基于机器学习的汽车吸能结构耐撞性智能预测方法

Machine Learning Method for Intelligent Prediction of the Crashworthiness of Automotive Energy Absorbing Boxes

摘要:

汽车零部件正向设计中，为快速预测所设计的吸能结构的碰撞吸能特性，以吸能盒为研究对象，通过有限元压溃变形仿真生成数据集，训练得到一种新的可识别几何结构和记忆时序特征的预测模型。模型通过基于图的编码器进行几何结构识别，采用长短期记忆网络和图卷积神经网络处理时序数据，并输出预测结果。对比表明：吸能盒压溃形态预测结果与有限元仿真结果一致，压溃变形量的预测精度可达95.33%，最大吸能值的预测精度可达99.98%。预测模型相较于有限元计算，其计算效率分别提高了174.5倍和210.5倍，可以快速准确地预测吸能盒的碰撞性能。

关键词:汽车吸能盒;耐碰撞性;有限元分析;机器学习;图卷积神经网络;长短期记忆神经网络

Abstract:

This study aims to achieve intelligent prediction of collision energy absorption characteristics of new structures in forward design of automotive parts. An energy-absorbing box is taken as the research object to generate training data sets by finite element crush deformation simulation. A graph-based encoder is adopted for geometric structure recognition. Long and short-term memory networks and graph convolutional neural networks were used to process adjacent temporal data. The novel neural network prediction system proposed can recognize geometric structures and memorize temporal data. The comparison between the model prediction results and simulation results shows that the predicted crush pattern of the energy-absorbing box is consistent with the finite element simulation results， and the prediction accuracy of the model for the crush deformation amount can reach up to 95.33%， while the prediction accuracy of the maximum energy absorption value can reach 99.98%. Compared with the finite element calculations， computational efficiency is 174.5 times and 210.5 times higher respectively， which manifested that the system can accurately and quickly predict the crash performance of the energy-absorbing box.

Key words:energy-absorbing box;crashworthiness;finite element simulation;machine learning;graph convolution networks;long short-term memory