Files in this item



application/pdfDELAY-THESIS-2016.pdf (15MB)
(no description provided)PDF


Title:Machine learning techniques for identifying railroad ballast degradation
Author(s):Delay, Benjamin L
Advisor(s):Ahuja, Narendra
Department / Program:Electrical & Computer Eng
Discipline:Computer Engineering
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Machine Learning
Ballast Degradation
Computer Vision
Abstract:Railroad ballast is a layer of uniform sized crushed aggregate particles placed between, below and around the crossties. Railroad ballast transfers the load from crossties to the subgrade layer, provides lateral track stability and facilitates the drainage of water. Repeated traffic loading and environmental factors cause particle breakage, abrasion and polishing, which eventually degrade the ballast and result in fouling conditions. Traditional ballast fouling assessment includes manual sampling and identifying particle size distributions using sieve analysis. Recently, automatic ballast sampling (ABS) methods have been introduced to the railroad industry to obtain a sample of ballast and underlying layers using an approximately 1 m (3.28 ft.) long heavy duty steel tube driven into the ballast layer to depths of up to 2 m (6.56 ft.). Currently, visual-manual classification methods are used by experts to identify fouling conditions and degradation trends in the collected ballast samples. This thesis presents multiple approaches developed for the objective classification of ballast degradation using a combination of advanced machine vision and machine learning techniques. Initially, various computer vision algorithms are used to generate features associated with images of ballast cross sections at different degradation levels. Next, the generated features are used alongside a visual classification database provided by experts to develop, train, validate, and test a feedforward artificial neural network (ANN) using a supervised learning method. This work is further extended by implementing convolutional neural networks (CNNs) to serve as automatic feature generators. Finally, this approach is used on another cross-sectional ballast dataset that more closely resembles the type of ballast cross sections that can be found in the field. The findings of this study show that the proposed CNNs with an optimized topology can successfully classify ballast fouling in an effective and repeatable fashion with reasonable error levels. Further improvement of this technology holds the potential to provide a tool for consistent and automated ballast inspection and life cycle analysis intended to improve the safety and network reliability of US railroad transportation systems.
Issue Date:2016-12-09
Rights Information:Copyright 2016 Benjamin L. Delay
Date Available in IDEALS:2017-03-01
Date Deposited:2016-12

This item appears in the following Collection(s)

Item Statistics