Scholarlink

The development of intelligent vehicles has accelerated the deployment of autonomous driving and advanced driver-assistance systems, raising stringent requirements for accurate and real-time state estimation. Beyond conventional kinematic parameters, vehicle mass and road slopes—both longitudinal and lateral—are essential for active braking, suspension regulation, and motion control, as they directly affect driving safety and ride comfort. However, these states are strongly coupled with pitch and roll dynamics, are sensitive to road and load variations, and remain difficult to measure directly. Traditional estimation approaches often encounter challenges such as significant model inaccuracies and slow dynamic response. This paper presents a hybrid estimation method that integrates a square root unscented Kalman filter with a lightweight neural network based on the Transformer architecture. A cross-attention mechanism is introduced to adaptively fuse the residuals of the physical filter and the neural model, combining physical consistency with nonlinear learning capacity. To support real-time deployment, the fused model is further compressed into a compact recurrent neural network with memory capability through knowledge distillation. Simulation results demonstrate that the proposed approach improves estimation accuracy for vehicle mass and road slopes by more than 25% compared to traditional filters, while maintaining low latency and high deployment feasibility.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Cross-Modal Fusion of Physical Filtering and Transformer-Based Neural Networks for Vehicle Mass and Road Gradient Estimation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hou Tianyu,
  • Zhou Shiji,
  • Rao Shengren

摘要

The development of intelligent vehicles has accelerated the deployment of autonomous driving and advanced driver-assistance systems, raising stringent requirements for accurate and real-time state estimation. Beyond conventional kinematic parameters, vehicle mass and road slopes—both longitudinal and lateral—are essential for active braking, suspension regulation, and motion control, as they directly affect driving safety and ride comfort. However, these states are strongly coupled with pitch and roll dynamics, are sensitive to road and load variations, and remain difficult to measure directly. Traditional estimation approaches often encounter challenges such as significant model inaccuracies and slow dynamic response. This paper presents a hybrid estimation method that integrates a square root unscented Kalman filter with a lightweight neural network based on the Transformer architecture. A cross-attention mechanism is introduced to adaptively fuse the residuals of the physical filter and the neural model, combining physical consistency with nonlinear learning capacity. To support real-time deployment, the fused model is further compressed into a compact recurrent neural network with memory capability through knowledge distillation. Simulation results demonstrate that the proposed approach improves estimation accuracy for vehicle mass and road slopes by more than 25% compared to traditional filters, while maintaining low latency and high deployment feasibility.