Name: GLEDSON FÁBIO COTRIM ROCHA
Type: MSc dissertation
Publication date: 27/08/2019
Advisor:
Name |
Role |
|---|---|
| RODRIGO DE ALVARENGA ROSA | Advisor * |
Examining board:
| Name |
Role |
|---|---|
| DANIEL CRUZ CAVALIÉRI | External Examiner * |
| RENATO ANTÔNIO KROHLING | Co advisor * |
Summary: As the rail is the costliest element on the permanent track, finding tools to estimate the
remaining life is important to make the most of them. Thus, it has hypothesized that
machine learning algorithms trained from historical databases can assist the planner
in performing this estimation. Two machine learning algorithms were tested: Artificial
Neural Network (RNA) and Nearest K-Neighbors (k-NN). The historical databases
used were from Vitória to Minas Railway (EFVM). The dataset has 1,275,034 records
for the period of 6 years and 8 months. The RNA had 5 neurons in the input layer: 1)
the degree of the curve; 2) internal or external rail curve; 3) the width of the rail; 4) the
height of rail; and 5) average weight carried. The intermediate layer, regardless of the
category for estimating the remaining rail life, had variations of 1, 2 and 3 layers and
variations in the number of neurons of 30, 50, 100, 200 and 400. The output layer
depends on the period category the remaining useful life of the rails: 1) month, with 80
neurons, 2) quarter, with 27 neurons, 3) semester, with 14 neurons and 4) year, with 7
neurons. For the k-NN algorithm, configurations ranging from k = 5, 7 and 9 were
tested. For both algorithms, k-fold cross validation was applied, with f = 10, and
performance was evaluated using the accuracy value and F1-score. The programming
language was Python and the Scikit-Learn library. RNA configurations and k-NN
configurations were compared and k-NN showed superior results to RNA. However,
both algorithms reached the objective proposed in this dissertation, which was the
estimation of the remaining rail life in order to help the railroad maintenance planner to
replace the rails, WHERE they obtained results in accuracy and F1-score higher than
80% for both algorithms for the semester and year period categories, these period
categories being most used by rail operators. The k-NN algorithm always obtained
better results when compared to RNA.
Keywords: Rail Track. Artificial neural network. k-Nearest neighbors. Railway superstructure.
