Reconstruction of high-speed train bogie hunting oscillations via a CNN-BiLSTM framework

Abstract

High-speed train bogie hunting waves exhibit strong nonstationary characteristics and often suffer from data loss. Traditional methods struggle to fit local transients, causing over-smoothing and information loss. To address this, a multi-scale spatio-temporal feature fusion model based on CNN-BiLSTM is proposed. The model employs a 1D-CNN to extract local shocks and high-frequency details, while a BiLSTM network captures bidirectional temporal dependencies, enabling end-to-end high-fidelity reconstruction. Validated on field-measured data, the method outperforms baseline deep learning models (BiLSTM, CNN-LSTM, LSTM, CNN) in reconstruction accuracy. The CNN-BiLSTM model achieves a Coefficient of Determination R 2 of 0.9365, significantly overcoming the smoothing distortion defects of traditional fitting methods. It accurately reproduces transient dynamics and maintains stable excitation energy distribution in the spatial-frequency domain. Results confirm that the fusion model possesses excellent generalization and robustness, enhancing the reliability of lateral stability assessment and fault diagnosis for high-speed train operation safety.