PhD Defense by Mariana Costa

Event Details
  • Date/Time:
    • Monday October 12, 2020
      2:00 pm - 3:00 pm
  • Location: REMOTE: BLUE JEANS
  • Phone:
  • URL: BlueJeans Link
  • Email:
  • Fee(s):
  • Extras:
No contact information submitted.

Summary Sentence: Interfacing Data Harnessing, Stochastic Modeling and Optimization for Maintenance Decisions for Railways

Full Summary: No summary paragraph submitted.

Thesis Title: Interfacing Data Harnessing, Stochastic Modeling and Optimization for Maintenance Decisions for Railways



Dr. David Goldsman, School of Industrial and Systems Engineering, Georgia Tech

Dr. António Ramos Andrade, Department of Mechanical Engineering, University of Lisbon - IST


Committee members:

Dr. Brani Vidakovic, Department of Statistics, Texas A&M University (Adjunct Professor at Georgia Tech ISyE)

Dr. Joel Sokol, School of Industrial and Systems Engineering, Georgia Tech

Dr. Nagi Gebraeel, School of Industrial and Systems Engineering, Georgia Tech


Date: Monday, October 12th, 2020

Time: 2 pm - 3 pm, EST (GMT-4)


Meeting URL (for BlueJeans):


Meeting ID (for BlueJeans):

877 493 991



The increasing demand for cost-effective and transparent solutions for the improvement of the maintenance decision-making process in railways fuels the development of more sophisticated and flexible models, which largely exploit the use of data analytics and optimization tools. At the same time,  recent advancements in technologies for railway condition monitoring and the availability of massive amounts of data allow for more accurate and reliable fault detection. One obstacle, however, is how to deal with the data provided by the monitoring equipment as well as the choice of suitable methods to translate the data into useful information for maintenance scheduling and prioritization. In light of this, three main stages of the maintenance decision-making process can be identified: i) data acquisition, ii) modeling approach and, iii) implementation of the policy. Deciding on which parameter(s) represent the real condition of the asset and accurately measuring them, guaranteeing appropriate instrument and good measurement precision concerns data acquisition (step i)). Next, step ii) implies the choice of a comprehensive model that can tackle all the constraints and uncertainties associated with the deteriorating system, while providing solutions (in terms of a maintenance policy) in a reasonable amount of time. Finally,  step iii) concerns the ease of implementation of the new maintenance policy, guaranteeing its practical applicability within the context of the train operating company under study.


This dissertation aims to provide contributions to these three aspects in terms of railway track and wheelset maintenance. For both deteriorating systems, the choice of an appropriate maintenance policy should balance the trade-off between maintenance costs and costs resulting from the poor-maintained asset, including those arising from potential safety hazards. This is discussed in the context of the three main stages mentioned above.


The dissertation is structured in five chapters. Chapter 1 provides the introduction, as well as a brief overview of each of the topics and results presented in the subsequent chapters. Then, chapters 2 and 3 focus on wheelset maintenance, and chapters 4 and 5 focus on railway track maintenance.


In chapter 2, the optimization of railway wheelset maintenance policy is discussed. This policy is developed based on a data-driven model encompassing estimation of wear rates and further application of a Markov Decision Process (MDP) approach to represent possible discretized wheel states, where the problem of maintenance planning is tackled from the perspective of immediate action cost-optimization. A bidimensional framework considering discrete intervals of wheel diameter along with a quantitative variable (kilometers since last turning/renewal) is used to represent the possible wheel states. In addition, the probability of a defect interfering with the wheel maintenance schedule is modeled by contemplating survival curves derived from a Cox Proportional-Hazards model. As a secondary goal, a comparison of the optimized policy with another wheel’s reprofiling policy that is also "easy to implement" is provided.


In chapter 3, an investigation around the uncertainty of wheelset inspection data is made. Previous research has highlighted the relevance of this topic in the decision-making process surrounding wheelset maintenance actions. In light of this, the investigation aimed to assess the agreement between data acquired from three different inspection devices, namely: i) manual (gauge device), ii) a laser device and iii) an under-floor wheel lathe. Three main wheelset parameters (flange thickness (Ft), flange height (Fh), and flange slope (qR)) are compared using a Linear Mixed Model (LMM) approach under several real-world limitations, such as those imposed by serially correlated, unbalanced, and unequally replicated data. Findings supported the use of LMM, showing its ability to capture and account for the differences among the various groups and highlighting statistical significant performances of the inspection devices.


In the context of the railway track, chapter 4 presents a spatiotemporal approach for the modeling and prediction of track geometry faults. Spatial-time data from a train operating company is considered through a 5-year inspection database. The track twist, defined as the amount by which the difference in elevation of rails increases or decreases in a given length of the track, is used as the main track quality parameter. The spatiotemporal approach considered two Kriging models with a Gaussian correlation function to study a strategic portion of a track used in heavy-haul transport. A CUSUM (Cumulative Sum) control chart approach is then applied to identify out-of-control track sections and a Logistic Regression model is used to get estimates of the probabilities of future out-of-control points based on the adopted thresholds. Finally, a simple MDP model based on out-of-control points is proposed to compare different maintenance policies aimed at cost minimization for different thresholds of twist standard deviation for different track sections grouping strategies.


Lastly, chapter 5 explores the use of Wavelet Analysis (WA) in the statistical modeling of railway track irregularities, namely (1) longitudinal level, (2) alignment, (3) cross-level, (4) gauge. WA is used to study and reconstruct the four different track geometry irregularity signals. This investigation aimed at finding wavelets that can appropriately describe each track irregularity signal studied and investigating whether the presence of some high amplitude wavelet coefficients in certain frequencies can be associated with higher vertical or lateral forces in the wheel-rail contact. The last step is accomplished by reconstructing the different irregularities signals using wavelet coefficients in various decomposition levels and studying their impact on Nadal’s safety criterion Y/Q through vehicle dynamics simulations and, therefore, it allows the establishment of a relationship between wavelets and a higher risk of derailment.

Additional Information

In Campus Calendar

Graduate Studies

Invited Audience
Faculty/Staff, Public, Graduate students, Undergraduate students
Phd Defense
  • Created By: Tatianna Richardson
  • Workflow Status: Published
  • Created On: Sep 29, 2020 - 10:04am
  • Last Updated: Sep 29, 2020 - 10:04am