See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/329377226

Principal components analysis and track quality index: A machine learning

approach

Article  in  Transportation Research Part C Emerging Technologies · April 2018

CITATIONS READS

5 648

1 author:

Ahmed Lasisi
University of Manitoba
28 PUBLICATIONS   75 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Master's Thesis on the Use of NDT for SHM View project

Principal Component Analysis in the Evaluation of Track Quality Index View project

All content following this page was uploaded by Ahmed Lasisi on 03 December 2018.

The user has requested enhancement of the downloaded file.

 Transportation Research Part C 91 (2018) 230–248

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier.com/locate/trc

Principal components analysis and track quality index: A machine

T
learning approach
⁎
Ahmed Lasisi, Nii Attoh-Okine
Department of Civil and Environmental Engineering, University of Delaware, Newark, DE, USA

A R T IC LE I N F O ABS TRA CT

Keywords: Track geometry data exhibits classical big data attributes: value, volume, velocity, veracity and
Track quality index variety. Track Quality Indices-TQI are used to obtain average-based assessment of track segments
Principal components and schedule track maintenance. TQI is expressed in terms of track parameters like gage, cross-
Machine learning level, etc. Though each of these parameters is objectively important but understanding what they
Railway engineering
collectively convey for a given track segment often becomes challenging. Several railways in-
cluding passenger and freight have developed single indices that combines different track
parameters to assess overall track quality. Some of these railways have selected certain para-
meters whilst dropping others. Using track geometry data from a sample mile track, we de-
monstrate how to combine track geometry parameters into a low dimensional form (TQI) that
simplifies the track properties without losing much variability in the data. This led us to principal
components. To validate the use of principal components as TQI, we employed a two-phase
approach. First phase was to identify a classic machine learning technique that works well with
track geometry data. The second step was to train the identified machine learning technique on
the sample mile-track data using combined TQIs and principal components as defect predictors.
The performance of the predictors were compared using true and false positive rates. The results
show that three principal components were better at predicting defects and revealing salient
characteristics in track geometry data than combined TQIs even though there were some cor-
relations that are potentially useful for track maintenance.

1. Introduction

This paper examines the potential of machine learning applications in railway track engineering. The railroad industry across the
world is experiencing an increased demand in operations and services due to the world’s increasing human needs and global trade
advancements. The US Class I freight railroads alone is worth over $60 billion operating approximately 140,000 miles of track (FRA,
2013). This is second to none in the world. Other major railway networks include but not limited to Russia, China, Japan, etc. are also
undergoing similar growth. The United Kingdom currently boasts of about 20,000 miles of track which currently ranks it in the first
20 rail networks in the world (Odlyzko, 2016). The global railway network is expected to reach an estimated 900,000 miles-route in
2020 (Statista, 2017). As promising as the industry is, a significant amount of the industry’s budget is continuously committed
towards repairs and maintenance. The annual capital program for US railroads alone is about $15 billion (Zarembski, 2015). A
railway disaster is often very ruinous and sometimes fatal. Track geometry-caused derailments constitute a significant portion of
railway accidents worldwide. In the United States alone, track-caused accidents have consistently constituted 30–40% of total

This article belongs to the Virtual Special Issue on “Big Data Railway”.
⁎
Corresponding author.
E-mail addresses: [email protected] (A. Lasisi), [email protected] (N. Attoh-Okine).

https://doi.org/10.1016/j.trc.2018.04.001
Received 26 July 2017; Received in revised form 14 March 2018; Accepted 2 April 2018
0968-090X/ © 2018 Elsevier Ltd. All rights reserved.
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

accidents for the past decade (FRA, 2015). This is only second to human as a primary cause category.
Understanding track geometry from a design or mechanistic perspective is often not far-fetched. However, it is difficult to monitor
the performance of track infrastructure and geometry throughout its life cycle with the same mechanistic models used during design
without considering inevitable influences of the loading environment, climate, track-train dynamics and a host of other factors (DYK,
2014). To continuously, assess track geometry performance, it is common practice to take ‘per-foot’ measurements of crucial track
geometry parameters for every mile of track using track geometry cars (Kelshaw, 1995). With varying geometry parameters and
wavelength complexities, measurements taken on every foot of mile across 140,000 miles of track in the US results in about a billion
observations for a given inspection date. If monthly routine inspections are scheduled, the data explodes to about several billion
observations. Other measurement systems include high resolution track imaging, wheel impact load detection, strain gauge mea-
surements, LIDAR and GPR-based ballast measurements, rail wear inspection system, etc. (Zarembski et al., 2013). The scale of data
measurements in US railroads has grown from megabytes and gigabytes in the past three decades to terabytes and petabytes today.
With this ‘mess’ of perpetual data collection, we are confronted with two major challenges that are characteristic of a typical ‘big data’
problem (Ishwarappa and Anuradha, 2015). Firstly, the problem of storage has been significantly confronted by modern cloud
solutions. Secondly, data analysis and processing time remains a major challenge in order to make sensible conclusions from the
entire ocean of data collected. Therefore, the fact that track geometry data exhibits classical big data attributes cannot be over-
emphasized.
In this paper we investigate the possibility of reducing multivariate track geometry indices into a low-dimensional form without
losing much information. Similar to the Pavement Condition Index in highways wherein weights are assigned to each parameter and
then summed up (Karim et al., 2016). However, our proposed approach takes cognizance of the fact the observed multidimensional
data often lies in an unknown subspace of two to three dimensions (Hastie et al., 2009). Hence, detecting this subspace in track
geometry data can significantly enable us to eliminate redundant information. This will make it possible to visualize multi-
dimensional track geometry data in two or three dimensions which was hitherto impossible with the raw parameters obtained from
track geometry cars. The second section of this paper focuses on introducing track geometry parameters, data collection and track
quality indices. The third section considers selected machine learning methods that are used to train, test and validate the use of
single and combined track quality indices including the proposed principal components. Low-dimensional representation of multi-
variate track geometry parameters in terms of principal components was validated and compared to existing TQIs in the penultimate
section. The last section of this paper highlights key findings with concluding remarks.

1.1. Overview of track geometry derailments

A major component of railway accidents and derailments are due to track geometry irregularities. These irregularities include
wide gage, excessive warp/twist, horizontal and vertical rail deformities which could also culminate in broken rails and derailments.
The Federal Railroad Administration (FRA) reported over 450 track geometry-caused derailments in 2016 according to major cause
categories. In the preceding year, the worth of track geometry derailments constituted about 15% of all major accidents (FRA, 2015).
While derailments are often misunderstood to imply all train accidents, there are in fact other train accidents but statistics show that a
significant proportion of accidents are derailments (Scientific American, 2016). As a result, major railways incur huge expenditures
on track maintenance activities alone. These activities include inspection programs, rail grinding, tamping, etc. so as to reduce the
risk of track geometry failure. In 2008 alone, about $8 billion was spent by US Class I railroads on track maintenance activities (Peng,
2011). A substantial portion of this cost go into geometry inspection. Historically, track geometry inspections were carried out
manually. In recent times, the use of automated inspection cars to collect and assess track geometry cars has significantly expedited
the inspection process spatially and temporally. The increase in the deployment of track geometry cars and several cutting-edge
technologies has aided the performance of risk-based scheduling of track maintenance activities with the philosophy to simulta-
neously minimize cost, optimize operations and reduce risk (Jamshidi et al., 2017).

1.2. Background

In the area of machine learning and track geometry data analytics, some researchers have focused on track classification, track
structure, components and defect detection (Molodova and Li, 2014). Significant contributions have also been made in the area of
predictive maintenance and condition monitoring. Yella et al. (2009) employed a classifier fusion to categorize sleepers in railway
tracks using machine vision. Using support vector machine and classifier fusion, the researchers arrived at a prediction accuracy of
92%. Fernando et al. (2011) worked on a machine vision inspection of railroad track sorting different track components from one
another. They outlined an algorithm using edge detection and texture information to inspect fasteners, cut spikes and switch com-
ponents. In the area of geometry forecasting, Peng (2011) proposed a short-range prediction method to forecast the amplitude of
future measurements by track geometry cars. With Bayesian statistics, a work which focused on a classification-based learning using a
tree augmented naïve Bayes track quality index (TAN-TQI) established better prediction accuracy than the short-ranged method (Bai
et al., 2016) used to predict track irregularity of each day in a future short period (Xu et al., 2012). Hu and Liu (2016) completed a
recent work on modelling track geometry degradation using support vector machine. The trained classifier was 70% accurate when
tested on a validation set. Understanding the behavior of track geometry data is an important step towards developing an effective
classification or regression tool. Investing some resources in this area prior classification would greatly improve on the results of the
above-cited papers. Li et al. (2008) conducted some correlation analysis between track geometry and surface defects like squats. The
research found a close correlation between short-wave track irregularities and the latter. Whether track quality is an indicator of

231
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

performance was the subject of the investigation by Hendry (2015), the results of the research proved to be affirmative. To assess the
importance of machine learning in predictive track maintenance, Li et al. (2014) investigated how rail network velocity can be
improved. While all of these studies aimed to solve different railway/track related problems, the ability to reduce track parameters
into simple variables can significantly simplify many complex problems formulated in the above-mentioned studies. Although, it has
been established that some of the challenges posed by point-based regression has been solved by profile-based prediction (Pedankar
et al., 2003) by segmenting tracks into chunks of specified length (e.g. 100 ft, 500 ft, 200 m, etc.), the challenge that remain to be
addressed is simplifying the variety of TQIs from these track segments into one or two meaningful variables. There is no doubt that
high dimensional data can be difficult to visualize or even understand but with effective dimension reduction tools that can possibly
explain a great portion of the data’s variance in 2D or 3D, it is possible to interact with track defects and data in a Cartesian
coordinate and possibly specify maintenance thresholds. Sequel to previous attempts to develop artificial indices by using linear
combination of track parameters, Liu et al. (2015) demonstrated an application of such for Beijing Metro. However, it was established
in this work that with two principal components only, over 90% variance of track geometry data was effectively summarized and
could therefore be a reasonable substitute for the multidimensional track quality indices. In the following section we introduce the
commonly used track quality indices just before discussing principal components.

2. Track geometry, defects and track quality index (TQI)

Track geometry is a description of the track in terms of its longitudinal (alignment), transverse (gage) and vertical properties
(surface/profile and cross level). Other track parameters combine these track irregularities in two-dimensions or more, e.g. vertical
and longitudinal (warp/twist). Track quality index on the other hand is a quantitative representation of ride quality in an attempt to
distinguish a good track from a ‘bad’ one. At this point, it is important to distinguish between track index, defects, irregularities and
how they contribute to derailments. Firstly, tracks are laid to meet very stringent construction standards. Wear and tear as a result of
track usage and tonnage results in deviations from construction standards. These deviations are often found in rails, track geometry,
structure, etc. Since track parameters are often defined by a nominal value which is the characteristic of an ideal track. Deviations
from these nominal values develop into track irregularities (Ciobanu, 2016). These irregularities grow gradually until it reaches an
unacceptable limit (maintenance threshold) that requires intervention. Nominal values for a parameter beyond this limit defines a
defect as seen in Fig. 1. Track geometry defects left to propagate is likely to lead to derailments as discussed in Section 1.
To evaluate, assess and make decisions based on each parameter per unit length of track is almost practically impossible because it
results in tremendous data-points and hypersensitivities in variations. Therefore, TQI is employed as an aggregate measure of a given
track geometry parameter over a specific length of track. Standard deviation, mean, power spectral density (PSD), etc. are among the
common average-based measures used as TQIs. Next, we discuss crucial track parameters and track quality indices expressed in terms
of individual parameters.
TQI = f (wi ) (1)
where TQI is the track quality indicator and wi is a single track parameter (Fazio, 1986).

2.1. Track geometry parameters and data

There is not a single set of track geometry parameters across the world. Europe and North America play a dominant role in all
major sectors of the industry (Thompson, 1989; Woodburn, 2014). Hence, while most of the track geometry parameters are similar
across the industry globally, there are subtle variations in the usage and understanding of these terms. A fundamental difference is
Gage for example. Variations such as chord length in taking measurements on vertical or horizontal irregularities also exist. For the
purpose of this research, the following track parameters have been chosen; Gage, Cross level, Surface, Alignment and Warp. They are
defined as follows: Gage is the distance between the inner faces of two adjacent rails measured at a certain distance (e.g. 5/8 in.)
below the rail surface. Surface (or profile) is a measure of uniformity in the vertical plane of each rail measured at the mid-point of a
given chord length (e.g. 31 ft or 62 ft). Cross level (or cant) is the vertical difference in height (or angle) between two adjacent rails
while Warp (or twist) measures changes in cross level (or cant) within a specific length interval (e.g. 62 ft) along the track length.
Alinement (or alignment) is the measure of uniformity in the horizontal plane of each rail measured at the mid-point of a given chord
length (FRA, 2002). The alternative terms in parentheses following the parameters used above are the European equivalents of the

Fig. 1. Track quality indices, tolerances and defects (Ciobanu, 2016).

232
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

North American track geometry parameters. While these terms are often used interchangeably they are not essentially the same. A
comprehensive detail of these terms, their subtle similarities and differences have been sufficiently discussed in Hay (1982) and
Esveld (2016). All of the said parameters are either collected manually or via automated track inspection vehicles which obviously is
the order of the day. For safety reasons, each of these parameters are better kept within defined thresholds. A violation of a track
parameter threshold is tantamount to a defect. While there are several types of thresholds (construction, maintenance and safety – in
order of increasing tolerance limits), it suffices to examine only safety in this work since it is the major threshold crucial to de-
railment.

2.2. Objective track quality indices (single)

For objective TQIs, surface, cross-level, gage, alignment, etc. are often expressed as separate TQIs to address different aspects of
track quality (FRA, 2005). A simple overview of the parameter-based TQIs for each parameter amongst other TQIs are discussed
below:

2.2.1. FRA length-based TQI

The Federal Railroad Administration (FRA) employs the basic concept of space curves generated on a foot by foot basis to
represent track quality (FRA, 2005). This makes it easy to express length-based parameters in different wavelength variations (e.g.
31 ft, 62 ft, or 124 ft). The length of a space curve is estimated as the sum of straight distances between adjacent data points (FRA,
2005).
n
Ls = ∑ Δx i2 + Δyi2
i=1 (2)
where Ls = traced length of space curve, Δx = sampling space; and Δy = difference of two consecutive measures. A set of TQIs
computed afterwards for gage, alignment, surface and cross level is given by the equation below:

L
TQIs = ⎛ s −1⎞ × 106
⎜ ⎟

⎝ Lo ⎠ (3)
where Lo is the theoretical length of section.

2.2.2. Canadian rail track quality index

Here, the Canadian National Railway provides both an objective and combined TQI approach. The railway considered a 2nd
degree polynomial of the standard deviation of each parameter measured within a specific track length and evaluate the TQI using
the equation below:
TQIi = 1000−C (σi2) (4)
where C = 700 for mainline track and i stands for each parameter. These TQIs are combined by averaging the TQIs for six parameters
excluding warp/twist (Liu et al., 2015).

2.2.3. Track roughness index (Amtrak)

This is defined as the average of squared measurements, x for a given parameter, i measured over a certain track section, j. It was
developed by Amtrak in 1998.
n
x ij2
TRIi = ∑ n
j=1 (5)

2.3. Artificial track quality indices (combined)

As mentioned in Section 2.2, the number of TQIs per parameter for each track section becomes challenging to handle. It grows
rapidly if several wavelengths of the same parameters are also considered and therefore not effective for maintenance decisions. To
solve this problem, many railroads have developed a way to combine all TQIs into a single measure of ride quality that can be used for
maintenance planning (as highlighted in the case of Canadian National Railway above). We have selected TQIs from both freight and
passenger railways in order to observe the inter-correlations and possibly discern any relationship they would have with the principal
components proposed in this work. Some of those considered are given below.

2.3.1. J-synthetic coefficient

This is a track quality indicator developed by the Polish railways for a quantitative evaluation of track condition (Madejski and
Grabczyk, 2002).
SZ + SY + SW + 0.5∗SE
J=
3.5 (6)

233
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

where SZ ,SY ,SW and SE are standard deviation values of vertical irregularities (surface), horizontal irregularities (alignment), twist/
warp, and gauge, respectively (Berawi et al., 2010).

2.3.2. Indian Track Geometry Index (TGI)

Developed by Indian Railways, TGI is given as:
2UI + TI + GI + 6AI
TGI =
10 (7)

where UI ,TI ,GI , and AI are the indices for unevenness (surface), twist, gauge, and alignment respectively (Berawi et al., 2010).

2.3.3. Chinese track quality index

The Chinese track quality Index is simply a sum of all the standard deviations (σi ) for individual TQIs of each parameter.
7
TQI = ∑ σi
i=1 (8)

Later on, we examine the performance of each of these TQIs as a predictor for defects in track sections and propose an alternative.

3. Selected machine learning tools

In one of the earliest definition of machine learning, it was defined as a sub-field of computer science that teaches computer to
learn without being explicitly programmed (Munoz, 2014). It is the ability of a system to learn and improve (performance) from
experience. Machine learning is increasingly gaining relevance in many engineering disciplines and civil or railway engineering is not
an exception (Bergmeir et al., 2013; Li et al., 2014). Hence, machine learning is closely related to computational statistics which
focuses on predictive analysis leveraging on modern day computing capabilities. A common synonym found in statistical learning is
also sometimes called data mining if the motive is to find critical anomalies in data or to provide exploratory data analysis. This type
of learning is referred to as unsupervised. Supervised learning however deals with inferential or predictive analysis based on ob-
servable outcome (Mannila, 2016). In the next subsection, an exploration of the potentials of unsupervised learning in track geometry
data was examined with principal components analysis after which some resources of supervised learning were deployed for binary
classification based on the presence of geometry defects.

3.1. Dimension reduction

Another synonym for dimension reduction is feature extraction or selection. It is the idea that data can be ‘squeezed’ in such a way
that the variables can be reduced and replaced with principal dimensions (or selected features) without losing relevant information in
the data. Factor analysis, t-distributed stochastic neighbor embedding, principal components, etc., are popular examples of dimension
reduction techniques. It suffices that we focus only on principal components in this work since dimension reduction has hardly been
considered in railway engineering.

3.1.1. Principal component analysis

Principal Component Analysis (PCA) finds a low dimensional representation of a dataset such that this representation contains as
much variation in the dataset as possible (i.e. it aims to maximize the variance over a set of linear combinations). The basic
mathematical formulation of the principal component analysis is given as follows:
Let X ∈ Rn × p with rows x1,…,x n ∈ Rn × p . We think of X as n inspection dates, observations of a random vector X1,…,Xn ∈ Rn × p . Since
TQIs are often represented in terms of standard deviation, scaling is not necessary. Objective is to find a linear combination
w1 X1 + ,… ,+w2 X2 of parameters with maximum variance, i.e. a direction in Rp where X varies the most.
We solve:
n
w =‖argmax
w ‖2 = 1 ∑ (x iT w )2
i=1 (9)
n
where w is an eigenvector associated to the largest eigenvalue (λ ) of XTX . Since ∑i = 1 (x iT w )2
is proportional to the sample variance of
the data, each column (parameter) has a zero mean. The PCA was used for each section of track to summarize all features into 1, 2 or
3 principal components depending on the percentage of variance explained. As such, the first 3 principal components explain
3
(∑i = 1 λii )
p × 100 % of variance. In this work, the application of PCA as a TQI is similar to the idea of Pavement Condition Index (PCI) in
(∑i = 1 λii )
highways. Rather than assign weights to parameters based on expert opinion or subjective biases has seen in Section 2.3 above, the
PCA does so in a principled manner as described in Eq. (9). For application, thirty-one (31) feature dimensions were originally gotten
from track geometry cars and initially screened to eleven (11), and by using principal component analysis (PCA), we further reduced
it to three (3) predictors. Three (3) scores were selected according to the minimum number of principal components that explains at
least 90% of variance at any given section of track. Details are provided in Section 4.

234
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Fig. 2. A simple illustration of support vector machine/classifier (James et al., 2013).

3.2. Classification model selection

Geometry parameters on a given section of track are either within acceptable safety thresholds or otherwise. This is a classical
binary categorical problem since the outputs are discrete as opposed to being continuous or quantitative. A few of the common
classification methods are thus introduced:

3.2.1. Linear discriminant analysis

Linear discriminant analysis (LDA) is a generalization of Fisher’s discriminant method used in statistics, pattern recognition and
machine learning to find a linear combination of features that separates two or more classes of objects or events, in this case, track
geometry defects. The resulting combination may be used as a linear classifier, or, more commonly, for dimension reduction before
classification (James et al., 2013). LDA is also closely related to principal component analysis (PCA) and factor analysis in that they
both look for linear combinations of variables which best explain the data according to a defined objective (Martinez and Kak, 2001).

3.2.2. Trees and random forests

While bagging and random forests are closely related, bagging is the acronym for bootstrap aggregation and bootstrap samples of
a data are samples of the original data collected with replacement. Classification or regression methods like trees are trained on
bootstrap samples which are collected in such a way that these samples end up being highly correlated.
Random forests as the name indicates randomize the sampling process by randomly selecting only a few of the parameters for tree
selection and randomly selecting each of them for each splitting criterion at every step of a tree. It turns out that random forests are
usually uncorrelated and as such, reduce the risk of overfitting the training data. Random forests were trained on the track geometry
data and the results would be discussed in the next section.

3.2.3. Support vector machines

Support Vector Machines are based on the concept of hyperplanes that define decision boundaries. A decision plane is one that
distinguishes between a set of objects having different class labels (e.g. sections with defects or not). In Fig. 2 for example (Hastie
et al., 2009), the objects belong either to class GREEN (without defect) or RED1 (with at least one defect).
The separating dark line defines a boundary above which all sections of a track without defects are GREEN and below which other
sections with defects are RED. In the case of the left pane in Fig. 2 below, the data is separable. The right panel shows the non
N
separable case in which the points labeled ξi are on the wrong side of the margin. To maximize the margin, a total cost ∑i = 1 ξi ⩽ c .
A simple case of an SVM formulation is given as follows:
1 N
Minimize 2 wT w + c ∑i = 1 ξi subject to the constraints:
yi (wT ϕ (x i ) + b) ⩾ (1−ξi ) (10)

ξi ⩾ 0 (11)

for i = 1,…,N
where c is the trade-off between the margin and misclassification error, w is the vector of coefficients, b is a constant, and ξi represents
a slack parameter for handling non separable data. Support vector machine was used in this study and it proved to be a very efficient
classifier especially with the cost parameter tuning. One additional advantage is that the SVM is easy to visualize with 2 principal
components only. However, a Gaussian or radial basis kernel was employed rather than the traditional linear or polynomial decision
boundary. This is defined as follows:

1
For interpretation of color in Figs. 2 and 11, the reader is referred to the web version of this article.

235
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

(x i−x j )2
Kij = e−
σ (12)

where Kij is the kernel matrix, σ or c is a positive constant and x i,x j are pairs of training observations. The reason for choosing a radial
kernel becomes clear after the principal component analysis.

4. Case study

4.1. Data description

Dataset was collected from a US Class I railroad. 1 mile of track was selected with 28 inspection dates. Specific information on the
particular section of track has been withheld due to privacy. The dates and the given mile was chosen because it had a section that
underwent a major maintenance work. This section of track had a history of sub-grade failure which was later improved by the
installation of a geocell layer that gave it the required stability to support load transfer. The mile of track was divided into segments of
specific lengths (150 ft and 500 ft) over which the TQIs were computed. This can further be used to assess quality of track based on
the measurements and defects collected over the given segments of track over time (El-Sibaie and Zhang, 2014). It is expected that the
segment length or locations do not change for accurate monitoring of the uncertainties and variations in track. See Fig. 3 below for
variations in Gage and Surface TQIs before and after maintenance. We believe that the miscellanies expressed in the data collected
after and prior the major maintenance are varied enough to support some of the objectives of this research.
Data was in a matrix format for each inspection date with about 31 features (e.g. gage, cross level, alignment, surface and profile
including wavelength variations) for every foot of a mile track. After data cleansing and preprocessing, adjacent points of track were
grouped into sections. This paved the way for profile-based prediction rather than point-based analysis. We considered two section
lengths, 150 ft and 500 ft. Therefore, the 150 ft and 500 ft section lengths are equivalent to 35 and 11 sections respectively. For a
500 ft section, each inspection date is an 11 sections by 31 parameters matrix. The data was further processed into section matrices
with 28 inspection dates by 11 parameters as features.

4.1.1. Exploratory data analysis

In this section, we explore the track geometry data and try to understand the underlying properties before model training and
validation. Fig. 3 shows the surface values before and after a major maintenance. The high variations in Fig. 3a were corrected by the
installation of a geocell layer and therefore mitigated the propagation of surface defects. Fig. 3b shows the reduced variations. This is
altogether summarized in the box plots before and after maintenance, Fig. 3c.
It is interesting to see from the scatter plot matrix in Fig. 4 that the 62 ft wavelength variations of surface parameters have a
smaller range and perhaps more normally distributed than Gage and 124 ft Surface. This undermines the claim that there is a better
structure to data belonging to a larger section of track as opposed to smaller sections. While Warp, Cross level and Alignment have
been excluded in this plot for simplicity, their dependencies with other parameters are considered in Fig. 5. As shown in Fig. 5a, Warp
62 ft appears to be closely related to a longer variation of Alignment while Cross level and Alignment 62 ft showed strong mutual
dependencies. However a global assessment of the correlations across all the sections shown in Fig. 5b isolated Alignment 62 ft from
the group of Warp, Cross level and Alignment 124 ft. This is an indication that a momentary variation in track properties for a
particular section could be misrepresentative of overall track behavior.

4.1.2. Scatter and correlation plots

Before discussing the correlation plots, it has been established that Gage and alignment are horizontal parameters while surface,
cross level and warp are vertical (Soleimanmeigouni et al., 2016). Gage could also be seen as a transverse track feature while
alignment and are both longitudinal as well. Bearing this in mind, we consider Fig. 5a before b. The correlation plot for Section 6 of
the track isolates gage and suggests it to be mostly negatively correlated with other parameters especially cross level while exhibiting
almost zero correlation with the Surface Right (62 ft) parameter. While this behavior needs further investigation, Table 2 shows that
five distinct gage defects were picked-up in the same section. The inter-correlations observed between the longer wave-form (124 ft)
of surface and alignment on both rails appear to be stronger than the short wave counterpart (62 ft). For example, the correlation
between Surface Left (62 ft) and Surface Right (62 ft) is less than that between Surface Left (124 ft) and Surface Right (124 ft). This is
expected because a more profile-based indicator is achieved as the wavelength increases. Hence, shorter wavelengths like 31 ft on
both rails are even expected to be less correlated. The correlation plot ordered the parameters in increasing order of inter-correlations
and clustered them in groups from top-left to bottom-right. Hence, cross level was grouped with alignment (both long and short
wave) which traditionally measures horizontal deviations along the track. By definition, cross level measures the vertical distance
between two adjacent rails and since surface assesses vertical deviations in a single rail along the track, it is intuitive to assume that
cross level would correlate more with surface than alignment. The implication of this is that the longitudinal component (i.e. it is
measured along the track) of cross level is suppressed by the vertical feature (height between rails). Fig. 5a also shows that cross level
is correlated with warp but not as much as the short wave form of alignment on both rails.
However, Fig. 5b shows that some of the conclusions in Fig. 5a can be misleading which is why it is dangerous to generalize
indiscriminately. Hence, every section of track exhibits peculiar heterogeneity despite general observations. For the whole track in
Fig. 5b all surface wave parameters are closely correlated since they are vertical, Gage has been grouped with the short wave form of
alignment even though it was isolated from any cluster in Fig. 5a. It is obvious that Fig. 5b complies more with the correlation

236
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Fig. 3. Exploratory data: effect of geocell installation on gage and surface parameters.

groupings for horizontal and vertical parameters-i.e. Gage and alignment are horizontal parameters while surface, cross level and
warp are vertical Soleimanmeigouni et al. (2016)-than Fig. 5a. Other exploratory plots are shown with the principal components in
Figs. 6 and 7.

237
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Fig. 4. Exploratory data: scatter plot matrix (all inches).

4.1.3. Principal components’ analysis

Table 1 is a summary of the principal components that effectively summarize over 90% of the variation from eleven (11) track
geometry parameters for each Section (150 ft and 500 ft) length. Bold-faced texts denote sections summarized by only one principal
component. Parameters are not scaled since they already exist as standard deviation values with a general unit expressed in inches.
From Table 1, it is possible to conclude that the first two or three principal components is sufficient to explain majority of the
data; and in some cases, only the first principal component is enough. A closer look reveals that over 90% of the variance is explained
by 3 principal components in 100% of the 500 ft sections and 89% of the 150 ft sections. Therefore, with a focus on the 500 ft section
length above, a separation of sections with geometric defects from those without based on the Federal Railroad Administration
thresholds (FRA, 2002) was accomplished. A section with defect here is defined as any point (in feet) within a section that violates at
least one of the thresholds specified by the FRA for the given class of track (FRA, 2002). Table 2 is a summary of defects found, their
location and corresponding frequencies.

4.1.4. Principal components’ bi-plots

From Fig. 6, the number texts on the plots represent inspection dates (month and year) while arrows represent 1st 2 principal
components’ loading vectors. One would observe that longitudinal parameters (alignment, warp) have more loadings on the first
principal component while the second principal component loadings are driven by the vertical plane of track (surface). The bi-plot
also shows that warp and alignment are highly correlated as suggested by the correlation plot in Fig. 5b. Tending in a different
direction, gage remains the most isolated parameter than any other. In addition, it can be deduced that an inspection date in the
direction of a parameter has a huge variation of that parameter for that inspection date.

238
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

$OLgQPeQW5LgKWfW

$OLgQPeQW5LgKWfW
$OLgQPeQW/efWfW
SurfaFe5LgKWfW

$lLgQPeQW/efWfW
SurfaFe5LgKWfW
SurfaFe/efWfW

SurfaFe/efWfW

WarSfW

Crosslevel
Gage

Gage

SurfaFe5LgKWfW


SurfaFe/efWfW


SurfaFe5LgKWfW

SurfaFe/efWfW 

WarSfW 

$lLgQPeQW5LgKWfW í

$lLgQPeQW/efWfW
í

$lLgQPeQW5LgKWfW
í

Crosslevel
í
$lLgQPeQW/efWfW
í

(a) Sample correlation plot for track geometry parameters (Section 6)

$lLgQPeQW5LgKWfW

$lLgQPeQW5LgKWfW
$lLgQPeQW/efWfW

SurfaFe5LgKWfW

$lLgQPeQW/efWfW
SurfaFe5LgKWfW
SurfaFe/efWfW

SurfaFe/efWfW
WarSfW
Crosslevel

Gage


Crosslevel

WarSfW


$lLgQPeQW5LgKWfW


$lLgQPeQW/efWfW

SurfaFe5LgKWfW 

SurfaFe/efWfW 

SurfaFe5LgKWfW í

SurfaFe/efWfW
í

Gage
í

$lLgQPeQW5LgKWfW
í
$lLgQPeQW/efWfW
í

(b) Correlation plot for track geometry parameters considering all sections in miletrack
Fig. 5. Exploratory data: correlation plots.

4.1.5. Principal components’ score and scree plots

The principal components’ score plot is a Cartesian representation of the product of principal components’ loadings and the
original observations in the data.

239
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

í04 í02 00 02 04 í08 í06 í04 í02 00 02

V23 Gage

6
V12

04

2
V13

02
V23 V16

4
V24 V7
V9
V10
V15 V11
V8
V14
AL_124
AR_124

02
Warp

00
V17

0
V25 V24 V20
V21

2
V19V18
V22
V25 SL_124
V12 V26 Xlevel AL_62
V16V13 V27
PC2

PC2
V15V14 V1
SR_124

í02
V28
Xlevel V26 V6

00
Gage V5
0 AL_62
V17 AL_124 V3V4

í2
V22V18
V8V20
V21 V19 Warp
AR_62 V7
V9 AR_124 V2
V2 V10
V11
V3
V5
V6
V4

í04
SL_62
í2

V27

í02
V1
AR_62

í4
V28 SR_62

í06
í4

í04
SR_62
SL_62
SL_124
í6

í08
í6
SR_124

í6 í4 í2 0 2 4 6 í6 í4 í2 0 2
PC1 PC1

Fig. 6. Biplots for a defective Section, 2 (left) and non-defective section, 5 (right).

Fig. 7. Principal component score plots for 1st 2nd PCs (left) and 1st and 3rd PCs (right).

Fig. 7 is a sample plot for sections with defects. It is clear from the score plot that the decision boundary for sections with defect is
not linear. We therefore hypothesize that a machine learning procedure capable of effectively separating sections with defects from
those without would require an elliptical/circular decision boundary if the principal component scores are not to be mapped onto a
higher dimensional space. To test this hypothesis, we train selected models in Section 4.2 below.
Scree plot samples for sections in both 150 ft and 500 ft lengths give elbows at two principal components as shown in Fig. 8. It is
also obvious that the first two principal components summarize about 90% of the data. As such, it is proposed that multidimensional
track geometry parameters could be effectively expressed as a bi-variate or tri-variate data.

4.2. Model training

Altogether, three sections were found to have at least one defect going by the 500 ft segments (Sections 5, 6 and 11). Afterwards, a
binary response variable was created based on presence of defect in a section. The number of observations (rows) are 28 inspection
dates for each of the Sections 5, 6 and 11 vertically stacked into a total of 84 inspections (28 for each section). This stacking was
necessary to provide enough observations for the model training and testing. The binary classification problem to be solved is
expressed as follows:

y = f (x ) (13)

where

240
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Table 1
Summary of principal components for each section.
Section Length = 150 feet, 35 Sections Section Length = 500 feet, 11 Sections

Section PCs, Variance Sum > 90% %Variance 1st PC (%) Section PCs, Variance Sum > 90% %Variance 1st 2PCs (%)

1. 1ST, 2ND & 3RD 94.68 50.48 1 1ST & 2ND 91.50 91.50
2. 1ST, 2ND & 3RD 95.38 51.25
3. 1ST, 2ND, 3RD & 4TH 94.34 41.49

4. 1ST 92.74 92.74 2 1ST 95.60 99.10

5. 1ST 94.13 94.13
6. 1ST & 2ND 96.67 86.60
7. 1ST, 2ND & 3RD 92.11 63.21

8. 1ST 94.00 94.10 3 1ST & 2ND 96.23 96.23

9. 1ST & 2ND 94.00 85.40
10. 1ST & 2ND 94.37 67.20

11. 1ST & 2ND 91.20 59.97 4 1ST 91.60 95.48

12. 1ST & 2ND 92.67 86.54
13. 1ST & 2ND 92.10 80.24

14. 1ST & 2ND 97.00 81.56 5 1ST, 2ND & 3RD 97.61 89.05
15. 1ST 91.79 91.79
16. 1ST, 2ND, 3RD & 4TH 95.70 62.84
17. 1ST & 2ND 92.45 81.06

18. 1ST & 2ND 90.48 59.10 6 1ST, 2ND & 3RD 94.96 85.40
19. 1ST, 2ND & 3RD 93.97 69.25
20. 1ST, 2ND & 3RD 95.49 55.17

21. 1ST, 2ND, 3RD, 4TH & 5TH 93.56 54.30 7 1ST, 2ND & 3RD 96.14 88.57
22. 1ST, 2ND & 3RD 92.93 74.81
23. 1ST, 2ND & 3RD 97.16 61.00

24. 1ST 93.20 93.20 8 1ST & 2ND 94.38 94.38

25. 1ST 95.43 95.43
26. 1ST & 2ND 95.55 81.55
27. 1ST, 2ND & 3RD 97.14 74.74

28. 1ST, 2ND & 3RD 94.20 71.56 9 1ST, 2ND & 3RD 98.00 86.28
29. 1ST & 2ND 97.55 87.61
30. 1ST & 2ND 90.48 65.66

31. 1ST, 2ND, 3RD & 4TH 96.84 50.73 10 1ST 91.30 97.33
32. 1ST & 2ND 96.21 85.91
33. 1ST 91.70 91.70

34. 1ST, 2ND & 3RD 92.90 52.99 11 1ST & 2ND 95.52 95.52
35. 1ST & 2ND 98.10 88.83

⎛ 0−no defects ⎞
⎜ 1−yes defect ⎟
y=⎜ : ⎟
⎜ : ⎟
⎝ 0−no defects ⎠ (14)

x x . . ,x1p
⎛ 11 12 ⎞
x= : : :
⎜x x . . ,x ⎟
⎝ n1 n2 np ⎠ (15)

The response variable above, y is a binary type where 1 or “Yes/True” stands for the presence of defect and 0 or “No/False” is no
defect. The predictor variables, x, provided from the Class I railroad (operating on FRA-Class 7 track) is about thirty parameters, some
of which are repeated due to varying wavelengths. However, other parameters like “Curve” and “Speed” are not TQIs while others
like “Cross level rate” were dropped in place of Warp 62 ft. After eliminating further redundancies, the selected TQIs (explanatory
variables) are as follows: Gage, Crosslevel, Surface Right 62 ft, Surface Right 124 ft, Surface Left 62 ft, Surface Left 124 ft, Alignment
Right 62 ft, Alignment Right 124 ft, Alignment Left 62 ft, “Alignment Left 124 ft and Warp 62 ft. Therefore, after having identified
which inspection dates had at least one defect according to the FRA thresholds (FRA, 2002), we employed machine learning tools to

241
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Table 2
Defects, section found and count.
S/No Parameters for Class 7 Track Sections with defects (500 ft) Location in feet(s) Count(s) Total

1. Gage Section 6 2967–2971 5 5

2. Alinement 62 ft NO NO 0
DEFECTS
3. Alinement 124 ft NO NO 0
DEFECTS DEFECTS
4. Crosslevel Section 5 2056–2057 2 70
Section 5 2108–2109 2
Section 11 5194–5263 61
Section 11 5233–5234 2
Section 11 5232–5234 3
5. Surface 62 ft NO NO 0
DEFECTS DEFECTS
6. Surface 124 ft NO NO 0
DEFECTS DEFECTS
7. Warp 62 ft NO NO 0
DEFECTS DEFECTS

TOTAL 75

Fig. 8. Variance (LHS) and Cumulative Variance (RHS) explained by principal components for sample sections in 150 ft (above) and 500 ft (below)
lengths.

learn from this dataframe and be able to predict should there be a new observation with a defect. The purpose of this is to select
which method/model works best before deploying it to validate principal components as artificial TQI.
The classification methods trained on these stacked sections are three, two of which are parametric (Linear Discriminant Analysis
and Support Vector Machine) while the other is non-parametric (Random Forest). The sections were combined so as to enable the
learning tool recognize sufficient defect pattern for better prediction. To evaluate the accuracy of the predictors, we employ two
measures: True positive rate (TPR) and False positive rate (FPR), the former measures the proportion of inspection dates with defects
in a section that are actually identified while the latter identifies the proportion of dates identified with defects that are not true. Our
focus here is to be able to identify defects and that’s why the TPR is our utmost priority. The FPR can be seen as a false alarm.
Based on Table 3, it is interesting to see that the support vector machine overfits the training data but performs almost at par with

Table 3
Error rates for selected training models.
S/No LEARNING TOOL TRAINING (%) TESTING (%)

Model evaluation TPR FPR ALL TPR FPR ALL

1.0 Linear Discriminant Analysis (LDA) 100 2.60 97.62 66.67 7.69 90.48
2.0 Support Vector Machine (SVM) 100 0.00 100 66.67 0.00 97.62
3.0 Random Forest 66.67 1.15 92.70 0.00 2.63 92.86

242
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Fig. 9. Model performance of selected machine learning tools.

the LDA on the test sample. This overfitting can be attributed to the limited observations in the training data. And because the
percentage of dates with defect is very low, it is possible for the training algorithm to classify with maximum accuracy. The SVM
performed woefully initially but after tuning out the parameters c (cost of constraint violation-regularization term) and gamma
(required by all kernels except a linear decision boundary) for a radial kernel through a ten-fold cross validation, it turned out with a
better classifier. Being non-parametric, the random forest could not spot any section with defects on the TPR for the test data even
with a thousand trees. A possible reason here is because it could not effectively learn from the data structure since the random trees
are uncorrelated. Hence, the SVM with a zero FPR on both training and testing data has been selected for model validation with PCA
and other TQIs listed earlier in Section 2. Fig. 9 provides a simple illustration of the model performance for the selected machine
learning tools.
Given that track sections with defects are significantly smaller than non-defective sections, the binary classification problem
solved is very highly unbalanced. Hence, an overall misclassification rate may not be appropriate, the more reason why the TPR has
been proposed for validation.

(a) Random Forest Variable Importance Plot

(b) SVM Classifier with two Principal Components

Fig. 10. Trained model outputs: random forest and support vector machine.

243
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Table 4
Prediction performance for selected TQIs with SVM.
Sections Alla J-Coeffc ITGIa CHTQIb CNTQIa 1st 3 PCsa

TPR FPR TPR FPR TPR FPR TPR FPR TPR FPR TPR FPR

Section 11 100 0.00 67.00 0.00 66.67 0.00 100 0.00 100 0.00 100 0.00
Section 5 100 0.00 50.00 96.15 0.00 0.00 50.00 0.00 100 0.00 100 0.00

a
cost = 3.4, gamma = 1.3.
b
cost = 4.3, gamma = 1.0.
c
cost = 2.9, gamma = 1.0.

Fig. 10a shows the important variable critical to defect classification in order of importance. The cross entropy and the Gini index
have been used as the node impurity measures in this case (Hastie et al., 2009). SVM classification plot-for a sample section with
defects using just two principal components in Fig. 10b-enunciates the fact that the defect decision boundary for a track geometry
data may not be linear, rather the plot suggests it to be radial or elliptical. Hence, any classification tool for track geometry defects
has to take cognizance of this to ensure accurate prediction.

4.3. Model validation

To validate the possibility of identifying sections with defects using principal components as predictors, we trained support vector
machine on 2 sections with defects using all the parameters, J-synthetic coefficient, the Indian Track Geometry Index (ITGI), Chinese
track quality index (CHTQI) and the Canadian National rail TQI (CNTQI). TPR and FPR for the corresponding parameters are
recorded in Table 4.
First of all, it is important to note that there are no false-positives except with J-Coefficient in Section 5, the table also shows that 3
principal components and the CNTQI are sufficient substitutes for all 11 parameters as predictors. While J-coefficient and CHTQI
could not do better than a random guess on the TPR in Section 5, the ITGI’s performance is surprisingly poor at the same section, even
with zero FPR. A 100% TPR from most of the predictors is nothing but a consequence of over-fitting. While the results are quite
encouraging, the SVM is not guaranteed to perform this incredibly on every new data set despite using a radial kernel. Another
explanation for this phenomenon is because of the limited dataset. The radial kernel successfully wraps around the inspections with
defects contained in the test set as shown in Fig. 10. Plots of the principal components, TQIs as a function of time (inspection dates) as
well as a correlation plot between the different artificial TQIs are given in Figs. 11 and 12.

4.4. Discussion

In this section, we discuss the output features of the artificial TQIs, how they vary with the presence of defects, time, and whether
they represent horizontal, vertical and transverse track properties based on the model results. We conclude with a discussion on the
principal component scores and how they represent most of the original track parameters than most of the artificial TQIs discussed in
Section 2. From Fig. 11, the red dots in each plot indicates the corresponding TQIs for an inspection date featuring at least one track
geometry defect. For the principal components, the plot suggests that the scores maintained a relatively smooth profile until track
geometry defects on the highlighted dates were observed. This observation also holds for the other TQIs as shown except for the
CNTQI and CHTQI. While the J-coefficient and the Indian TGI had hitherto maintained similar profile, their graphs crossed at
inspection date ‘1115’. However, the same inspection date featured a spike for the CNTQI and CHTQI despite the fact that they are
profile-based rather than point-based observations as explained in Section 4. These spikes suggest an indication for outliers that could
be further investigated. A closer look at this date revealed that certain cross level defects were picked up. With the principal com-
ponents scores in Fig. 12a, we observe a trend similar to those of the original single TQIs in Fig. 13. Especially with surface and the
first principal components. The correlation plots in Fig. 11 shows the salient dependencies between the mix of track geometry
parameters and principal components. The plot was ordered by hierarchical clustering based on a package developed on the R
programming statistical software (Wei and Simko, 2016). The most-correlated parameters are ordered from top-left to bottom-right in
decreasing order of interdependence.
Fig. 11, we observe that the first principal component can be said to represent vertical track parameters as it is closely positively
correlated to both surface wavelengths, buttressing the findings in previous studies that surface parameters are sufficient to assess
ride quality (Audley and Andrews, 2013; Khouy, 2013). Secondly, the second principal components explains longitudinal variations
of the track parameter which is seen from its close correlation to both waveforms of alignment on both rails. As for the combined
TQIs, PC2 from Fig. 11 suggests a strong correlation with CNTQI but negative correlations with the Chinese TQI, Polish J-coefficient
as well as the Indian TGI. This finding implies that while the afore-stated railways might have used their respective TQIs to measure
ride quality, the TQIs actually emphasize longitudinal deviations. Hence, their use should be appropriately channeled, perhaps to
monitor longitudinal degradation. What more, the third principal component appears to be a measure of transverse irregularities
(Gage and cross level), see Fig. 11.

244
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

S8)_R_124

S8)_L_124
ALI_R_124

S8)_R_62
ALI_L_124

S8)_L_62
WARP_62
ALI_R_62

ALI_L_62

XLEVEL
J_Coeff
CNTQI

CHTQI
GAGE
PC2

PC3
TGI

PC1
1
CNTQI

PC2
08
ALI_R_124

ALI_L_124 06
J_Coeff

TGI 04

ALI_R_62
02
ALI_L_62

GAGE
0
PC3

XLEVEL
í02
CHTQI

WARP_62 í04
S8)_R_62

S8)_L_62 í06

PC1
í08
S8)_R_124

S8)_L_124
í1

Fig. 11. Hierarchical correlation plot for all artificial TQIs.

We can therefore summarize the 3D properties of track geometry parameters with 3 principal components with the first being
vertical, second being longitudinal and the third principal component represents transverse irregularities. With this, we propose that
the use of 3 principal components can simplify track geometry parameters instead of dealing with the numerous explanatory vari-
ables and wavelength variations (sometimes up to 30) often rolled out by track geometry cars. Traditionally, most of the combined
TQIs used by the railways selected for this study selects about 7 of these variables and neglect the rest due to subjective reasons. But
with principal components we can still keep most of the variance and the three dimensional irregularities in our track geometry data
with comparatively higher objectivity and much simplicity.

5. Concluding remarks

This paper formally described the work on principal components and track quality indices. To summarize heterogeneous track
geometry data, some railways assign weights to selected track geometry parameter. This assignment is followed by the sum of all the
products of the weights and the parameters to arrive at a value that is used as a measure of overall track quality. While the assigned
weights are often subjective, the parameters selected vary from one railway to the other. Also, relevant information is lost through
neglected parameters and subjective weight assignment. In order to prevent this, the use of principal components as combined TQIs
was proposed in this work. This made it possible to simplify track geometry data in a way that most of the variance in the data is
captured.
The use of principal components as TQIs was tested using classical machine learning algorithms and the following conclusions are
highlighted: (1) With a sample mile track, ⩾ 90% variance in the geometry data was explained by 1st 3 components in 100% of 500 ft
sections and 88% of 150 ft sections. The first principal component captured track variations in the vertical plane, the second principal
component in the longitudinal plane and the third correlated well with transverse irregularities. This information can be used to plan
maintenance activities such as tamping or stone-blowing (1st PC and 2nd PC) as well as gage correction (3rd PC). (2) Support Vector
Machine (SVM) was the most effective learning tool for classifying track sections with geometry defects among other selected ma-
chine learning tools; and (3) Using principal components and other combined TQIs from different railways, SVM predicted track
defects better with 3 principal components and Canadian TQI than any other TQIs considered in this study. The prediction perfor-
mance was measured using TPR (True Positive Rate) and FPR (False Positive Rate) since the defect data is highly unbalanced.
This approach will help railways and track engineers assess track geometry monitoring from a different perspective as a novel

245
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

(a) Principal Components versus Inspection Date

(b) Polish J-Coefficent and Indian TGI versus Inspection Date

(c) Chinese TQI versus Inspection Date

(d) CN Rail TQI versus Inspection Date

Fig. 12. Various track quality indices as a function of inspection dates.

method of combined/artificial TQI for maintenance scheduling. This work is a first step in incorporating dimension reduction in track
geometry data analysis using classical techniques. Future work will consider the development of thresholds for principal components
through correlation studies with vertical or lateral accelerations on train; and the use of classifier fusion to obtain better predictions.
Because dimension reduction/feature extraction with machine learning have not been widely adopted in track geometry data and
analysis, there is great potential for optimized maintenance scheduling under this approach.

246
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Fig. 13. Timeline variations of selected parameters in given sections.

Acknowledgment

The authors would like to acknowledge the University of Nevada, Las Vegas (UNLV/USDOT) University Transportation Center
(UTC) on Improving Rail Transportation Infrastructure Sustainability and Durability for partly sponsoring this research.

References

Audley, M., Andrews, J., 2013. The effects of tamping on railway track geometry degradation. Proc. Inst. Mech. Eng., Part F: J. Rail Rapid Transit 227 (4), 376–391.
Bai, L., Liu, R., Sun, Q., Wang, F., Wang, F., 2016. Classification-learning-based framework for predicting railway track irregularities. Inst. Mech. Eng., Part F: J. Rail
Rapid Transit 230 (2), 598–610.
Berawi, A.R.B., Delgado, R., Calçada, R., Vale, C., 2010. Evaluating track geometrical quality through different methodologies. Int. J. Technol. 1 (1), 38–47.
Bergmeir, C., Sáinz, G., Martínez Bertrand, C., Benítez, J.M., 2013. A study on the use of machine learning methods for incidence prediction in high-speed train tracks.
In: Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) 7906 LNAI, pp. 674–683.
Ciobanu, C., 2016. Evaluation of the Track Quality. < https://pwayblog.com/2016/09/11/evaluation-of-track-quality/?blogsub=confirming#blog_
subscription-3 > .
DYK, B.J.V., 2014. Characterization of the Loading Environment for Shared-Use Railway Superstructure in North America (Ph.D. Thesis). University of Illinois Urbana
Champaign.
El-Sibaie, M., Zhang, Y.-J., 2014. Objective track quality indices. J. Transp. Res. Board 1863. <http://trrjournalonline.trb.org/doi/pdf/10.3141/1863-11> .
Esveld, C., 2016. Modern Railway Track, second ed. MRT-Productions.
Fazio, J.L.C., 1986. Track quality index for high speed tracks. Transp. Eng. 1 (112), 46–61.
Fernando, L., Hart, J., Al., E., 2011. Machine Vision Inspection of Railroad Track. Tech. Rep., University of Illinois Urbana-Champaign.
FRA, 2002. Track Safety Standards Compliance Manual. Track and Rail and Infrastructure Integrity Compliance Manual 4.
FRA, 2005. Development of Objective Track Quality Indices. US Department of transportation Federal Railroad Administration Research Results, 6–9. < https://www.
fra.dot.gov/Elib/Document/2146 > .
FRA, 2013. Track Safety Standards; Improving Rail Integrity; Final Rule. < https://www.gpo.gov/fdsys/pkg/FR-2014-01-24/pdf/2014-01387.pdf > .
FRA, 2015. FRA Office of Safety Analysis. < http://safetydata.fra.dot.gov/officeofsafety/default.aspx > .
Hastie, T., Tibshirani, R., Friedman, J., 2009. The Elements of Statistical Learning, vol. 1 Springer. <http://www.springerlink.com/index/10.1007/b94608> .
Hay, W.W., 1982. Railroad Engineering, second ed. Wiley.
Hendry, M., 2015. JRC2015-5654. Joint Rail Conference, 1–7.
Hu, C., Liu, X., 2016. Modeling track geometry degradation using support vector machine technique. In: The ASME Joint Rail Conference, 2016. Article Number: V
001T01A011 (JRC2016-5739).
Ishwarappa, Anuradha, J., 2015. A brief introduction on big data 5Vs characteristics and hadoop technology. Procedia Comput. Sci. 48 (C), 319–324.
James, G., Witten, D., Hastie, T., Tibshirani, R., 2013. An Introduction to Statistical Learning, first ed. Springer. <http://www-bcf.usc.edu/∼gareth/ISL/
ISLRFirstPrinting.pdf> .
Jamshidi, A., Faghih-Roohi, S., Hajizadeh, S., Núñez, A., Babuska, R., Dollevoet, R., Li, Z., De Schutter, B., 2017. A big data analysis approach for rail failure risk
assessment. Risk Anal. 37 (8), 1495–1507.

247
A. Lasisi, N. Attoh-Okine Transportation Research Part C 91 (2018) 230–248

Karim, F.M.A., Rubasi, K.A.H., Saleh, A.A., 2016. The road pavement condition index (PCI) evaluation and maintenance: a case study of Yemen. Organ., Technol.
Manage. Construct.: Int. J. 8 (1), 1446–1455.
Kelshaw, M.T.T.-L.Y.R., 1995. Developments and applications in the Amtrak track geometry measurement car. Nondestructive Evaluation of Aging Railroads 2458.
< http://spie.org/Publications/Proceedings/Paper/10.1117/12.212678 > .
Khouy, I.A., 2013. Cost-Effective Maintenance of Railway Track Geometry, A Shift from Safety Limits to Maintenance Limits (Ph.D. Thesis). Luleå University of
Technology.
Li, H., Parikh, D., He, Q., Qian, B., Li, Z., Fang, D., Hampapur, A., 2014. Improving rail network velocity: a machine learning approach to predictive maintenance.
Transp. Res. Part C: Emerg. Technol. 45, 17–26.
Li, Z., Zhao, X., Esveld, C., Dollevoet, R., Molodova, M., 2008. An investigation into the causes of squats—correlation analysis and numerical modeling. Wear 265
(9–10), 1349–1355.
Liu, R.K., Xu, P., Sun, Z.Z., Zou, C., Sun, Q.X., 2015. Establishment of track quality index standard recommendations for beijing metro. Discr. Dyn. Nat. Soc.
Madejski, J., Grabczyk, J., 2002. Continuous geometry measurement for diagnostics of tracks and switches. In: Proceedings of the International Conference on
Switches.
Mannila, H.., 2016. Data mining: machine learning, statistics, and databases. In: Int’l Conf. Scientific and Statistical Database Management. IEEE Computer Society.
Martinez, A.M., Kak, A.C., 2001. PCA versus LDA. IEEE Trans. Pattern Anal. Mach. Intell. 23 (2), 228–233.
Molodova, M., Li, Z., 2014. Automatic detection of squats in railway infrastructure maria. IEEE Trans. Intell. Transp. Syst. 15 (5), 1980–1990.
Munoz, A., 2014. Machine Learning and Optimization. Courant Institute of Mathematical Sciences. <https://www.cims.nyu.edu/∼munoz/files/ml_optimization.
pdf> .
Odlyzko, A., 2016. The early British railway system, the Casson counterfactual, and the effectiveness of central planning. Essays Econ. Bus. History 34 (1), 60–94.
Pedankar, N., Rao, B.V., Kapur, P., 2003. Profile-based prediction: an enhanced methodology for prediction of track degradation. In: American Railway and
Maintenance of Way Association (AREMA) 2003 Conference Proceedings, pp. 1543–1548.
Peng, F., 2011. Scheduling of Track Inspection and Maintenance Activities in Railroad Networks (Ph.D. Thesis). University of Illinois at Urbana-Champaign.
Scientific American, 2016. Broken Rails Are Leading Cause of Train Derailments. < http://www.scientificamerican.com/article/broken-rails-are-leading-cause-of-
train-derailments/ > .
Soleimanmeigouni, I., Ahmadi, A., Kumar, U., 2016. Track geometry degradation and maintenance modelling: a review. Proc. Inst. Mech. Eng., Part F: J. Rail Rapid
Transit 0 (0), 1–30.
Statista, 2017. Transportation and Logistics, Development of Metro Networks Worldwide. < http://www.statista.com/statistics/619184/metro-networks-worldwide-
track-distribution/ > .
Thompson, L.S., 1989. Railways In Eastern Europe. Tech. Rep., The World Bank.
Wei, T., Simko, V., 2016. An Introduction to Corrplot Package. < https://cran.r-project.org/web/packages/corrplot/vignettes/corrplot-intro.html > .
Woodburn, A., 2014. Development of rail freight in Europe: What regulation can and cannot do United Kingdom Case Study. CERRE: Center Regul. Europe 32
(December), 1–43.
Xu, P., Liu, R., Sun, Q., Wang, F., 2012. A novel short-range prediction model for railway track irregularity. Discr. Dyn. Nat. Soc. http://dx.doi.org/10.1155/2012/
591490.
Yella, S., Pasha, S.M., Dougherty, P.M., 2009. Classifier fusion for condition monitoring of wooden railway sleepers. In: Systems, Man and Cybernetics, 2009. SMC
2009. IEEE International Conference on 1 (May), pp. 2–4.
Zarembski, A.M., 2015. Some examples of big data in railroad engineering. In: Proceedings – 2014 IEEE International Conference on Big Data, IEEE Big Data 2014, pp.
96–102.
Zarembski, A.M., Grissom, G.T., Euston, T.L., 2013. Use of ballast inspection technology for the prioritization, planning and management of ballast delivery and
placement. In: American Railway Engineering Association Annual Conference.

248

View publication stats