United States Department of Transportation
i
A Stochastic Rail Wear Forecast Using Mixture Density Networks and the Laplace Distribution
-
2019-10-31
[PDF-3.12 MB]
Details:
-
Creators:
-
Corporate Creators:
-
Corporate Contributors:University of Nevada. University Transportation Center on Improving Rail Transportation Infrastructure Sustainability and Durability ; United States. Department of Transportation. Office of the Assistant Secretary for Research and Technology ; United States. Department of Transportation. University Transportation Centers (UTC) Program
-
Subject/TRT Terms:
-
Resource Type:
-
Geographical Coverage:
-
Edition:Final Report 5
-
Corporate Publisher:
-
Abstract:The US rail network consists of more than 232,000 miles of track (as of 2016). More than 160,000 miles are classified as Class I railway (belonging to a railway with operating revenue greater than $458M). (Association of American Railroads, 2017). In 2016, Class I railroads replaced or installed nearly 7,000 miles of rail. (Association of American Railroads, 2016), which represents 1.5% of the total amount of rail in the US. This corresponds to an average (steady state) rail life of approximately 33 years. Rail wear research has been conducted since the 1860’s. Much research has been performed in the area of mechanics of rail wear and the parameters that most contribute to the rate of wear. Early research by Archard (Archard, 1953) focused on applied load and material properties (rail metallurgy/hardness) and developed wear coefficients for understanding the relative wear performance of various materials under different loading conditions. With the advent of automated inspection cars, railways had a method to measure and monitor rail wear over time (and accumulated traffic) more accurately and efficiently then measuring by hand. This however resulted in large amounts of data that were only used for go/no-go (threshold) analyses. Some work was performed to automate the process and determine wear rates from the measured wear data, but was limited in scope. This report formulates, solves and implements a stochastic rail wear forecasting model using mixture density networks (MDNs). MDNs take advantage of artificial intelligence, using a neural network, to define a median and shape parameter as output for a defined distribution, given causal factors (independent variables). In the case of rail wear rates, it was shown in a previous report that the rail wear rates follow a two side exponential distribution (Laplace distribution). Specifically, a stochastic modeling approach is presented and applied to more than 270 miles of railway track to develop a stochastic forecast of future rail requirements based on rail wear.
-
Format:
-
Funding:
-
Collection(s):
-
Main Document Checksum:
-
Download URL:
-
File Type:
