【摘要】 提出一种基于多头自注意力机制Transformer的震致轨道变形预测方法。基于OpenSees建立五跨高铁简支梁桥有限元模型，从太平洋地震中心数据库随机选取50条符合桥址场地特征的地震动，采样时间间隔统一调整为0.02 s，将其调整到多遇地震水平，以其作为横向地震输入，开展横向地震作用下的确定性分析，构建轨道-桥梁系统震致轨道变形数据库。结合地震动与梁端关键截面横向变形的时空特征，构建基于多头自注意力机制的Transformer预测模型，将解码器替换为全连接层，采用十折交叉验证进行训练。通过均方误差、平均绝对误差及决定系数评估模型精度。研究结果表明：该模型能有效预测桥上轨道震致变形时程，具备较高的精度与较强的泛化能力，可为地震下桥上轨道震损与行车安全评估提供技术支持。更多还原

【Abstract】 A prediction method for seismic-induced track deformation based on the multi-head self-attention mechanism Transformer model was proposed. A finite element model of a five-span simply-supported high-speed railway bridge was established using OpenSees. Fifty ground motions conforming to the bridge site characteristics were randomly selected from the pacific earthquake engineering research(PEER) center database, and their sampling time intervals were uniformly adjusted to 0.02 s. These ground motions were scaled to the frequent earthquake level and utilized as lateral seismic inputs to conduct deterministic analyses during transverse earthquakes. Subsequently, a database of seismic-induced track deformation for the track-bridge system was constructed. By incorporating spatiotemporal characteristics of ground motions and transverse deformations at key bridge-end sections, a Transformer prediction model based on the multi-head self-attention mechanism was developed, with its decoder replaced by fully connected layers. The model was trained using ten-fold crossvalidation. Prediction accuracy was evaluated through mean square error, mean absolute error and coefficient of determination. The results show that the proposed model can effectively predict the time histories of seismicinduced track deformation on bridges, exhibiting high accuracy and robust generalization capability. This methodology can provide technical support for assessing track seismic damage and train operational safety on bridges during earthquakes.更多还原