In this paper, a novel method is proposed to predict the nonlinear seismic response of train-bridge coupled (TBC) systems by utilizing a long short-term memory (LSTM) recurrent neural network (RNN) based surrogate model. The surrogate model employed in this paper adopts a unidirectional multi-layer stacked LSTM architecture and implements sliding time windows for recursive calculation. The evaluation metrics used to assess the model’s performance have been enhanced to account for sensitivity variations in response amplitudes and to mitigate the phase-sensitive issues encountered with traditional evaluation metrics. Furthermore, network hyperparameters are carefully selected and presented for reference, and the surrogate model’s generalization ability is examined under different seismic scenarios. The results demonstrate that the LSTM-RNN-based model exhibits excellent computational accuracy and robustness when confronted with various types of seismic waves and system parameters. This approach offers fresh insights in situations where conventional numerical methods face limitations, such as rapid seismic response predictions in urban areas and simplifications for seismic design of high-speed railways. Overall, this paper contributes to the state of the art by introducing a novel approach that effectively predicts the nonlinear seismic response of TBC systems, addressing the increasing complexity and demands for accuracy and efficiency.

中文翻译：

---

使用基于 LSTM-RNN 的代理模型对地震作用下桥梁上的高速列车进行安全分析

本文提出了一种利用基于长短期记忆（LSTM）递归神经网络（RNN）代理模型来预测车桥耦合（TBC）系统非线性地震响应的新方法。本文采用的代理模型采用单向多层堆叠LSTM架构，并实现滑动时间窗口进行递归计算。用于评估模型性能的评估指标已得到增强，可以考虑响应幅度的敏感性变化，并减轻传统评估指标遇到的相位敏感问题。此外，网络超参数经过精心选择并提出以供参考，并在不同地震场景下检验替代模型的泛化能力。结果表明，基于 LSTM-RNN 的模型在面对各种类型的地震波和系统参数时表现出优异的计算精度和鲁棒性。这种方法为传统数值方法面临局限性的情况提供了新的见解，例如城市地区的快速地震响应预测和高速铁路抗震设计的简化。总体而言，本文通过引入一种有效预测 TBC 系统非线性地震响应的新方法，解决了日益增加的复杂性以及对准确性和效率的要求，为最新技术做出了贡献。