为研究高速铁路轨道板与钢轨之间的时空关联规律,提出变分模态-转换器(Variational Mode Decomposition-Transformer,VMD-T)反演模型,该模型通过分解轨道板振动加速度来反演估计钢轨振动加速度.首先,对数据进行预处理并利用双门限法检测振动端点,分离振动信号与静默信号、干扰信号,再将提取后的振动信号整合输入到VMD-T模型.其次,利用VMD模型将轨道板振动加速度数据分解成一系列不同的子模态,并网格遍历搜索与钢轨振动加速度相关系数最大的子模态,以降低原始数据的复杂度以及非平稳性,提升Transformer模型的特征抽取能力.然后,通过Transformer模型对搜索出的轨道板振动加速度子模态与钢轨振动加速度数据进行反演估计训练.最后,将该模型应用于某城际高速铁路轨道板与钢轨实测振动加速度数据反演估计试验.现场高铁试验结果表明:与单一Transformer模型相比,VMD-T模型均方根误差(Root Mean Squared Error,RMSE)、绝对平均误差(Mean Absolute Error,MAE)及决定系数(R2_score)分别提升近20％、11％及48.1％,特征学习能力更强,反演估计效果更佳,初步实现钢轨垂向振动加速度幅值非接触式监测估计.
Inversion estimation and field verification of rail vibration acceleration based on vibration acceleration of track slabs
To explore the spatiotemporal correlation patterns between high-speed railway track slabsand rails, this paper proposes a Variational Mode Decomposition-Transformer (VMD-T) inversion model, which estimates rail vibration acceleration from decomposed track slab vibration accelerationdata. Firstly, the data undergoes preprocessing, including vibration endpoint detection using a dual-threshold method to separate the vibration signals from silent and interference signals. The extracted vibration signals are integrated and fed into the VMD-T model. Subsequently, the VMD model breaks down the track slab vibration acceleration data into various sub-modes. A grid search identifies the sub-modes with the highest correlation coefficients with rail vibration acceleration, aiming to sim-plify the original dataset and reduce its non-stationarity, thereby enhancing the Transformer model's feature extraction capabilities. Following this, the Transformer model undertakes inversion estimation training on the identified track slab vibration acceleration sub-modes and rail vibration acceleration data. Finally, the model's utility is demonstrated through inversion estimation trials using actual vibra-tion acceleration data from the track slabs and rails of a specific intercity high-speed railway. Field tests on high-speed rail reveal that, compared to a standalone Transformer model, the VMD-T model con-siderably improves in terms of Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the determination coefficient (R2_score) by approximately 20％, 11％, and 48.1％ respectively, indi-cating a stronger capability in feature learning and inversion estimation. This study preliminarily achieves non-contact estimation monitoring of rail vertical vibration acceleration magnitudes.
万方数据
维普
