Tire–Road Friction Estimation for Four-Wheel Independent Steering and Driving EVs Using Improved CKF and FNN
Tire–Road Friction Estimation for Four-Wheel Independent Steering and Driving EVs Using Improved CKF and FNN
复制标题
DOI：
10.1109/tte.2023.3289140
复制
发表时间：
2024-03
影响因子：
7
通讯作者：
Guowang Zhang;Xiang-yu Wang;Liang Li;Xun Zhao
中科院分区：
文献类型：
--
作者：
Guowang Zhang;Xiang-yu Wang;Liang Li;Xun Zhao
轮胎-路面摩擦系数（TRFC）的估计一直是车辆控制领域的热点。这个问题在具有四轮独立转向和驱动（4 WISD）的电动车辆（EV）中也很重要。本文提出了一种基于改进的容积卡尔曼滤波器（CKF）和模糊神经网络（FNN）的TRFC估计方法。首先，基于先验知识的CKF（PCKF）被设计用于观察作用在轮胎上的力。PCKF算法利用先验知识对快速变化的状态进行跟踪。仿真结果表明，PCKF比CKF和扩展卡尔曼滤波器（EKF）具有更高的精度。然后，自组织非对称模糊神经网络（SOAFNN）的设计，以适应轮胎的特性，而不是轮胎模型。采用非对称隶属度函数（AMF）来减少网络参数的个数，并采用自组织方法来确定网络的最佳规模。结合PCKF和SOAFNN可以估计TRFC值。除了一种常见的情况外，我们还研究了4 WISD电动汽车特有的两种情况，这些情况在过去的研究中被忽略了。实验结果表明，该方案在三种情况下都具有比基线更高的精度。PCKF和SOAFNN的均方根误差（RMSE）值分别比CKF和递归最小二乘（RLS）的RMSE值降低了21.80%和19.71%。
Estimation of the tire–road friction coefficient (TRFC) has always been a hotspot in the field of vehicle control. This problem is also significant in electric vehicles (EVs) with four-wheel independent steering and driving (4WISD). In this article, a novel scheme based on an improved cubature Kalman filter (CKF) and a fuzzy neural network (FNN) is proposed for estimating the TRFC. First, a prior knowledge-based CKF (PCKF) is designed to observe forces acting on tires. PCKF can perform good tracking under rapidly changing states by using prior knowledge. Simulations show that the PCKF is more accurate than CKF and extended Kalman filters (EKFs). Then, a self-organizing asymmetric FNN (SOAFNN) is designed to fit the characteristics of tires instead of tire models. An asymmetric membership function (AMF) is used to reduce the number of parameters, and the self-organizing method is used to get the best size of a network. The TRFC value can be estimated by combining PCKF and SOAFNN. Apart from a common condition, we also study two more conditions unique to 4WISD EVs, which were simply ignored in past studies. The experimental results show that the proposed scheme has higher accuracy than those of the baselines under the three conditions. The root-mean-square error (RMSE) values of PCKF and SOAFNN are reduced by 21.80% and 19.71% than those of CKF and recursive least squares (RLSs), respectively.