Data-driven high-precision model identification and time series prediction for maglev train suspension systems

In response to low modeling accuracy and poor robustness in the suspension control systems of EMS-type maglev trains caused by high dimensionality, strong nonlinearity, and open-loop instability, a high-precision model identification and time series prediction method was proposed based on a long short-term memory (LSTM) improved by the Bayesian optimization (BO) algorithm. First, an experimental platform for a single-degree-of-freedom magnetic suspension system was built. A controller was used to track complex excitation signals, including sine, pulse, and white noise. Time series data of the control voltage and suspension gap covering the dynamic characteristics of the system were collected. The input and output sequences were reconstructed using the sliding window method. Second, an LSTM model was constructed to capture the time-dependent features of the suspension system. In response to the difficulty of manual parameter tuning in balancing efficiency and performance, the BO algorithm was introduced. A Gaussian process was selected as the probabilistic surrogate model, and the expected improvement was used as the acquisition function. The mean absolute error of the model on the validation set was set as the objective function. A global optimization was performed for the number of hidden layer neurons, initial learning rate, and L2 regularization coefficient of the LSTM model. Research results show that after BO, the model can accurately predict the key state changes of the suspension gap. The prediction error stabilizes within ±0.1 mm, and the error distribution conforms to a normal distribution, indicating that the structural deviation of the model is eliminated. Compared with the unoptimized LSTM model, the mean absolute error on the test set is reduced by 48.04%, and the root mean square error is decreased by 41.83%. This method effectively solves the difficulty of hyperparameter setting in suspension system modeling, and significantly improves the prediction accuracy and generalization ability of the model, providing a reliable model foundation for the design of high-precision suspension controllers and real-time condition monitoring of maglev trains.