汽车工程 ›› 2026, Vol. 48 ›› Issue (3): 638-650.doi: 10.19562/j.chinasae.qcgc.2026.ep.001

• • 上一篇    

融合CNN-BiLSTM与AUKF的车辆质心侧偏角估计

金琪1,赵治国1(),姜超2,周宇星2,赵坤1,夏雪1   

  1. 1.同济大学汽车学院,上海 201804
    2.上汽集团创新研究开发总院,上海 201804
  • 收稿日期:2025-05-20 修回日期:2025-07-09 出版日期:2026-03-25 发布日期:2026-03-19
  • 通讯作者: 赵治国 E-mail:zhiguozhao@tongji.edu.cn
  • 基金资助:
    国家自然科学基金(52172390)

Vehicle Sideslip Angle Estimation Based on CNN-BiLSTM and AUKF Fusion

Qi Jin1,Zhiguo Zhao1(),Chao Jiang2,Yuxing Zhou2,Kun Zhao1,Xue Xia1   

  1. 1.School of Automotive Studies,Tongji University,Shanghai 201804
    2.SAIC MOTOR R&D Innovation Headquarters,Shanghai 201804
  • Received:2025-05-20 Revised:2025-07-09 Online:2026-03-25 Published:2026-03-19
  • Contact: Zhiguo Zhao E-mail:zhiguozhao@tongji.edu.cn

摘要:

质心侧偏角是表征车辆运动稳定性的关键变量之一。现有估计方法多依赖高精度车辆动力学与轮胎模型,导致其在复杂非线性工况下的估计精度难以保证。为提升质心侧偏角估计的准确性与鲁棒性,本文提出一种基于卷积神经网络-双向长短时记忆网络(CNN-BiLSTM)与自适应噪声无迹卡尔曼滤波(AUKF)相融合的车辆质心侧偏角估计方法。首先,构建基于CNN-BiLSTM的分位数估计模型,结合车辆运动学模型与分位数回归损失函数,实现对质心侧偏角均值及分位区间的估计;其次,设计基于AUKF的动力学状态观测器,利用CNN-BiLSTM观测器的估计结果更新观测协方差矩阵,实现质心侧偏角的准确估计;最后,通过CarSim/Simulink联合仿真和实车试验对所提算法进行验证。结果表明,本文提出的CNN-BiLSTM与AUKF融合的车辆质心侧偏角估计算法,在不同工况与路面附着条件下估计结果准确,显著优于基于动力学模型的估计方法,具有较高估计精度和较强鲁棒性。

关键词: 质心侧偏角估计, 卷积神经网络, 双向长短时记忆神经网络, 自适应无迹卡尔曼滤波

Abstract:

The centroid sideslip angle is one of the key variables characterizing the stability of vehicle motion. Existing estimation methods mostly rely on high-precision vehicle dynamics and tire models, making it difficult to guarantee estimation accuracy under complex nonlinear operating conditions. To improve the accuracy and robustness of centroid side-slip angle estimation, this paper proposes a vehicle centroid side-slip angle estimation method that integrates the convolutional neural network-bidirectional long short-term memory network (CNN-BiLSTM) with the adaptive noise unscented Kalman filter (AUKF). Firstly, a quantile estimation model based on CNN-BiLSTM is constructed, combining the vehicle kinematic model and the quantile regression loss function to estimate the mean and quantile intervals of the centroid sideslip angle. Secondly, a dynamic state observer based on AUKF is designed, using the estimation results of the CNN-BiLSTM observer to update the observation covariance matrix, achieving accurate estimation of the centroid sideslip angle. Finally, the proposed algorithm is verified through CarSim/Simulink co-simulation and real vehicle tests. The results show that the CNN-BiLSTM and AUKF integrated vehicle centroid side-slip angle estimation algorithm proposed in this paper has accurate estimation results under different operating conditions and road adhesion conditions, significantly outperforming the estimation methods based on dynamic models, with high estimation accuracy and strong robustness.

Key words: sideslip angle estimation, CNN, BiLSTM, adaptive unscented Kalman filter