基于机器学习的汽车侧气帘性能预测方法

doi:10.3969/j.issn.1674-8484.2026.03.003

摘要/Abstract

摘要：

为解决汽车零部件制造企业在车型研发过程中，侧气帘性能仿真计算迭代次数多、耗时长的问题，提出基于机器学习的汽车侧气帘性能预测方法，分别对侧气帘充气体积和展开形状的关键性能进行建模。该文依托企业积累的大量仿真数据，提出了适用于实际工况的专用图像特征提取方法，通过比选采用极端梯度提升（XGBoost）算法进行充气体积预测，通过比选采用基于残差U形神经网络（ResUNet）的双头输出神经网络实现展开形状预测。研究表明：侧气帘充气体积预测模型在测试集上误差小于5%的样本占比达98.2%，展开形状预测模型交并比（IoU）大于0.8的样本占比达93.3%，均满足工程应用精度要求。与传统仿真方法相比，AI模型将侧气帘性能预测时间从2天缩短至秒级，显著提升了性能预测的迭代效率，为汽车安全部件研发提供了高效、可行的技术路径，对企业数字化转型具有重要的实践价值。

关键词: 汽车安全装置, 汽车安全气帘, 机器学习, 性能预测

Abstract:

To address excessive iterations and long computation time in the performance simulation of automotive side curtain airbags during vehicle development, a machine learning-based prediction approach was proposed to model two key performance indicators: inflation volume and deployment shape. Leveraging a large repository of enterprise simulation data, a domain knowledge-driven image feature extraction method was proposed, and the XGBoost algorithm was selected for inflation volume prediction through comparative evaluation. Meanwhile, a dual-head neural network based on ResUNet was adopted for deployment shape prediction. The results show that 98.2% of the test samples for the inflation volume prediction model have an error of less than 5%, and 93.3% of the samples for the deployment shape prediction model achieve an intersection-over-union (IoU) greater than 0.8, both meeting engineering accuracy requirements. Compared with traditional simulation methods, the proposed AI models reduce prediction time from two days to the order of seconds, significantly improving iteration efficiency. An effective and feasible technical pathway is provided for automotive safety component development, offering substantial practical value for enterprise digital transformation.

Key words: automotive safety systems, automotive safety airbag, machine learning, performance prediction

中图分类号:

叶晔, 刘二勇, 陈翼雄. 基于机器学习的汽车侧气帘性能预测方法[J]. 汽车安全与节能学报, 2026, 17(3): 314-321.

YE Ye, LIU Eryong, CHEN Yixiong. Machine learning-based method for predicting the performance of automotive side curtain airbags[J]. Journal of Automotive Safety and Energy, 2026, 17(3): 314-321.

图/表 8

图1 仿真气袋模型形状

AI模型	RMSE	MAE	R²	MAPE / %
XGBoost	1.029 8	0.595 4	0.996 7	1.22
LightGBM	1.093 8	0.744 0	0.996 2	1.53
Random Forest	1.174 9	0.613 9	0.995 7	1.24
SVR	1.559 0	0.652 6	0.992 4	1.37
MLP	1.657 2	0.723 3	0.991 4	1.59

表1 不同AI模型侧气帘充气体积预测精度

图2 不同AI模型侧气帘充气体积预测精度

图3 基于自动编码器的神经网络结构

图4 基于ResUNet的神经网络框架和结构

图5 不同IoU对应的展开形状预测结果

AI 模型	平均值	标准差	预测精度占比/ %
AI 模型	平均值	标准差	< 0.7	0.7~0.8	0.8~0.9	> 0.9
自动编码	0.824 5	0.057 5	1.58	35.70	55.62	7.10
ResUNet- 双头位移预测	0.855 2	0.043 9	0.20	9.86	75.15	14.79
ResUNet- 双头生成预测	0.862 3	0.039 6	0.20	6.51	74.56	18.74
ResUNet- 单头位移预测	0.859 7	0.040 6	0.59	9.07	80.28	10.06
ResUNet- 单头生成预测	0.859 9	0.037 9	0.59	6.11	85.21	8.09

表2 不同AI模型侧气帘展开形状预测精度(采用IoU值评价)

图6 不同模型预测结果对比 (a)设计输入；(b)仿真真值；(c)自动编码器网络预测结果；(d) ResUNet双头模型预测的位移场；(e) ResUNet双头模型：位移预测头；(f) ResUNet 双头模型：生成预测头；(g) ResUNet单头模型(仅位移)；(h) ResUNet单头模型(仅生成)

参考文献 23

[1]	侯志平, 朱海涛, 刘灿灿, 等. 基于自适应自编码器假人力学响应降维和重构方法[J]. 汽车安全与节能学报, 2024, 15(3): 337-343.
	HOU Zhiping, ZHU Haitao, LIU Cancan, et al. Dimension-ality reduction and reconstruction method of dummy biomechanics response based on adaptive autoencoder[J]. J Autom Safe Energ, 2024, 15(3): 337-343. (in Chinese)
[2]	荣海, 蒋建中, 姚再起, 等. 基于神经网络代理模型的车身多学科轻量化优化设计[J]. 汽车安全与节能学报, 2025, 16(1): 32-42.
	RONG Hai, JIANG Jianzhong, YAO Zaiqi, et al. Multidis-ciplinary design optimization of a vehicle body for light-weight based on a neural network surrogate model[J]. J Autom Safe Energ, 2025, 16(1): 32-42. (in Chinese)
[3]	Ghojogh B, Samad M N, Mashhadi S A, et al. Feature selection and feature extraction in pattern analysis: A literature review[J/OL]. arXiv (CoRR). (2019-05-07) https://arxiv.org/abs/1905.02845.
[4]	Le T H, Bui L. Face recognition based on SVM and 2DPCA[J/OL]. arXiv (CoRR). (2011-10-25) https://arxiv.org/abs/1110.5404.
[5]	Athiwaratkun B, Kang K. Feature representation in convol-utional neural networks[J/OL]. arXiv(CoRR). (2015-07-08) https://arxiv.org/abs/1507.02313.
[6]	Arjomandi R M, Salomonsson K, Cenanovic M, et al. Correlation-based feature extraction from computer-aided design, case study on curtain airbags design[J]. Comput Indu, 2022, 138: 103634.
[7]	Arjomandi R M, Cenanovic M, Salomonsson K. Image regression-based digital qualification for simulation-driven design processes, case study on curtain airbag[J]. J Engi Desi, 2023, 34(1): 1-22.
[8]	LeCun Y, Boser B, Denker J S, et al. Backpropagation applied to handwritten zip code recognition[J]. Neural Comput, 1989, 1(4): 541-551.
[9]	Goodfellow I J, Pouget-Abadie J, Mirza M, et al. Genera-tive adversarial nets[J]. Advan Neural Info Proc Syst, 2014, 27: 2672-2680.
[10]	Diakogiannis F I, Waldner F, Caccetta P, et al. ResUNet-a: A deep learning framework for semantic segmentation of remotely sensed data[J]. ISPRS J Photo Remo Sens, 2020, 162: 94-114.
[11]	Berahmand K, Daneshfar F, Salehi E S, et al. Autoencoders and their applications in machine learning: a survey[J]. Artif Intel Rev, 2024, 57(2): 28.
[12]	Isola P, ZHU Junyan, ZHOU Tinghui, et al. Image-to-image translation with conditional adversarial networks[J/OL]. arXiv (CoRR). (2016-12-21) https://arxiv.org/abs/1611.07004.
[13]	ZHU Junyan, Park T, Isola P, et al. Unpaired image-to-image translation using cycle-consistent adversarial networks[C]// Proc IEEE Int’l Conf Comput Visi, Venice, Italy, 2017: 2223-2232.
[14]	Kim S, Kreis K, Li Daqing, et al. Polymorphic-GAN: Generating aligned samples across multiple domains with learned morph maps[J/OL]. arXiv (CoRR). (2022-06-06) https://arxiv.org/abs/2206.02903
[15]	Balakrishnan G, ZHAO Amy, Sabuncu M R, et al. Voxel-morph: a learning framework for deformable medical image registration[J]. IEEE Trans Medi Imag, 2019, 38(8): 1788-1800.
[16]	刘卫朋, 李旭, 任子文, 等. 多尺度残差可变形肺部CT图像配准算法[J]. 华南理工大学学报(自然科学版), 2024(10): 135-145.
	LIU Weipeng, LI Xu, REN Ziwen, et al. Algorithm for multiscale residual deformable lung CT image registration[J]. J South China Univ Tech (Nat Sci Edit), 2024(10): 135-145. (in Chinese)
[17]	Del Río J E, Valenzuela R, Lancashire R, et al. Machine-learning-accelerated simulations for the design of airbag constrained by obstacles at rest[J]. Stap Car Crash J, 2024, 67(1): 1-13.
[18]	中国汽车技术研究中心有限公司. [S]. 天津: 中国汽车技术研究中心有限公司, 2023.
	China Automotive Technology and Research Center Co., Ltd. [S]. Tianjin: China Automotive Technology and Research Center Co., Ltd, 2023. (in Chinese)
[19]	Martin S. The k-d tree data structure, a proof for neighbor-hood computation in expected logarithmic time[J/OL]. arXiv (CoRR). (2019-03-12) https://arxiv.org/abs/1903.04936
[20]	ZHAO Hang, Gallo O, Frosio I, et al. Loss functions for image restoration with neural networks[J]. IEEE Trans Comput Imag, 2017, 3(1): 47-57.
[21]	Akiba T, Sano S, Yanase T, et al. Optuna: a next-generation hyperparameter optimization framework[J/OL]. arXiv (CoRR). (2019-07-25) https://arxiv.org/abs/1907.10902.
[22]	Sorensen T. A method of establishing groups of equal amplitude in plant sociology based on similarity of species content and its application to analyses of the vegetation on Danish commons[J]. Biologiske Skrifter, 1948, 5: 1-34.
[23]	Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[J/OL]. arXiv (CoRR). (2014-12-14) https://arxiv.org/abs/1411.4038.