A machine learning optimisation integration for enhanced railway crossing safety

Rail–road level crossings remain among the most accident-prone points in surface transportation, yet scarce public budgets often prevent blanket deployment of expensive engineering counter-measures. This study presents a two-stage, data-driven framework that (i) predicts the likelihood of accidents at individual crossings and (ii) selects the most cost-effective safety interventions under real-world constraints. In stage 1, a nationwide database covering 597 Turkish crossings (2015–2023) was cleaned, balanced and analysed with several classification algorithms. A random forest model, tuned using randomised search, achieved 88% accuracy and an area under the curve of 0.91, identifying crossing type, vehicle–train interaction, navigation moment, road-traffic volume and train visibility distance as the dominant risk factors. Stage 2 embeds these insights in a mixed-integer, multi-objective optimisation model that maximises safety benefit while respecting budget, capacity and traffic-flow limits. Solved with the branch-and-reduce optimisation navigator (Baron)/GAMS software, the model recommends installing automated barriers at the highest-risk sites, delivering the greatest risk reduction within the current budget envelope. The integrated framework links predictive analytics with actionable optimisation, enabling authorities to target limited resources where they avert the most accidents per unit cost. Beyond railway safety, the approach is transferable to other infrastructure domains that face similar risk-versus-budget dilemmas.