Random Forest–Based Covariate Shift in Addressing Nonstationarity of Railway Track Data

Abstract

The multimodal transportation network in which freight rail plays an essential role continues to enhance the United States’ contributions in the global market. For years, track geometry defects data have been often gathered by visual inspections. However, automated track vehicles are now deployed for the same purpose to save time and cost. One of the limitations of an automated vehicle is the likelihood of nonstationarity of the gathered data due to external influence. The effect of nonstationarity may lead to the wrong representation of track conditions and thereby increases the possibility of false model output. This study applies supervised machine learning (ML) methods to detect the nonstationarity of the geometry data. The methods include random forest (RF), logistic regression (LR), and support vector machine (SVM). The authors vary the train test and validation ratio in phases to ascertain each machine learning method’s accuracy and adaptability to different distributions. In the first phase, the random forest and the support vector machine show an accuracy of 97.1%, while the logistic regression reveals 96% accuracy. In the second and third phases, the random forest method gives a better result than other supervised learners with accuracies of 97% and 97.1%, respectively. Similarly, for validation, the random forest performs better than other classifiers. Conclusively, the developed models’ application indicates that the random forest is a more effective approach to detecting the nonstationarity of track geometry data.