Research on rotation cycle feature extraction and vertical force estimation for smart tires

Abstract

To address the problem of real-time tire vertical force estimation, this paper proposes a method based on radial acceleration signals acquired from smart tire sensors. An adaptive negative peak detection algorithm is first developed to identify the contact minimum, which serves as a reference for cycle segmentation, enabling adaptive determination of the tire rotation cycle without requiring a rotational speed sensor. Based on this procedure, five low-dimensional features are constructed for each rotation cycle, including the cycle start value, leading edge value, trailing edge value, minimum value, and contact duration. These features effectively characterize the dynamic behavior of the tire under varying load and vehicle speed conditions. A dataset is then established using these features, and the estimation performance of three neural network models—LSTM, GRU, and MLP—is comparatively evaluated. Experimental results from bench tests demonstrate that LSTM and GRU achieve comparable accuracy and robustness in capturing the nonlinear relationship between the extracted features and vertical force, while MLP exhibits significantly larger errors. Considering model complexity and computational efficiency, the GRU model is selected as the final solution. The proposed method enables accurate estimation using only five features per cycle, significantly reducing storage and computational requirements while maintaining strong adaptability across operating conditions.