Analysis of relationship between track geometry defects and measured track subsurface condition using different machine learning methods

Date
2022
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This research aims to enhance the further study of the relationship between track geometry defects and measured track substructure condition and to provide a mechanism to allow railroads to take proper preventive maintenance at a specific time for specific segments. For this study, a large amount of data was collected from railroad sites in Lincoln, Nebraska and Kansas City, Missouri on a major Class1 freight Railroad with the analysis concentrating on six different sites coming from St. Joseph and Creston subdivisions: Parkville (fouled), Parkville (control), Hickman (fouled), Hickman (control), Waldron, and Roca. The purpose of this research is not only to compare the results with previous studies done on Amtrak's Northeast Corridor, but also to apply more complex exploratory analysis, statistical modeling, and machine learning methodologies to develop more accurate predictive models using smaller segment lengths (e.g. 60 ft segments). The goal is to develop a more accurate predictive model that can deal with more varieties of collected data presented on multiple and different track sections that have shared values and attribute. ☐ Included in the investigation and analysis are variables such as Ballast Fouling Index and Fouling Depth Layer from Ground Penetrating Radar (GPR) inspection and Profile (measured over a 62-foot chord) as well as Cross-level from Track Geometry inspection. The resulting rate of degradations (slopes) was analyzed using several machine learning approaches such as Multivariate Linear Regression (MLR), Logistic Regression (LR), and Bayesian Linear Regression (BLR). After evaluating the different modeling approaches, the focus was placed on BLR machine learning technique to determine relationships for the changes in defects over time for each segment in the sites mentioned above and to forecast the time it would take for each segment to reach to track maintenance threshold as defined by railroad standards. The results and sensitivity analysis not only showed that increasing the Ballast Fouling Index and decreasing the Fouling Depth Layer (FDL) increases the Track Geometry Defect growth rate but also showed when preventative maintenances need to be performed.
Description
Keywords
Railroad, Machine learning, Preventive maintenance
Citation