Railway track covariate shift formulation and analysis using machine learning techniques
Date
2022
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
The railroad industry accounts for approximately 40% of intercity freight volume in the United States - more than any other mode of transportation. Among the operational benefits of the railway infrastructure are road decongestion, highway fatalities, reduction in greenhouse gases, logistics and maintenance costs. An effort to keep the infrastructure and its components the focus of railroad research in recent times has been on developing advanced mathematical and machine learning models that best describe the condition of the track. The literature continues to discuss superiority of one method over the other, data uncertainty (non-stationarity), which can impact the overall analysis, are less considered. To avoid any forms of error while extracting useful information from the data, there is a need to explore alternatives that effectively process the large-scale track data without necessarily undermining the data attributes. This dissertation aims to evaluate freight and passenger tracks conditions, which includes surface defects image data and track geometry parameters using the Covariate-Shift Detection technique. The Covariate-Shift method is combined with machine learning (ML) techniques to resolve data inadequacies, shift in distribution during the training and testing stages. The implementation of the covariate shift detection is considered as a multi-stage analysis using US Class I railroad data. This data includes both the surface defects (corrugation, insulation joints, wear, and squat) and the geometry data (surface, cross-level, gage, and alignment). The initial stage involves using the covariate shift framework with supervised machine learning approach SVM, Decision Tree, and Random Forest. Principal Component Analysis (PCA) and T-Stochastic Neighbor Embedding (TSNE) are implemented as a tool to reduce the high dimensionality and uncertainties associated with the data. The final stage of the analysis focuses on detecting surface defects and improving track performance using a deep learning model and Generative Adversarial Network (GAN). Finally, the study illustrates that covariate-shift can be harmonized with machine learning and deep learning architecture to resolve track data uncertainties and models.
Description
Keywords
Covariate-shift, Data non-stationarity, Generative adversarial network, Machine learning technique, Railroad infrastructure, Railway track geometry