Abstract:  In view of the facts that direct tire load measurement was expensive and complex and the traditional load identification method had low accuracy and poor robustness, A method for radial load identification of rubber-tyred vehicles was proposed based on 1D CNN and BiGRU. Taking the prior information of tire radial load data into full consideration, the feature set was constructed based on multi-source information such as vehicle vibration response, body poses and running states. The effective feature subset was retained by feature selection, and multi-time-step input-single-time-step output samples were constructed for network training. The multi-dimensional spatial features of the signals were extracted by 1D CNN and input into BiGRU to capture the bidirectional temporal features. Finally, the load prediction results were output by the model. The theoretical model was modified by combining prediction accuracy, generalization performance and robust performance. Taking APM300 vehicle as an example for load identification, compared with the traditional algorithm, the proposed method may reduce the errors of load identification effectively, is suitable for different operating conditions, and may overcome different degrees of measurement noise, which has practical application values in the engineering fields.

Key words: load identification, rubber-tyred vehicle, one-dimensional convolutional neural network(1D CNN), bidirectional gating recurrent unit(BiGRU)

摘要： 针对轮胎载荷直接测量昂贵复杂及传统载荷识别方法精度低、鲁棒性差的现实，提出了一种融合一维卷积神经网络（1D CNN）和双向门控循环单元（BiGRU）的胶轮车辆轮胎径向载荷识别方法。充分考虑轮胎径向载荷数据的先验信息，以车辆振动响应、车体位姿、运行状态等多源信息构建特征集并经特征选择保留有效的特征子集，构造多时间步输入单时间步输出的样本用以网络训练。运用1D CNN提取信号的多维度空间特征并输入BiGRU中双向捕获时序特征，得到载荷预测的结果，结合预测精度、泛化性能、鲁棒性能修正理论模型。以APM300型车辆为例进行载荷识别，与传统算法相比，所提方法有效降低了载荷识别的误差，适用于不同运行工况，且能克服不同程度的测量噪声，在工程领域有现实应用价值。

关键词: 载荷识别, 胶轮车辆, 一维卷积神经网络, 双向门控循环单元