Correlating geometry and tie defects using track automated inspection data
Date
2018
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This dissertation focuses on the effect of tie conditions on geometry defect probability using automated inspection data. The analysis is based on data provided by Georgetown Rail’s AURORA tie inspection system and from CSX’s track geometry cars. The goal is to provide a tool that will be helpful to railroad engineers in scheduling tie and geometry maintenance and prioritizing railway segments that have higher rates of defects in terms of tie conditions. This dissertation combines engineering judgments and statistical analysis to develop analytical models to estimate the probability of developing geometry defect rates as a function of tie condition. ☐ The dissertation proposes different levels of statistical analysis that range from simple descriptive statistics and exploratory data analysis to different machine-learning techniques and deep-learning models. Predictive models of geometry defects were developed as part of this thesis. Also, geometry defect distribution parameters were estimated, and mixed models of defect distributions were presented. Furthermore, convolutional neural networks (CNNs) models were developed, as well as the outputs of the models used to build multiple regression models. Additionally, various data analysis issues were addressed in this dissertation. ☐ This dissertation’s contribution includes predictive models of track geometry defects as a function of tie condition and position. The three proposed models provide approaches to predicting the probability of geometry defects as functions of (a) tie conditions, (b) tie positions, and (c) tie conditions and positions. The three models showed an ability to identify locations of increased probability of developing defects, particularly for locations with poor tie conditions. The main and secondary objectives of this thesis, such as proving the existence of an effect of tie conditions on track geometry, developing models of geometry defect distributions in reference to tie conditions, and utilizing high-level machine-learning techniques, were met. Also, this dissertation includes a practical example of implementing the proposed models.