A Small–Large Model Collaborative Framework for Safety-Constrained and Interpretable Rail Replacement Decision-Making in Heavy-Haul Railways

Abstract

With the increasing transportation load of heavy-haul railways, rail damages continuously accumulate and exhibit significant spatial heterogeneity. Conventional rail replacement strategies based on fixed cycles, expert experience, or simple threshold rules often fail to balance safety and economy, leading to both over-maintenance and under-maintenance. To address these issues, this paper proposes a small–large model collaborative decision framework for heavy-haul railway rail replacement. The small model serves as the optimization core, formulating rail replacement as a safety-constrained Pareto decision problem that jointly considers post-decision damage safety and the service economy of replaced rails under engineering feasibility constraints. It constructs safety constraints, quantifies maintenance feasibility requirements, and generates a Pareto set of candidate replacement plans. The large language model does not directly solve the optimization problem; instead, it performs second-stage processing on the structured Pareto set by interpreting the safety and economic implications of different candidates, checking the consistency between numerical results and textual descriptions, providing a rule-based default recommendation, and explaining representative alternatives for human review. Experiments on Shuohuang Railway data show that the proposed method achieves full post-decision safety compliance while maintaining high average cumulative gross tonnage of replaced rails. Compared with conventional periodic and threshold-based rail replacement strategies, the proposed framework provides a better balance between safety and service-life utilization. Ablation results further demonstrate that the small model guarantees computational rigor and constraint compliance, whereas the large model improves interpretability, interactivity, and reviewability. Overall, the proposed framework supports the transition from cycle-driven rail replacement to condition- and value-driven, human-centered intelligent maintenance decision-making.