Irregularity prediction of slab track for high-speed railway based on IPSO-LSTM

To accurately predict the development trend of slab track irregularity for high-speed railways, a track quality index (TQI) prediction model (IPSO-LSTM) was built to improve particle swarm optimization (IPSO) and long short-term memory network (LSTM). The track irregularity detection data obtained by the track inspection vehicles were preprocessed by outlier elimination and noise reduction to generate time series data of TQI. Then the standardized TQI samples were used to carry out model training and irregularity prediction. Comparisons with other commonly used prediction methods were made. The results show that long short-term memory network has the function of memorizing historical information and can predict the development trend of nonlinear time series well. The difficulty faced by LSTM in hyperparameter selection, such as the number of hidden layer neurons and learning rate is solved. The model prediction performance is enhanced by adopting IPSO. For the track irregularity data at K5+000 to K7+000 section of a high-speed railway for 4 years, the IPSO LSTM model has the highest prediction accuracy for TQI, followed by the autoregressive integral moving average model (ARIMA). And the BP neural network is not much different from the grey model. The average relative error and root mean square error of IPSO-LSTM are 0.035 and 0.135, respectively. Compared with ARIMA, BP neural network and gray model, the average relative error is reduced by 22%～45%, and the root mean square error is reduced by 26%～45% for IPSO-LSTM. It verifies the validity of IPSO-LSTM model for predicting the irregularity of slab tracks. The IPSO-LSTM model is expected to provide a new technical support for controlling the development of slab track quality for high-speed railways. © 2023, Central South University Press. All rights reserved.