Suspension Gap Prediction of High-Speed Maglev Train Based on MconvTimeNet

The suspension gap between the train and the track is one of the most critical parameters in high-speed maglev transportation systems. Accurate and efficient prediction of the suspension gap is essential for the safe operation of the train. This paper proposes a suspension gap prediction method based on a deep hierarchical fully convolutional network model (MConvTimeNet), which adaptively segments and combines the time-series state data of the suspension system, eliminating the impact of sparse semantics. The model is designed with fully convolutional blocks based on deep convolution and pointwise convolution operations, using learnable weights to extract high-frequency and low-frequency information from the time series. Finally, Multi-Head Self-Attention is employed for multivariate feature fusion to achieve multi-step suspension gap prediction. The model was tested and validated using multivariate time-series state data generated from a real vehicle suspension system. The results show that the root mean square error (RMSE) and mean absolute error (MAE) of the MConvTimeNet model are 1.718 and 1.0085, respectively. Compared with the traditional ResNet model, the MConvTimeNet model reduces the RMSE and MAE by $\mathbf{2. 8 3 \%}$ and $\mathbf{2 3. 6 0 \%}$, respectively.