Railway Track Geometry Irregularity Exceedance Prediction Based on CNN-BiLSTM-Attention and Neural-Wiener Process Fusion with Degradation Feature Diversity

Railway track geometry prediction faces heterogeneity-data sparsity challenges: degradation dynamics vary, while inspections are sparse and irregular. Preventive maintenance further creates label-scarce, interrupted histories, invalidating direct prediction. In this study, a four-stage framework is proposed: An optimization model of degradation period division is designed based on adaptive detection of change point on trend curve. A CNN-BiLSTM-Attention model is constructed for multi-index prediction in a collaborative way. The probability distribution of the first arrival time on thresholds is estimated by integrating neural network and Wiener process. Finally, the prediction is re-corrected through a feature transformation approach. Based on six-year measured data on Wuhan-Jiujiang Railway in China and a cross-domain validation, the experiments show that the proposed method has distinct improvements compared with the existing methods in error controls for predicting both the value of track geometry and the exceedance time. This study provides theoretical and engineering support for preventive maintenance of railway tracks.