Approximate Bayesian computation and Bayesian nonparametric techniques in railroad track geometry modeling
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Rail infrastructure plays a crucial role in the economic development of a country through the facilitation of freight and passenger transport. The modern society recognizes the railway transportation system as an important and invaluable element of its structure. In the United States, the $80-billion freight rail industry alone provides over 167,000 jobs while providing additional advantages like reduction in fuel expenses, and traffic congestion. In order to sustain and improve these advantages, a substantial effort needs to be expended to keep the rail infrastructure in good condition. This can be achieved through determining which rail infrastructure should be maintained, when this maintenance should take place and how to go about maintenance to make sure that the rail system stays efficient. ☐ The track is an essential component of rail infrastructure hence a great deal of maintenance effort and significant proportion of the maintenance budget is allocated to it. Additionally, the quality of track geometry is directly linked to vehicle safety, reliability and ride quality. The performance of track is therefore considerably hindered when track geometry indicators deviate from the specified and approved limits due to loads and continuous usage. Track condition is often assessed in order to determine whether maintenance interventions are needed or otherwise. Track Quality Indices (TQIs) have been established to be sufficient indicators of track condition. TQIs are computed for track geometry parameters like gage, crosslevel, alignment and profile. This study therefore presents different approaches of predicting TQIs for the different geometry parameters from historical data. ☐ Probabilistic approaches have been favored over deterministic approaches because they recognize the role that randomness plays in the prediction of TQIs. Recently, there has been the application of Bayesian statistical methods in track degradation models. However, most models rely heavily on likelihood functions which are not available. There is therefore the need to explore alternative methods of building Bayesian models which do not depend on intractable likelihood functions. Information on track quality is typically combined to form large datasets. This data can be processed and filtered to obtain data/ information that is needed for a particular analysis task. The processed data can become small, making it undesirable to use prediction methods which work better on large datasets. Therefore, there is a need to explore methods that work well on small datasets, and measure the uncertainty associated with model predictions. Consequently, Approximate Bayesian Computation (ABC), also known as the likelihood-free method, is utilized in this study to formulate TQI models. The Bayesian Nonparametric method – Gaussian Process Regression is also applied to predict TQIs as it offers the advantage of being able to produce good predictions even with small datasets. ☐ The results from this study are very promising and indicate that ABC and Gaussian Processes are viable methods for developing TQI models. Maintenance managers can therefore make informed decisions on which track segments to prioritize during maintenance, based on model TQI predictions, optimizing time and financial resources, and improving track safety.
Description
Keywords
Approximate Bayesian Computation, Gaussian Processes, Machine learning, Railway transportation, Track geometry