Short-term prediction of railway track degradation using ensemble deep learning

Abstract

Short-term prediction of track degradation facilitates flexible and efficient maintenance, thereby meeting the railway system's escalating demands for track safety and smoothness. However, the track condition evolution presents challenges to accurate prediction, with diverse influential factors resulting in heterogeneous degradation patterns across space and time. In a short-term context, time series derived from historical records are length-limited, with sparse sampling points complicating feature identification. Actual activities, particularly minor repairs, lack strict periodicity, leading to irregular spans in continuous degradation curves, yielding nonuniform samples. This study leverages dynamic inspection and influential factors to propose an ensemble learning using the Transformer model. The outer framework employs unsupervised learning to group the sections based on specific time periods and track lengths. It assigns fuzzy logic categories to these groups to capture differentiated patterns and guides the division of samples into fuzzy subsets and assigns them to the learners corresponding to each cluster. The loosely coupled structure aids task decomposition and enhances local performance. The inner model refines the Transformer design for a new scenario, introducing a prediction objective transformation based on the interdependencies among multidimensional indicators to strengthen feature extraction. The prediction performance is evaluated using over 2 years of records from 560 km railway lines, offering insights for improving onsite track management.