Date of Award

May 2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Civil Engineering

First Advisor

Sattar Dorafshan

Abstract

Ballast contamination is a significant and growing challenge that causes railway track deterioration, poor drainage functionality, and early failure of railroad components. Traditional, human, destructive, field sampling, and most nondestructive evaluation (NDE) inspection methods are ineffective. They are plagued by drawbacks, such as inefficiency, subjective inspection outcome, alteration of existing structures, undermining the stability of the railroad substructure, traffic disruptions, and safety concerns, or providing limited information about the condition of the ballast. Therefore, a reliable noncontact technique is needed to evaluate, characterize, monitor, and predict the presence of contamination in railroad ballast. The conventional contamination indices only account for solid fouling without considering trapped moisture in fouled railroad ballast. This study presents the findings of deploying advanced nondestructive evaluation (NDE) technique-NIR spectroscopy, statistical-based and data-driven AI/ML models to detect and predict the combined presence of fouling and moisture contaminants in railroad ballast. The study also evaluated different ballast and fouling materials, with varying amounts of fouling and moisture contamination at different depths of contamination. This study presents the results of the spectral characterization, a proposed wet-fouled contamination index, and prediction outcome of the combined presence of fouling and moisture contamination using linear regression, principal component regression (PCR), and partial least square regression (PLSR) models. Moisture and fouling contents were varied from 10%-100%. Volumetric properties, such as the void ratio of the critically contaminated ballast (CCB), were determined. The fouling and moisture content for the CCB were determined and used to benchmark other wet-fouled conditions. The scan data were preprocessed and extracted the peaks of the first derivative (R_1st), and the Savitzky–Golay (S-G) first and second derivatives (D1 and D2) filters. The unprocessed raw reflectance spectral was also considered for model development to compare with other outcomes. After model calibration, leave-one-out-cross-validation (LOOCV) and ten k-folds were adopted for cross-validation. Linear regression models were developed for the extracted peaks of the R_1st and the contamination index for different fouling and moisture content combinations. The worst model had less than 0.5 coefficient of determination (R2) for a combination of datasets containing all investigated depths of fouling. The factors that affect the model’s predictability and generalization were evaluated on different combinations and complexity of the dataset comprising of varied ballast and fouling. The PCR and LOOCV outperformed the PLSR model in all cases. The models with excellent performance metrics on the four data-combination datasets (G/cl-L/cl-G/co-L/co) were selected. The outcome of the models have ranges of RMSEcal = 0.0010 to 0.001, RPDcal = 25.72 to 322.90, R2cal = 0.998 to 1.000, RMSEcv = 0.015 to 0.063, RPDcv = 3.902 to 15.377, R2cv = 0.929 to 0.996, and computation time = 9.2 to 92.3 seconds.

Download

Share

COinS