基于串联深度神经网络的跨坐式单轨车辆轮胎径向载荷识别模型  

Radial Tire Load Identification Model of Straddle-Type Monorail Vehicle Based on Concatenated Deep Neural Network

作　　者：任利惠 周荣笙 季元进 曾俊玮 REN Lihui;ZHOU Rongsheng;JI Yuanjin;ZENG Junwei(College of Transportation,Tongji University,Shanghai 201804,China;Experimental Teaching Center of Transportation Engineering,Tongji University,Shanghai 201804,China)

机构地区：[1]同济大学交通学院,上海201804 [2]同济大学道路与交通工程教育部重点实验室,上海201804

出　　处：《中国铁道科学》2025年第1期136-148,共13页China Railway Science

基　　金：国家自然科学基金资助项目(52205121);中国博士后科学基金博士后创新人才支持计划(BX20200240);中国博士后科学基金资助项目(2020M671207)。

摘　　要：针对识别跨坐式单轨车辆轮胎径向载荷时直接测量法成本昂贵、定制复杂,而基于物理模型的方法稳定性差、计算量大、精度不足的问题,建立车辆动力学模型,兼顾物理关系合理性和测量便捷性,选取可通过能观性分解得到的车体和构架振动加速度以及易直接测量的位移、转角和角速度等车辆姿态信息构建数据集,并验证动力学模型的准确性;预处理数据集时,向其中混入噪声增强数据鲁棒性,进行归一化处理便于数据计算,扩充时间步长增强数据的时序关联性;在此基础上,构建基于一维卷积神经网络(1DCNN)和双向门控循环单元(BiGRU)串联深度神经网络的轮胎径向载荷识别模型,采用Hyperband算法进行模型的超参数优化,在学习率、批量大小和优化器种类最优下通过设置合理的卷积核尺寸和门控循环单元个数规划各层数据维度,在1DCNN中引入逐点卷积和膨胀卷积以提升模型识别效果,并从准确性、鲁棒性和泛化性3个方面对模型的载荷识别效果进行评估。结果表明:与传统模型相比,基于1DCNN-BiGRU的载荷识别模型均方误差较低,低于0.106,准确性较高;数据混入信噪比低至27 dB噪声时仍具有较好的识别效果,鲁棒性较强;在不同的曲线半径、曲线超高率和惯性参数扰动工况下仍能维持较好的识别效果,泛化性较好。To address the issues in identifying the radial tire load of straddle-type monorail vehicles-such as the high cost and complexity of direct measurement methods,and the poor stability,large computational load,and insufficient accuracy of physics-based models,a vehicle dynamic model is established.The model balances the rationality of physical relationships and the convenience of measurements.Vehicle posture information,including body and frame vibration acceleration that can be obtained via observability decomposition,and easily measurable parameters like displacement,rotation angle,and angular velocity,are selected to construct a dataset and validate the accuracy of the dynamic model.During dataset preprocessing,noise is added to enhance the data robustness,normalization is performed to facilitate data calculation,and time step expansion is carried out to strengthen the temporal correlation of the data.Based on this,a tire radial load identification model is built using a deep neural network consisting of a 1D convolutional neural network(1DCNN)and a bidirectional gated recurrent unit(BiGRU)in a serial architecture.The Hyperband algorithm is employed to optimize the hyperparameters of the model.By setting optimal learning rates,batch sizes,and optimizer types,and planning the data dimensions of each layer with appropriate convolutional kernel sizes and the number of GRU units,pointwise and dilated convolutions are introduced into the 1DCNN to improve model identification performance.The model′s load identification performance is evaluated from the perspectives of accuracy,robustness,and generalization.The results show that,compared to traditional models,the 1DCNN-BiGRU-based load identification model achieves a root mean square error lower than 0.106 with higher accuracy.The model still maintains good recognition performance under noise conditions with a signal-to-noise ratio as low as 27 dB,demonstrating strong robustness.Furthermore,under varying operational conditions,such as different curve radii,cant defic

关 键 词：载荷识别 跨坐式单轨车辆 卷积神经网络 双向门控循环单元 超参数优化 车辆动力学模型 

分 类 号：U270.11[机械工程—车辆工程]