Ballast condition and degradation evaluation using computer vision techniques: Algorithms and applications

Luo, Jiayi

Permalink

https://hdl.handle.net/2142/124126 Copy

Description

Title

Ballast condition and degradation evaluation using computer vision techniques: Algorithms and applications

Author(s)

Luo, Jiayi

Issue Date

2023-12-14

Director of Research (if dissertation) or Advisor (if thesis)

Tutumluer, Erol

Doctoral Committee Chair(s)

Tutumluer, Erol

Committee Member(s)

• Barkan, Christopher Paul Lyman
• Ahuja, Narendra
• Edwards, John Riley
• Sussmann, Ted

Department of Study

Civil & Environmental Eng

Discipline

Civil Engineering

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Ballast
• Inspection
• Fouling Condition
• Degradation
• Railroad
• Engineering
• Computer Vision
• Deep Learning
• Object Detection, Image Segmentation
• Transfer Learning
• Semi-Supervised Learning
• Regression Analysis

Abstract

Railroads serve as one of the most efficient means of transportation within the U.S. The track substructure plays a significant and irreplaceable role in the overall performance of a railway track system in response to repeated train loading. A majority of the railroad structures are ballasted tracks, which may contain up to six basic components including rail, crossties, ballast, subballast, subgrade, and embankment. Among these, railroad ballast stands as a key component of the railroad track substructure. It is comprised of large, uniformly graded, and fully crushed coarse aggregate materials placed between and immediately underneath the crossties, providing drainage and structural support for the track system. Ballast degradation denotes the process where the material deteriorates as the voids in unbound aggregate layers are filled with relatively finer materials, or fouling agents, commonly resulting from the breakdown of ballast, contamination, and subgrade soil intrusion. As the ballast ages, it is progressively subjected to fouling and degradation caused by particle breakage and abrasion, thus leading to poor drainage, rapid and excessive settlement, and reduced lateral stability. To measure ballast fouling and degradation levels, and correlate it with ballast and track performance, several ballast fouling indices have been proposed. These indices fall into different categories: weight-based, volume-based and imaging-based. Depending on the specific analysis scenarios, different indices should be employed to provide accurate estimation of the ballast condition. Current ballast evaluation methods mainly rely on visual inspections, laboratory analysis of sampled ballast, and Ground Penetrating Radar (GPR) techniques. While visual inspections are subjective and sample collection has its limitations, GPR has been effective in assessing ballast conditions to some extent but falls short in detailed geotechnical analysis. Over the years, field inspection systems have evolved, with vision-based methods being developed for analyzing track conditions, but these primarily focus on track superstructure components, lacking in substructure evaluations like ballast fouling evaluation. Non-vision methods like GPR and Light Detection and Ranging (LiDAR) have been used for substructure inspections; however, they lack detailed geotechnical information about the ballast, accompanied by various resolution concerns and limitations. This doctoral research study focused on developing a novel ballast evaluation system by leveraging deep learning-based computer vision techniques to evaluate ballast condition and degradation level in the field. The proposed system entails the development of both software and hardware components to facilitate automated field ballast data collection and in-depth ballast condition evaluation. An imaging-based index, Percent Degraded Segments (PDS), is introduced to correlate image segmentation results with ballast degradation indicators, primarily Fouling Index (FI). The deep learning algorithms are developed and implemented, specifically Mask Region-based Convolutional Neural Network (Mask R-CNN), for precise ballast instance segmentation and establishing regression models to link PDS and FI. An automated Ballast Scanning Vehicle (BSV) is designed and constructed to acquire high-quality field data during operation, which is then analyzed using a companion data analysis software, I-BALLAST, for streamlining the evaluation procedure and minimizing user-dependency. Model improvements are pursued through (i) exploring advanced deep learning architectures like Vision Transformer, and (ii) a semi-supervised learning framework, Ballast Semi-Supervised Framework (BSSF), to enhance ballast segmentation performance. Several field experiments have been conducted during the development and validation of the ballast evaluation system. Two field experiments, in conjunction with Ballast Shoulder Cleaner (SBC), in Indiana and Ohio Norfolk Southern (NS) railroads, have been conducted to test the functionalities of the developed BSV and collect field data. The fully developed ballast evaluation system, after integrating the BSV and I-BALLAST, is successfully validated through a field experiment at the High Tonnage Loop (HTL) in the Transportation Technology Center (TTC), demonstrating its potential to provide accurate and robust ballast condition evaluations and geometric analyses over long continuous stretches of track, crucial for optimizing rail maintenance and rehabilitation planning on-site. In the longer term, the fully developed ballast evaluation system could serve as the key component of a comprehensive Ballast Management System (BMS) to study the deterioration mechanisms and improvement of ballasted track designs and ballast maintenance practices.

Graduation Semester

2024-05

Type of Resource

Thesis

Copyright and License Information

Copyright 2024 Jiayi Luo

Owning Collections