Deep learning-based site-specific prediction of rail joint-gap variations on curved ballasted track using field-measured rail data

Rail joint-gap variations due to thermal and geometric factors are the most prevalent durability challenge for concrete sleepers and track slabs, undermining overall track integrity. This study presents a deep neural network (DNN) model to predict the rail joint-gap variations on curved, ballasted bridge tracks using rail-based field measurements. For model training, 2570 data points (gap, temperature, cant, gauge, and position) were used. Three datasets (right-rail, left-rail, and combined) were used. Model performance was evaluated via MSE, MAE, and R 2 ; the DNN achieved an MAE of nearly 1.3 mm (< 15 % of the mean gap) and R 2 of 0.86 (right) and 0.80 (left). Residuals were near-normal, with over 95 % of predictions within ± 4 mm; bootstrap sampling yielded a narrow 95 % confidence interval for R 2 (0.83–0.89). SHAP analysis showed rail temperature as the primary predictor, with gauge, cant, and position contributing modestly. Incorporating on-site monitoring of fastening health and ballast behavior could further reduce errors. This system can reduce inspection hazards, enhance track safety, and contribute to proactively managing the integrity of railway concrete structures. • Developed a DNN using 2570 noise-augmented in-situ measurements to predict rail joint-gap variations on curved, ballasted bridge tracks. • Achieved MAE near 1.3 mm (< 15 % of mean gap) with R 2 of 0.86 (right rail) and 0.80 (left rail). • Residuals within ± 4 mm for > 95 % of predictions; bootstrap sampling (n = 1000) confirmed a narrow R 2 confidence interval of 0.83–0.89. • SHAP analysis identifies rail temperature as the primary predictor, while gauge, cant, and position exert secondary influences on gap variability.