基于多重滤波联合观测的混动拖拉机轮土模型在线辨识

Online Identification of Hybrid Tractor Tire-soil Interaction Model Based on Multiple Filtering Joint Observation

摘要:

精准的轮土模型是优化混动拖拉机牵引力控制进而提升综合能效的基础，其中，精确的纵、垂向力估计是保证模型辨识精度的关键。然而，作业时地形变化、犁体与土壤互作的未知负载变化造成拖拉机各轮的纵、垂向力波动剧烈，增加了准确估计的难度。为此，以分布式混合动力拖拉机(DHET)为研究对象，提出了一种融合多重滤波联合观测的轮胎-土壤模型在线辨识方法。首先，利用卡尔曼滤波(KF)估算轮胎的纵向力，结合动力学模型采用容积卡尔曼滤波(CKF)算法估算轮胎的垂向力；然后，结合Brixius模型和整车纵向动力学，构建基于无迹粒子滤波(UPF)的轮土模型在线辨识器，通过无迹变换(UT)设计粒子提议分布，保留非线性轮土模型中的高阶信息，并使用粒子滤波(PF)算法对待估参数进行更新，完成对模型的在线辨识。硬件在环(HIL)试验结果表明，该方法能够在机动指数变化的路面条件下，快速且准确地实现辨识，移动系数均方根估计误差不超过1.2。试验结果显示，在两种地况下，辨识模型边界牵引系数误差在±0.015范围内，前轮试验数据分布占比达到84.45%和88.16%，后轮分别为86.72%和85.38%，验证了该方法在不同地况下均具备较高精度和鲁棒性。研究结果为拖拉机最优驱动力分配策略提供了理论支持，并有助于提升牵引效率和燃油经济性。

Abstract:

Accurate tire-soil models are fundamental to optimizing the traction force control of hybrid tractors, thereby improving overall energy efficiency. Among these, precise estimation of the longitudinal and vertical forces is crucial for ensuring model identification accuracy. However, during operation, terrain variations and the unknown load changes caused by the interaction between the plow body and the soil lead to significant fluctuations in the longitudinal and vertical forces on each tire, which increases the difficulty of accurate estimation. To address this, a tire-soil interaction model identification method based on the fusion of multiple filters and joint observation was proposed, taking the distributed hybrid electric tractor (DHET) as research object. Firstly, the longitudinal force of the tire was estimated by using the Kalman filter (KF), and the vertical force was estimated by using the cubature Kalman filter (CKF) algorithm combined with the dynamic model. Then, based on the Brixius model and the vehicle’s longitudinal dynamics, an online identification device for the tire-soil model was constructed by using the unscented particle filter (UPF). The unscented transformation (UT) was used to design the particle proposal distribution, preserving higher-order information in the nonlinear tire-soil model, and the particle filter (PF) algorithm was applied to update the estimated parameters, completing the model’s online identification. Hardware-in-the-loop (HIL) test results showed that the method can quickly and accurately achieve identification under varying terrain conditions, with a root mean square error not exceeding 1.2. Experimental results indicated that, under two types of soil conditions, the identification model’s boundary traction coefficient error was within ±0015, with the front wheel test data distribution accounting for 84.45% and 88.16%, respectively, and the rear wheel data for 86.72% and 85.38%, verifying that this method maintained high accuracy and robustness under different soil conditions. The research findings can provide theoretical support for the optimal drive force distribution strategy of tractors and contribute to improving traction efficiency and fuel economy.