PI-MSANet: a physics–data-driven method for vehicle dynamic response estimation under settlement deformation

View at publisher Summarise with VIRAT
Authors: Shijie Zhou, Hua Peng, Huihuan Zheng, Xiao He, Wenhui Ma
Cited by: 0
References: 32

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

During the operation of rail transit, the track structure may experience settlement deformation due to various factors, thereby impacting the operating vehicles. This study proposes a novel deep learning model – the Physics-Informed Multi-Scale Attention Neural Network (PI-MSANet) – to estimate vehicle vibration responses under settlement deformation conditions. The model takes track irregularities with settlement deformation as inputs, while the predicted outputs include the vibration responses of the vehicle body and bogies. During the training process, physical constraints are incorporated to ensure that the multiple predicted outputs comply with the vibration mechanism. In addition, a multi-scale attention module is designed to extract track irregularity features across different scales. To verify the effectiveness of each module, a series of ablation experiments were conducted. The results demonstrate that the incorporation of physical constraints not only ensures the physical consistency of the predictions but also significantly enhances training stability, while the multi-scale attention module greatly improves prediction accuracy. Finally, the PI-MSANet was employed to estimate vehicle vibration responses on complete tracks with additional settlement deformation. The model achieved R2 values for all variables exceeded 0.97, and could accurately reflect the effects of settlement wavelength, settlement amplitude, and vehicle speed.

Fields of Study & Subjects

Railway Engineering and Dynamics Structural Health Monitoring Techniques Model Reduction and Neural Networks

Publication details

Journal / source
Vehicle System Dynamics
Pages
1–21
Publication date
22 March 2026
Language
EN