ZHAIG Z. Prediction on seat’s anti-whiplash-injury performance based on deep learning[J]. Automotive Engineering,2022,44(10):1600-1608.(in Chinese)
GAOW Z, ZENGX J, WANGD H,et al. Research on robustness analysis method of automobile restraint system based on ODYSSEE[J]. Intelligent Manufacturing,2023(Sup.1): 110-115.(in Chinese)
HUANGZ H, TANGH B, WANGX S,et al. Research on occupant injury prediction method in vehicle collision based on deep learning[J]. Automotive Engineer, 2024(4): 8-11.(in Chinese)
China New Car Assessment Program (C-NCAP) Management Rules 2024 Appendix A Regulation of Normal 100% Overlap Rigid Barrier Crash Test [S]. Tianjin: China Automotive Technology & Research Center Limited Company, 2024: 26-27.(in Chinese)
[14]
YASUKIT. Application of reduced model to simulations of occupant protection and crashworthiness[C]//The 15th International LS-DYNA® Conference & Users Meeting. Detroit, USA: Livermore Software Technology Corporation (LSTC), 2018.
[15]
HOCHREITERS, SCHMIDHUBERJ. Long short-term memory[J]. Neural Computation, 1997, 9(8):1735-1780.
LINX, ZHANGH, MAY L, et al. Electric vehicle charging load prediction based on improved LSTM neural network[J]. Modern Electronics Technique, 2024, 47(6): 97-101.(in Chinese)
[18]
王玉堃 .LSTM和GRU神经网络在股指高频数据预测中的研究[D].济南:山东大学,2020.
[19]
WANGY K. Research on prediction of high frequency data of stock index based on LSTM and GRU neural network[D]. Jinan:Shandong University, 2020.(in Chinese)
FANGP J, CAIY F, CHENL,et al. Neural network lateral dynamics modeling and control based on ED-LSTM for intelligent vehicle[J]. Chinese Journal of Theoretical and Applied Mechanics, 2022, 54(7): 1896-1908.(in Chinese)
LIUJ, LEIH, LÜJ,et al .SOC estimation of lithium-ion batteries based on CNN-LSTM[J].Electric Drive,2024,54(2):26-31.(in Chinese)
[24]
SUBRAMANIANV. Deep learning with PyTorch:a practical approach to building neural network models using PyTorch[M]. Birmingham: Packt Publishing,2018: 1-262.