A deep learning model for track irregularity long-term evolution of ballastless high-speed railway

Ballastless track of China’s high-speed railway has experienced a service time of almost two decades. The evolution path of track irregularity could be a mixture of periodic and trend patterns, with varied deterioration rates at different stages. While traditional methods face difficulties in modelling and predicting ballastless track irregularity, recent deep learning techniques provide new possibilities for a more efficient and efficacious solution. Therefore, a track irregularity evolution model, which belongs to the Transformer family with an encoder–decoder architecture, is proposed and termed TIEM. In TIEM, auto-correlation is introduced to learn sequence-wise correlations in historical track irregularity time-series, in place of the classical point-wise attention mechanism. Moreover, a specially designed functional block of convolutional decomposition is proposed to provide more flexibility in refining periodic patterns, as well as a shared feed-forward block for enhancing trend patterns. Several datasets, containing both deterministic amplitudes and stochastic spectra of track irregularity, are founded on the inspection data of four representative high-speed lines since operation, through proper means of characterization. Supported by the datasets, the necessity of functional blocks and the superiority of TIEM are validated through an ablation study and comparative experiments. After careful selection of hyper-parameters, the mean absolute error and mean absolute percentage error reach around 0.3 mm and 9%, respectively, for the prediction of track irregularity amplitude, outperforming traditional fully connected neural networks and long short-term memory, as well as recently emerged models belonging to the Transformer family. Besides, the proposed model also shows satisfactory performance in predicting multi-dimensional track irregularity spectrum.