汽车工程 ›› 2026, Vol. 48 ›› Issue (3): 663-675.doi: 10.19562/j.chinasae.qcgc.2026.03.016

• • 上一篇    

基于CNN-Transformer的车辆侧倾动力学建模及实验验证

曹守启1,高雅琪1,周国峰1(),陈渐伟2,周志松3,姜加胜1   

  1. 1.上海海洋大学工程学院,上海 201306
    2.火箭军工程大学导弹工程学院,西安 710025
    3.香港中文大学天石机器人研究所,香港
  • 收稿日期:2025-03-14 修回日期:2025-06-09 出版日期:2026-03-25 发布日期:2026-03-19
  • 通讯作者: 周国峰 E-mail:gfzhou@shou.edu.cn
  • 基金资助:
    上海市农业科技创新项目(沪农科I2023006)资助。

Modeling and Experimental Validation of Vehicle Roll Dynamics Based on CNN-Transformer

Shouqi Cao1,Yaqi Gao1,Guofeng Zhou1(),Jianwei Chen2,Zhisong Zhou3,Jiasheng Jiang1   

  1. 1.School of Engineering,Shanghai Ocean University,Shanghai 201306
    2.College of Missile Engineering,Rocket Force University of Engineering,Xi'an 710025
    3.T Stone Robotics Institute,The Chinese University of Hong Kong,Hong Kong
  • Received:2025-03-14 Revised:2025-06-09 Online:2026-03-25 Published:2026-03-19
  • Contact: Guofeng Zhou E-mail:gfzhou@shou.edu.cn

摘要:

高精度车辆侧倾动力学建模对提升车辆主动安全控制性能至关重要。然而,车辆系统具有强非线性与参数不确定性,基于传统机理分析的建模方法难以准确估计侧倾角和侧倾角速度。针对此问题,本文提出一种融合多尺度卷积神经网络与Transformer的数据建模方法。该模型利用多尺度卷积核提取含噪声数据中的多频域特征,提高模型对噪声干扰的鲁棒性;同时,结合Transformer的注意力机制,有效捕捉侧倾动力学中的长时序依赖关系,进一步提升建模精度。为验证模型性能,本研究基于CarSim高保真仿真平台和实车道路实验数据,将其与传统Transformer、LSTM、GRU等数据驱动模型以及物理模型进行对比分析。实验结果表明,所提出的CNN-Transformer混合模型在侧倾角和侧倾角速度预测任务中表现最优,预测决定系数R2 均高于0.974 5,实现了对车辆侧倾动力学的准确建模。

关键词: 多尺度特征提取, Transformer, 多头注意力机制, 侧倾动力学

Abstract:

High-precision vehicle roll dynamics modeling is crucial for enhancing the performance of active safety control systems. However, the vehicle system is characterized by strong nonlinearity and parameter uncertainty, making it difficult for traditional mechanism-based modeling methods to accurately estimate roll angle and roll angle velocity. To address this issue, this paper proposes a data modeling method that integrates multi-scale convolutional neural networks with Transformer. This model utilizes multi-scale convolutional kernels to extract multi-frequency domain features from noisy data, enhancing the model's robustness against noise interference. Meanwhile, by combining the attention mechanism of Transformer, it effectively captures the long-term temporal dependencies in roll dynamics, further improving the modeling accuracy. To verify the model's performance, this study conducts a comparative analysis with traditional data-driven models such as Transformer, LSTM, GRU, and physical models based on high-fidelity CarSim simulation platform and real vehicle road test data. The experimental results show that the proposed CNN-Transformer hybrid model performs optimally in the tasks of roll angle and roll angle velocity prediction, with prediction determination coefficients R2 all above 0.974 5, achieving accurate modeling of vehicle roll dynamics.

Key words: multi-scale feature extraction, Transformer, multi-head attention mechanism, rollover dynamics