Optimizing Railway Track Tamping and Geometry Fine-Tuning Allocation Using a Neural Network-Based Solver

Abstract

This paper introduces a Neural Network Solver (NNS) for Railway Geometry Rectification Linear Program Model (RGRLPM), integrating tamping and fine-tuning operations for millimeter-precision adjustments. The NNS, enhanced by a grad norm process for faster convergence, achieves rectification plans three times faster than the simplex method. Dynamic programming is applied to allocate adjustments between tamping and fine-tuning. Experiments reveal that reducing 10 m and 5/30 m chord offset limits to 0.4 times improves dynamic performance over manual schemes. At a 0.2 reduction factor, cumulative rectification decreases by 5.6%, and the Sperling index drops by 26.9%, highlighting superior efficiency and dynamic outcomes.