Predicting vehicle-body vibration accelerations with track irregularities based on deep learning approaches

The vehicle-Track system is a complex nonlinear dynamic system. Conventional assessment of track condition based only on track irregularities is insufficient. Vehicle-track interaction should also be considered in track monitoring. It is a current trend to use vehicle-body vibration accelerations as an auxiliary evaluation index of track condition, which requires building predictive models to forecast vehicle-body accelerations from track irregularities. We proposed a FETimesNet-GRU (Frequency-Enhanced TimesNet-Gated Recurrent Unit). The FETimesNet transforms a 1D series into a set of 2D tensors based on multiple periods, utilizing 2D convolutional kernels to extract both local and global features. The GRU builds the dependency relationship between track irregularities and vehicle vibration accelerations. The experiments performed on the measured data of high-speed railway lines in China, show that the model outperforms CNN-GRU and Transformer, by 7.2%∼25.3% and by 1.6%∼21.1% for vertical vehicle-body vibration acceleration, and by 2.9%∼33.9% and by 1.0%∼34.6% for lateral vehicle-body vibration acceleration. Moreover, we built multi-train models for five train types and identified the worst-case scenario to guide track maintenance, which facilitated a comprehensive and more cautious track state assessment.