Railway track geometry degradation: A review of prediction methods

• Synthesizes track-geometry degradation models across five major methodological families. • Benchmarks methods using five criteria: accuracy, uncertainty quantification, scalability, interpretability, and cost. • Distinguishes segment-level indices, point-wise forecasting, and European Standard EN 13848-5 exceedance events. • Identifies why localized exceedances are under-detected (label scarcity and smoothing). • Proposes two-layer labeling to separate exceedance events from anomalies. • Outlines deployment-ready directions: low-computational-cost hybrid models, edge analytics, interpretable uncertainty. Maintaining track geometry within tolerance is essential for safe, efficient, and economical railway operations. This review—framed around both aggregate degradation trends and isolated defects—synthesizes five modeling families (empirical, mechanistic, statistical, probabilistic, and machine learning) and evaluates them against five criteria: predictive accuracy, uncertainty quantification, network-scale scalability, interpretability, and computational cost. Empirical and mechanistic models offer transparent, physics-based insights at low computational cost, yet typically lack dynamic train–track interaction modeling and explicit uncertainty estimates. Statistical and probabilistic approaches (stochastic processes, Markov chains, Bayesian inference) support long-term planning and maintenance scheduling but require extensive high-quality data and careful calibration. Machine-learning methods—particularly random forests and convolutional neural network–long short-term memory hybrids—achieve high accuracy on complex spatio-temporal patterns; however, they depend on European Standard EN 13848-1/-5–compliant preprocessing, well-labeled datasets, and dedicated interpretability measures. A persistent cross-cutting gap is the under-detection of isolated defects, driven by label scarcity and the smoothing inherent in section-level indices (e.g., track quality index/standard deviation). Priority directions include: (i) low-computational-cost hybrid frameworks via knowledge distillation and quantization for network-wide deployment; (ii) multivariate Bayesian and point-process models to jointly represent standard deviation of longitudinal level/standard deviation of horizontal alignment and localized jumps; (iii) internet of things/edge analytics for low-latency detection; and (iv) interpretable uncertainty (e.g., Monte Carlo dropout, local interpretable model-agnostic explanations/Shapley additive explanations). Balancing accuracy, scalability, real-time operation, uncertainty quantification, and interpretability is pivotal for transitioning from reactive maintenance to intelligent, self-adaptive asset management.