多轴特种车辆的数据建模方法及横向动力学应用

多轴特种车辆的动力学模型具有强非线性，精细化的物理建模需要准确的模型参数和动力学方程以映射车辆动力学的特性。在无精确的车辆先验物理参数信息和动力学函数关系条件下，为提高车辆动力学建模的保真度，针对某型五轴特种车辆的横向动力学行为，提出了一种基于神经网络的数据建模方法。网络框架主体呈闭环结构，网络输出的状态信息同时作为输入用于预测下一时刻的状态，实现了数据建模递归更新；针对闭环网络模型，设计了闭环结构的训练策略，在网络模型中引入中间变量，使得网络在训练阶段仍然保持闭环结构；网络模块采用循环门控单元(Gate Recurrent Unit)和全连接网络(Full Neural Networks)的组合方式；数据集由经过实车验证的Trucksim仿真模型生成，分析结果表明：在无精确车辆先验信息条件下，物理建模难以准确预测出车辆的状态信息，数据模型具有更好的保真度。闭环训练方法可以使得闭环结构的网络具有更好的保真度，对于横向速度和横摆角速度预测的最大绝对值误差仅为0.079 km/h和0.342°/s,相比于开环训练的结果，最大误差降低了58.40%和49.48%。

Data Modeling of Multi-Axle Special Vehicles and Lateral Dynamics Applications

The dynamic model of multi-axle special vehicles has strong nonlinearity. The modeling method based on physical laws requires precise model parameters and complex mathematical equations to reflect the characteristics of vehicle dynamics. In the absence of accurate prior physical parameter information of the vehicle and dynamic function relationship, to improve the fidelity of vehicle dynamics modeling, a data modeling method based on neural networks is proposed for the lateral dynamic behavior of a five-axle special vehicle. At the same time, it is used as an input to predict the state of the next moment, and the recursive update of data modeling is realized; for the closed-loop network model, a training strategy is designed for the closed-loop structure, and intermediate variables are introduced into the network model, so that the network still maintains the closed-loop structure during the training phase; the network module adopts a combination of Gate Recurrent Unit (GRU) and Full Neural Networks (FNN); the data set is generated by the TruckSim simulation model that has been verified by real vehicles. The results show that it is difficult for physical modeling to accurately predict vehicle state information without accurate prior vehicle information, and the data model has better fidelity. The closed-loop training method can make the network with a closed-loop structure have better fidelity. The maximum absolute errors of the prediction of lateral velocity and yaw velocity are only 0.079km/h and 0.342°/s; compared with the results of open-loop training, the maximum errors are reduced by 58.40% and 49.48%.