Prediction of Suspension Gap for High-Speed Maglev Trains Based on KANResnet Neural Network

The suspension control system of high-speed maglev trains is open-loop unstable and exhibits nonlinear characteristics. To achieve stable suspension control under high-speed conditions, it is necessary to accurately predict the suspension gap. This paper proposes a suspension gap prediction model for high-speed maglev trains based on a modified Residual Network (Resnet) using Kolmogorov-Arnold Networks (KAN), termed KANResnet. By introducing the KAN module to replace the fully connected layers in the Resnet, the model extracts the dependencies between variables such as the suspension gap, current, and voltage, as well as their temporal dependencies. This allows the integration of multi-variable related features and temporal features to complete the suspension gap prediction task. Using real data collected from actual maglev lines, the root means square error (RMSE) and mean absolute error (MAE) of the KANResnet model are 1.746 and 1.160, respectively. Compared with the traditional Resnet, the RMSE and MAE of the KANResnet model are reduced by 1.24% and 12.12%, respectively. The experimental results show that the proposed model can achieve accurate predictions of the suspension gap for high-speed maglev trains.