Hybrid Reduction Techniques With Covariate Shift Optimization in High-Dimensional Track Geometry

Abstract

Abstract In discussions of track geometry, track safety takes precedence over other requirements because its shortfall often leads to unrecoverable loss. Track geometry is unanimously positioned as the index for safety evaluation—corrective or predictive—to predict the rightful maintenance regime based on track conditions. A recent study has shown that track defect probability thresholds can best be explored using a hybrid index. Hence, a dimension reduction technique that combines both safety components and geometry quality is needed. It is observed that dimensional space representation of track parameters without prior covariate shift evaluation could affect the overall distribution as the underlying discrepancies could pose a problem for the accuracy of the prediction. In this study, the authors applied a covariate shift framework to track geometry parameters before applying the dimension reduction techniques. While both principal component analysis (PCA) and t-distributed stochastic neighbor embedding (TSNE) are viable techniques that express the probability distribution of parameters based on correlation in their embedded space and inclination to maximize the variance, shift distribution evaluation should be considered. In conclusion, we demonstrate that our framework can detect and evaluate a covariate shift likelihood in a high-dimensional track geometry defect problem.