Seminar Report on

WHEEL DEFECT DETECTION WITH MACHINE

LEARNING
Submitted by

ABHIMANUE K S (NCE18MC001)
in partial fulfillment of the requirements for the award of the degree
of BACHELOR OF TECHNOLOGY in Mechatronics Engineering

Guide

Mr. NIVETH L
Assistant Professor
Department of Mechatronics Engineering

DEPARTMENT OF MECHATRONICS ENGINEERING

NEHRU COLLEGE OF ENGINEERING AND RESEARCH CENTRE
THRISSUR-680597

DECEMBER 2021
 DEPARTMENT OF MECHATRONICS ENGINEERING
NEHRU COLLEGE OF ENGINEERING AND RESEARCH CENTRE
THRISSUR–680597

Certificate

This is to certify that this report entitled “WHEEL DEFECT DETECTION

WITH MACHINE LEARNING” is a bona fide record of the seminar
presented by ABHIMANUE K S(NCE18MC001) towards the partial fulfillment
for the award of Degree of Bachelor of Technology in Mechatronics
Engineering under APJ Abdul Kalam Technological University, Kerala, during
the year 2021.

Guided By Head of Department

Mr Niveth L Dr Bobby ND
(Assistant Professor) (Professor)
Dept. Of Mechatronics Dept. Of Mechatronics
Engineering Engineering
Seminar Report Wheel Defect Detection With Machine Learning

DECLARATION

I undersigned hereby declare that the seminar report “WHEEL DEFECT

DETECTION WITH MACHINE LEARNING”, submitted for partial
fulfilment of the requirements for the award of degree of Bachelor of
Technology of the APJ Abdul Kalam Technological University, Kerala, is a
Bonafede work done by me under supervision and guidance of Mr. NIVETH L,
Asst. Prof Mechatronics Department. This submission represents my ideas in
my own words and where ideas or words of others have been included, I have
adequately and referenced the original sources. I declare that I have adhered to
ethics of academic honesty and integrity and have not misrepresented or
fabricated any data or idea or fact or source in my submission. I understand that
any violation of the above will be a cause for disciplinary action by the institute
and/or the University and can evoke penal action from the sources which have
thus not been properly cited or from whom proper permission has not been
obtained. This report has not been previously formed the basis for the award of
any degree, diploma or similar title of any other University

Place: PAMPADY ABHIMANUE K S

DATE: NCE18MC001

Department of Mechatronics, NCERC

ii
Seminar Report Wheel Defect Detection With Machine Learning

ACKNOWLEDGEMENT

My endeavor stands incomplete without dedicating our gratitude to everyone

who has contributed a lot towards the successful completion ofour project work.
First of all, I offer our sincere thanks to our parents for their blessings. I am
indebted to God Almighty for blessing us with His grace and taking our
endeavor to a successful culmination. I submit this project work at the lotus feet
of Late Dr. P. K. Das, Founder Chairman, Nehru Educational and Charitable
Trust. We express our profound gratitude to Adv. Dr. P. Krishnadas,
Managing Trustee, Nehru Educational and Charitable Trust. I would also like to
express my sincere thanks to Dr. P. Krishnakumar, CEO & Secretary, Nehru
Educational and Charitable Trust. I am really indebted to all the staff and faculty
members of our college for all helps they have extended to us. I am also grateful
to Prof. Dr. Ambikadevi Amma.T, Principal of our college, for supporting us
all along. I express sincere thanks to Mrs. Bobby ND, Head of the Department,
Mechatronics for giving us the correct format to write this report and correcting
the same. I express our sincere thanks and gratitude to, Assistant Professor,
Mechatronics for the guidance given to us and steering us to the successful
completion of this project. I finally, thank my friends and all our well-wishers
who had supported us directly and indirectly during my seminar.

ABHIMANUE K S
NCE18MC001
S7 MECHATRONICS
NCERC

Department of Mechatronics, NCERC

iii
Seminar Report Wheel Defect Detection With Machine Learning

ABSTRACT

Wheel defects on railway wagons have been identified as an important source of damage
to the railway infrastructure and rolling stock. They also cause noise and vibration
emissions that are costly to mitigate. We propose two machine learning methods to
automatically detect these wheel defects, based on the wheel vertical force measured by a
permanently installed sensor system on the railway network. Our methods automatically
learn different types of wheel defects and predict during normal operation if a wheel has a
defect or not. The first method is based on novel features for classifying time series data
and it is used for classification with a support vector machine. To evaluate the performance
of our method we construct multiple data sets for the following defect types: flat spot,
shelling, and non-roundness. We outperform classical defect detection methods for flat
spots and demonstrate prediction for the other two defect types for the first time.
Motivated by the recent success of artificial neural networks for image classification, we
train custom artificial neural networks with convolutional layers on 2-D representations of
the measurement time series. The neural network approach improves the performance on
wheels with flat spots and non-roundness by explicitly modeling the multi sensor structure
of the measurement system through multiple instance learning and shift invariant
networks.

Department of Mechatronics, NCERC

iv
Seminar Report Wheel Defect Detection With Machine Learning

CONTENTS

ACKNOWLEDGEMENT ................................................................................ iii

ABSTRACT ........................................................................................................ iv

Chapter 1 .............................................................................................................. 7

INTRODUCTION ............................................................................................... 7

Chapter 2 .............................................................................................................. 9

LITERATURE SURVEY ................................................................................... 9

Chapter 3 ........................................................................................................... 17

MEASUREMENT SYSTEM AND DEFECT TYPES .................................. 17

Wheel Load Checkpoint ................................................................ 17

Railway Wheel Defects ................................................................. 18

Chapter 4 ........................................................................................................... 20

TIME SERIES REPRESENTATION FOR DEFECT DETECTION ........ 20

Chapter 5 ............................................................................................................ 21

AUTOMATIC REPRESENTATION LEARNING ....................................... 21

Chapter 6 ............................................................................................................ 24

CLASSIFICATION OF WHEEL DEFECTS ................................................ 24

Chapter 7 ............................................................................................................ 25

DATA SETS AND MODELS ........................................................................... 25

Chapter 8 ............................................................................................................ 27

EXPERIMENTAL RESULTS ......................................................................... 27

Chapter 9 ............................................................................................................ 29

Department of Mechatronics, NCERC

v
Seminar Report Wheel Defect Detection With Machine Learning

CONCLUSION .................................................................................................. 29

Chapter 10 .......................................................................................................... 30

REFERENCES .................................................................................................. 30

Department of Mechatronics, NCERC

vi
 Wheel Defect Detection With Machine Learning
Seminar Report

Chapter 1
INTRODUCTION
Early detection of serious wheel defects on freight trains are an essential part in preventing
damage to the railway infrastructure and in providing the train operators with timely
information on necessary repairs, that can prevent further deterioration of the wheels.
Wheel defects of railway vehicles directly cause an increase in attrition of and damage to
the railway infrastructure, e.g., the track systems or the civil engineering works, thereby
adding additional costs to maintenance and repair and leading to a reduced lifetime and
availability of rolling stock. The life span of the railway infrastructure is significantly
shortened by the negative effects of wheel defects. The life span of railway bridges for
instance is calculated with an assumed maximal dynamical load of 21 tons. Due to wheel
defects the actually occurring dynamical load can be up to 50 tons, or 270% higher than
the theoretically assumed maximum, thus shortening the life span. Wheel defects also
accelerate crack-growth on the rail tracks and lead to premature failure of the rail system.

Another important effect caused by wheel defects are ground vibration and noise
emissions. In the European Union (EU) Project “Railway Induced Vibration Abatement
Solutions” (RIVAS)1 27 partners from nine countries investigated the source and
mitigation measures for noise and vibration emissions. They found that reducing wheel
defects by wheel maintenance significantly reduces vibration and noise emissions directly .
Therefore, it is recommended to use timely and targeted maintenance of train wheels as an
economic means to reduce emissions . This measure is all the more important as the
density and usage of modern railway networks is steadily increasing and failures quickly
disrupt operation of the whole network or parts of it. Since 2008, all states in the EU are
advised to employ noise emission ceilings. Switzerland started a noise abatement program
based on emission ceilings that requires the infrastructure manager to curb emissions
above the ceiling. This abatement programme leads to total costs of 1.5 billion.

In this paper we propose a method of detecting defective wheels. This classification

method promises to increase the reliability of the railway infrastructure, to reduce the cost
of freight train operation and to save additional investments on noise protection measures.
To reach this goal without the costly construction of further measurement sites or newly
built sensors, we propose the use of statistical methods that allow us to automatically
inspect the existing data and extract the information about defective wheels that is already
present.

Our proposed methods do neither require a model of the measurement system, nor of train
dynamics or wheel defects. The methods enable us to predict defects on wheels where
there is no prior understanding of how these defects manifest themselves in the
measurements. The methods detect and classify different types of defects based on
measurements during normal operation where the trains pass the measurement sites in full
operational speed. The features that we have developed for the use in supervised learning
are general and can in principle be used for any time series data and are not restricted to

Department of Mechatronics, NCERC Page 7

 Wheel Defect Detection With Machine Learning
Seminar Report

specific defect types. In a second step we automatically learn features directly from the
raw measurement signal.

A. Contribution
Our main contribution are two methods for automatic railway wheel defect detection and
classification through vertical force measurements of trains running in full operational
speed. For the first method we design novel wavelet features for time series data from
multiple sensors and we learn a classifier using a support vector machine. For the second
method we design and train convolutional neural networks for different wheel defect types
by deep learning. We evaluate our novel and other classical methods for wheel defect
detection on two labeled data sets with different types of wheel defects, that we have
constructed from calibration runs and from maintenance reports.

B. Related Work
While there has been research on machine learning methods for railway track inspection
or condition based maintenance , to our knowledge machine learning methods for railway
wheel defect detection have not been developed so far. There has been some research on
sensor systems for wheel defect detection on freight trains. Nenov et al. analyses the signal
from acceleration sensors and demonstrate that they can visually see a difference between
the measurements of wheels with flat spots and good wheels but they do not propose a
method for detection. Another related work advocates the use of Fibre Bragg Gating
sensors for defect detection of rails to monitor track conditions. The authors investigate the
wavelet decomposition of pressure signals but they do not propose a method or threshold
for automatic defect detection. Jianhai et al. use continuous wavelet analysis of
acceleration sensor data to visually inspect the measurements and conclude that there is a
difference in the coefficients for wheel with flat spots and defect-free wheels.

Different kinds of track scales are in use in the field. They can in principle be used to
detect flat spots. But to our knowledge they do not use machine learning to train a defect
classifier. A general advantage of our proposed system is that the measurement system is
relatively inexpensive, but we can show that it can still be used to detect wheel defects,
thanks to our proposed machine learning methods.

Department of Mechatronics, NCERC Page 8

 Wheel Defect Detection With Machine Learning
Seminar Report

Chapter 2
LITERATURE SURVEY

This paper has described the recent engagement with a railroad company to develop a
continuous programmed vision based rail assessment framework. Specifically, the system
is during a position to robustly distinguish significant rail parts with high precision and
efficiency supported visual, location, DMI, and contextual information. The anchor
exception are detected at both tie and compliance levels. Quantitative qualitative analysis
performed on an outsized video informational collection caught with various track and
lighting conditions have exhibited extremely encouraging execution. The foremost
challenge is to handle scenarios during which heavy shadows and light-weight
overexposure exist within the videos. The extensive testing covering longer railroad tracks
with varying defect conditions. Inspection for other rail objects like spikes, spike holes,
and joint bars must be also developed. Finally, to reinforce the algorithms with a
potentially modified imaging system to accommodate a faster and more desirable
inspection speed. Controlled illumination of the rail infrastructure is going to be also
explored to avoid ambient lighting artifacts. Finally, 1the object detection and optimization
approaches that are proposed here are often either applied readily or with minor tunings to
other rail fastening systems and while the vision algorithms that we developed are finely
tuned toward rail track structure and components, the encouraging results we achieved
have demonstrated the applicability of machine vision technology to real applications
within the general transportation domain like the advanced driver assistance system.

In this paper our two-stage pipeline method based on object localization and CNN is tested
on our rail image dataset and it is validated that both the recall and precision of our rail
image classification receive an impressive and robust performance compared with raw rail
image owing to the presentation of he the new loss function and object localization
algorithm. The results also indicate that our localization strategy is significantly for CNN
to learn more characteristic features for classification. Besides, the utilization of transfer
learning strategy also decreases the computational complexity. It is normal that the method
can be utilized for actual rail defect detection automatically. However, our object
localization method is specifically designed for railway image given the rail geometric
attributes and this detection method cannot guarantee the real-time work.

In this paper the type of failures on rail surfaces is determined. The rail surface is detected
from the rail view so that the fault type can be detected. Otsu method was used in
determining the outside of the rail Otsu method is widely used segmentation method.
Feature signals are obtained by calculating the variance values through the ray surface
image. Faults are detected through these signals. The feature type is extracted from the
attribute signals and the type of fault is determined together with the fuzzy logic. When
examining existing studies in the literature, there are many methods for detecting faults
occurring on the track surface. But there are no studies to determine the type of failure. In
this study, both malfunctions are detected and six different fault types are detected on the
track surface. The enhanced Canny edge detection strategy preserves the advantages of the
traditional Canny edge detection technology, and makes many improvements on the basis
Department of Mechatronics, NCERC Page 9
 Wheel Defect Detection With Machine Learning
Seminar Report

of the traditional algorithm. The algorithm first uses adaptive smoothing filter to filter out
salt and pepper commotion and can better protect the details of the image. Using the
geometric characteristics of edge noise to distinguish between true and false edges, can
better remove false edges. The algorithm can accurately locate the edge of the image, can
effectively suppress the noise and texture caused by false edges. The enhanced Canny
algorithm is employed for artificial object edge detection, so as to suppress the influence of
noise on image edge detection and edges from extraneous objects. Compared with the
traditional algorithm, the new algorithm has better effect on image background de-noising
and can retain better results on the target object.

In this paper , it is shown that fiber optic sensing technology can be adequately used for
robust detection of wheel defects in rapid railroad frameworks. Wheels with defects can
wear the rail, reduce the comfort of the passengers and in some extreme cases cause train
derailment. Undoubtedly a fiber optic solution to this problem is close to ideal, since noise
problems are eliminated and the identification of the defective wheels is significantly more
reliable than using conventional technology, strongly affected by electromagnetic noise.
We have shown here that wheel flats can be determined with a single FBG sensor,
interrogated at very high sampling rates (several kHz). Wheel flats can be identified as
high energy impacts in the rail. A detailed analysis methodology has been explained, and
the implications for weight in motion systems have also been presented.

A machine learning method was utilized to screen the condition of damage of steel
examples exposed to cyclic contact weariness tests and to make a nonstop estimation
without coming into contact with the examples analyzed. Six unique arrangements of
examples were thought of and the torque and vibration information acquired from their
individual tests were investigated. Diverse measurable and phantom qualities were gotten
from each dataset, to speak to their data content. The relationship existing between the
qualities was concentrated to choose just the basic ones. The worldwide information
volume was decreased by a primary part investigation. The information in this manner got
were along these lines handled, and arranged, building up a K-means calculation. The
outcomes were figured in probabilistic terms with the guide of participation graphs,
disarray lattices and likelihood bends. The outcomes got from this first investigation are
profoundly promising. The algorithm is able to assess the damage status of the specimens,
as well as detect the operating conditions of the machine. A considerable bit of this
technique is the capacity to appraise quantitatively the damage by ascertaining the
separation of the present information from explicit references. The exhibitions of this
methodology were numerically assessed as far as affectability, explicitness, and exactness.

This paper analyzed two major kinds of surface defects, rail head spalling and breaks in
the surface. According to characteristics of different defects, two methods based on region
width and region position are applied effectively to eliminate interfering objects. Level of
wear of railhead and length of splits, as evaluation of defects of rail head section, are
calculated and outputted automatically by image processing operations. The adaptability of
algorithm needs to be proved by processing more images next. Further research is also
about to develop the algorithm and process consecutive images captured by CCD scan
camera in real time.
Department of Mechatronics, NCERC Page 10
 Wheel Defect Detection With Machine Learning
Seminar Report

A sensitivity-based Tikhonov Regularization algorithm in the EMT system is introduced in

this paper. According to the simulation study, the Tikhonov Regularization algorithm
tackles the poorly presented backwards issue. Because its image reconstruction quality is
better than that of the LBP algorithm. Also, it is more efficient than iterative algorithms. In
general, the Tikhonov regularization strategy is a better EMT image reconstruction
algorithm and extremely suitable for online industrial applications. It is necessary for the
future research to choose particular constraint and a suitable regularization parameter for
variable reconstruction conditions, which is of great importance to the quality of the
reconstruction results.

To address the issues of automation and efficiency in tread defects detection, a wheel tread
extraction and registration method was proposed.The paper introduces the structure and
working principle of tread defect detection system, and then the algorithm based on ellipse
and line extraction, image registration and image difference was studied. Experimental
outcome show that using this method treads could be segmented accurately. And
registration matching rate between the under detecting tread and the template could reach
96%. Missing rate for the real-existing defects detection decreased significantly so this
method could help us get tread defects location automatically and rapidly. As the
information about the defects can be achieved from the located defects, we can determine
the type of the defects by using a classifier such as SVM. We may also use amounts of
labelled images to train a CNN model to predict the type of the defects.

The major goal of this paper is to prevent the collisions of trains with both humans and
animals. The major problem is due to the unmanned railway crossings. The major problem
is due to the unmanned railway crossings. Unmanned crossings are the places where roads
and railway tracks intersect with no monitoring. This research paper was mainly created in
order to avoid the horrifying accidents that have been occurring in India and around the
world by providing a means of providing the motor men information of what is on the
track ahead of them. Our idea portrays the use of Machine Learning to detect the objects
and then passing an alert to the stations which can then be conveyed to the motormen. The
results were simulated and achieved with Python and Tensor flow.

Department of Mechatronics, NCERC Page 11

 Wheel Defect Detection With Machine Learning
Seminar Report

Department of Mechatronics, NCERC Page 12

 Wheel Defect Detection With Machine Learning
Seminar Report

Department of Mechatronics, NCERC Page 13

 Wheel Defect Detection With Machine Learning
Seminar Report

Department of Mechatronics, NCERC Page 14

 Wheel Defect Detection With Machine Learning
Seminar Report

Department of Mechatronics, NCERC Page 15

 Wheel Defect Detection With Machine Learning
Seminar Report

Department of Mechatronics, NCERC Page 16

 Wheel Defect Detection With Machine Learning
Seminar Report

Chapter 3
MEASUREMENT SYSTEM AND DEFECT TYPES

A. Wheel Load Checkpoint

The infrastructure division of the Swiss railway operator SBB operates and maintains the
one of the most heavily used railway network of the world. In 2010, 95.4 km of trains
travelled one kilometer of track on average; this value documents the highest utilisation of
network capacity in the world . Automatically monitoring trains and network are
thus important to minimise the risk of incidents that quickly affect the scheduling of trains
on the network. SBB infrastructure operates an integrated wayside train monitoring system
that controls safety relevant aspects of the railway traffic and infrastructure.

As part of this system, the wheel load checkpoints (WLC) measure vertical force through
strain gauges installed on the rails. These devices are used for observing maximal axle
load, maximal train load, load displacement and grave wheel defects. Our study
investigates the use of machine learning methods to defect and classify wheel defects
based on the data obtained through these wheel load checkpoints.
Each WLC consists of four 1m long measurement bars with four strain gauges (referred to
as sensors in the following) per measurement bar. Since on each side two measurement
bars with 4 sensors are installed, each wheel that runs over the WLC is measured eight
times at different parts of the wheel. Fig. 1 shows schematically the measurement of one
wheel by one measurement bar. In this example a defect is directly observed by the
measurement of the first sensor.

Department of Mechatronics, NCERC Page 17

 Wheel Defect Detection With Machine Learning
Seminar Report

See Fig. 2 for a diagram of one sensor. The strain gauges are installed perpendicular on the
centerline of the railroad track and they are combined into one vertical wheel force
measurement. One sensor covers approximately 30cm of the wheel circumference.The
wheel load checkpoints are installed on multiple strategic sites on the railway network: ten
on the border to Switzerland at the entrance to the railway network maintained by SBB and
a dozen within the network.

B. Railway Wheel Defects

A relatively well understood wheel defect type is the flat spot or wheel flat. This defect
occurs when the wheel stops rotating (for instance during an emergency brake) and is
dragged along the track. Fig. 3 shows an image of a flat spot on a railway wheel of SBB
and the corresponding idealized measurement obtained by the WLC if the flat spot directly
hits a sensor of the measurement system. Grave wheel flats can be detected by looking at
simple statistics of the measurement if the defect hits the sensor perfectly. To be able to
detect flat spots that are less grave or that do not hit a sensor directly, more advanced
machine learning methods are required. We demonstrate such cases on our first data set.

Apart from flat spot, other common wheel defects on railway vehicles are non-roundness
and shelling . Wheels with non-roundness have a high influence on the vibration and noise
emitted by a passing train and, therefore, they are an important type of defect to detect.
Non-roundness, in contrast to shelling and flat spot, is a non-discrete type of defect. This
characterization means that the defect affects a large part of the wheel and changes its
shape in a non-local way. We create an additional data set that contains the defect types
flat spot, non-roundness and shelling and then, we compare the performance of our two
machine learning methods in predicting these three defect type.

Department of Mechatronics, NCERC Page 18

 Wheel Defect Detection With Machine Learning
Seminar Report

Department of Mechatronics, NCERC Page 19

 Wheel Defect Detection With Machine Learning
Seminar Report

Chapter 4
TIME SERIES REPRESENTATION FOR DEFECT
DETECTION

An important step in any machine learning method is finding a representation of the

original measurements that supports discrimination between different classes. For instance:
the mean of the measurement signal of a wheel with or without a flat spot coincide if the
weight of the axle is the same and the defect perfectly hits a sensor. The standard deviation
on the other hand differs significantly, since the force exerted on the track is much higher
for a wheel with the flat spot than for non-defective wheels. For other types of defects like
shelling this observation does not hold, as the variance of the measured force does not
significantly differ from a non-defective wheel, but there is a clear difference in higher
frequency bands of the measurement, c.f. Fig. 4. These observations suggest to
decompose the signal by a multiscale wavelet analysis in order to extract indicative
frequency features for time series data.

Department of Mechatronics, NCERC Page 20

 Wheel Defect Detection With Machine Learning
Seminar Report

Chapter 5

AUTOMATIC REPRESENTATION LEARNING

An alternative to predefined feature representations are provided by deep neural networks

that learn the features from data in a task specific way to maximize correct classification.
In this section we introduce a learning method to automatically infer a representation of
the measurements for the classification of wheel defects based on deep artificial neural
network models (DNN). These models have gained considerable popularity in recent
years, mostly due to their success in image classification and segmentation tasks in speech
recognition and quite recently in reinforcement learning for playing.

DNN for wheel defect detection alleviates the burden of the modeller to manually
construct features and allows to learn representations from time series directly. Another
benefit is the flexibility that comes with designing decision functions as stacked activation
layers. This flexibility allows us to design a network specifically for certain defect types.

A. 2-Dimensional Time Series Representation

Motivated by the success of convolutional neural networks on image classification tasks

we propose the use of 2D representations of the measurement signals for wheel defect
detection. Recently Gramian Angular Fields (GAF) have been proposed as a 2-
dimensional encoding of time series data. This representation has been shown to capture
cross-temporal dependencies and to enhance classification performance when used as
input to a convolution network. A GAF is constructed by first transforming the time series
to polar coordinates and then computing trigonometric sums between all points. As a
second 2D representation we also considered transforming the time series into the image of
its 2D graph. This procedure is motivated by the fact that a human expert would also look
at such a two-dimensional representation to classify wheel defects. The addition of the
value of the signal as the second dimension allows the network to learn different filters for
different values of the signal at the same point in time.

B. DNN Network Architecture

We use a Convolutional Neural Network (CNN) based architecture to automatically

extract the discriminating features. Here, we considered the 8 signals of the WLC as
different channels. Our networks are composed of two modules: the mono channel feature
extracting layers and cross channel feature extracting layers respectively from bottom
(input layer) to top (classification layer). The mono channel feature extracting layers take
each channel independently and compute high level features in parallel that can then be
processed by the cross channel feature extracting layers. Furthermore, the weight of
the mono channel feature extracting layers are shared across all channels, allowing it to
learn from all channels at once. This approach is both computationally efficient, and also
well suited for the data set. Since each channel represents a load measurement of the wheel
Department of Mechatronics, NCERC Page 21
 Wheel Defect Detection With Machine Learning
Seminar Report

from one sensor of the WLC the network learns features from the signals and also a
relationship between the signals.

1) Mono Channel Feature Extracting Layer: This module is a traditional CNN, composed
of a sequence of convolutional layers, eventually followed by a fully connected layer:
a) Convolutional layer: A convolutional layer is a combination of a number of filtering
layers, each followed by a non-linearity and a pooling layer. The settings chosen for
each of these layers are specified below. The filtering layer outputs convolutional products
of the input by learnable filters with a fixed receptive field. Every filter layer is followed
by an activation function. We use a Parametric Rectified Linear Unit (PReLU), as it better
back-propagates the gradient compared to the tangent hyperbolic or sigmoid functions,
which can easily saturate. The PReLU non-linearity also prevents neurones from “dying
out” as can be the case for the popular ReLU units, by introducing a learnable non-zero
slope to the negative side of the input.

PReLU(x) = max(0, x) + a · min(0, x) ,

where a is an adaptable parameter. The pooling layers reduce the resolution of the input
time series and the learned features at each layer of the deep neural network. This max-
pooling allows the classification to be robust to small variations of learned features at each
layer. In all of our convolutional layers, we used a pooling layer with filters of size 2 × 2
applied with downsampling ratio of two, taking the maximum value among the four pixels
in its receptive field.

b) Fully connected layer: Neurons in a fully connected layer have full connections to all
units in the previous layer. The layer outputs biased linear combination of its input,
followed by a non-linearity. As a non-linearity we used the hyperbolic tangent function
(tanh).

2) Cross Channel Feature Extracting Layer:

a) Cyclic permutation network: The cyclic permutation network (Fig. 5) is designed to
learn cross-sensor features invariant to a cyclic permutation of the eight recordings.
Depending on its phase, a given wheel can generate a set of possible recordings, which is
approximately stable by cyclic permutation of the eight recordings. This network
architecture serves the purpose to encode this characteristic of cyclic invariance. The
network works in the following way:

Department of Mechatronics, NCERC Page 22

 Wheel Defect Detection With Machine Learning
Seminar Report

C. Top Layer Features Learned by the DNN

In this section we look at the features learned by the DNN and compare the fifilters learned
by the network on the 1-dimensional or 2-dimensional time series representation. The
results in this section were obtained by training on data set 2 (SectionVI-C) and defect type
flat spot. Examples of top-layer filters learned by the DNN directly on the 1-dimensional
time series, as well as the features extracted by them are shown in Fig. 7. We can observe
that the network has been trained to detect a short quick oscillation in the time series. The
extracted features on the defective input clearly shows the successful training of the model
in detecting defect regions.

Department of Mechatronics, NCERC Page 23

 Wheel Defect Detection With Machine Learning
Seminar Report

Chapter 6
CLASSIFICATION OF WHEEL DEFECTS

Detection and classification of wheel defects amounts to infer from a vertical force
measurement x of a wheel if a wheel is defective or not. Mathematically, a function
f (· ) either encode the binary information, that a defect is present or absent, or its defect
class when we can differentiate the defect category. To achieve this goal we use sets of
measurements of wheels to train decision functions for certain defect types and for non-
defective wheels. We then use this training set of measurements and labels (the type of
defect) to automatically find a function that is expected to predict the defects of wheels not
seen during training accurately.

A. Support Vector Machine

One of the most popular models to find such a function are Support Vector Machines
(SVM) . A SVM finds a linear function parameterized by the vector w that maximally
separates the two classes during training. It achieves this separation by maximizing the
margin between the points of the two classes in feature space, or equivalently by
minimizing the regularized empirical risk.

B. Classification with DNN

It is based on artificial neural network. It is used for an art and visual division regression in
medical image analysis and in computer vision. Represent wheel defect in 2D and it will
compare the obtained data from the wheel defects with the correct data then gives the
result whether the wheel is defective or not.

Department of Mechatronics, NCERC Page 24

 Wheel Defect Detection With Machine Learning
Seminar Report

Chapter 7
DATA SETS AND MODELS

Two data sets from different sources are assembled to evaluate the performance of
different methods for wheel defect detection and classification and to train various
classifiers. For both data sets the signals that we use to predict a wheel defect are measured
by the wheel load checkpoint. The annotations or labels that provide the information about
the defectiveness and defect class of a wheel are collected from different sources. These
data sets contain information about different types of defects as described in the following.
We also describe what models and features we will use for the respective data sets in this
section.

A. Models and Features

On the first data set we compare the Wavelet-SVM with benchmark flat spot prediction
methods. The second data set serves to demonstrate that the Wavelet SVM can accurately
classify all three defect types. We also compare the performance of the deep learning
models on different time series representations by showing that the cyclic permutation
network outperforms the simpler neural networks and also the Wavelet-SVM for non-
roundness. For flat spots, the neural network with features learned on the 2D time series
representation also outperforms the Wavelet-SVM.

B. Data Set 1: Calibration Run

To acquire a first training data set for flat spots, two wheels on different wagons were
artificially damaged. The wagons were then added to a calibration train that was run over
different measurement sites with different velocities and from both directions to calibrate
the wheel load check points. This resulted in 1600 measurements, 50% of which are from a
wheel with a flat spot.
We also consider another method to detect flat spots in this data set, that is not based on
machine learning. It is a conservative threshold on the dynamic coefficient: a general
measure of spread within one time series. For each sensor this coefficient is given by

where max and x refer to the maximum and average value of a sequence of measurements
x, respectively.

Department of Mechatronics, NCERC Page 25

 Wheel Defect Detection With Machine Learning
Seminar Report

C. Data Set 2: Reprofile Events

To generate data for training and testing a classifier that can predict additional types of
wheel defects, we aggregated the time and date of reprofile events and linked them to
railway wagons. We used two sources for these events: the protocols of repair workshops
of freight trains and the regular maintenance measurements of passenger trains. These
were annotated with a defect class by an expert before re-profiling the defective wheels.
We then categorized measurements of the wheel load checkpoints of the same wagons
around the date of re-profiling. Measurements up to a week before re-profiling were
considered defective (according to the class label given by the expert), while
measurements up to a week after re-profiling were considered defect free. Using this
procedure we were able to obtain a large data set of annotated measurements from wheels
of different defect classes over the span of multiple years. 1836 measurements are
evaluated for flat spot detection, where 588 cases are classified as defective. For shelling,
we received 6070 measurements, with 2678 being defective. For the non-roundness defect
class, 688 cases out of 920 measurements are defective.

Department of Mechatronics, NCERC Page 26

 Wheel Defect Detection With Machine Learning
Seminar Report

Chapter 8
EXPERIMENTAL RESULTS

For performance evaluation of the methods we compute three metrics: accuracy, precision
and recall. Whereas accuracy gives the total fraction of correctly classified wheels,
precision measures the fraction of correctly predicted defects out of all predicted defects
and recall the fraction of correctly predicted defects out of all defects.

A. Model Selection and Evaluation

For all the experiments in this section the performance shown is computed on a test set that
was not used for training or model/parameter selection. To make the evaluation robust
against chance we repeat each experiment multiple times on new random train/test splits
and report average and standard deviation over these repetitions. For data set 1 we only
report the average as the standard deviation was not reported for the benchmark method.
B. Data Set 1

In a study prior to this publication , this data set was used to empirically demonstrate the
effectiveness of a new algorithm for MIL. Krummenacher et al. beat stateof-the art MIL
algorithms on this data set and get a classification accuracy of 70% with ellipsoidal
multiple instance learning (eMIL). In this study features based on the Global
Alignment (GA) kernel for time-series were used.

Department of Mechatronics, NCERC Page 27

 Wheel Defect Detection With Machine Learning
Seminar Report

C. Data Set 2 - SVM

Equipped with our general method of constructing features from multiple wheel vertical
force measurements and learning a classifier from them we are now ready to predict other
types of wheel defects as well. We also evaluate the DNN based method in this section.
The SVM classifier are trained on the labels obtained by this method for the defect types
flat spot, shelling and non-roundness.

In Table II the performance on the reserved test set is reported for each defect type
including standard deviation over the permutations. The performance on shelling is the
best out of the three defect types. This observation can be explained by the fact that the
training set for this defect type was by far the largest, so we were able to train a classifier
with higher accuracy. This defect type also affects the wheel globally, so it is harder to
miss for the sensors than a flat spot. To improve the performance on flat spot and non-
roundness we trained custom deep neural networks and give the results in the next
section.

D. Data Set 2 - Deep Learning

Using the same data set as in the previous section we evaluate the deep learning method
on the two defect types flat spot and non-roundness. To simplify the experiments we do
not include additional features like speed, measurement site or template fit, but only
consider the wheel vertical force measurements from the WLC sensors. Therefore, the
performance of the SVM is slightly worse compared to the previous section.

In comparison with the Wavelet-SVM the cyclic DNN shows higher accuracy and
precision and reduced variance. Unlike the DNN for flat spot we only trained the cyclic
DNN for non-roundness directly on the 1D time series, as the increase in parameters due to
the concatenation of measurements of the sensors prohibited efficient training of the model
on the 2D representation.

Department of Mechatronics, NCERC Page 28

Seminar Report Wheel Defect Detection With Machine Learning

Chapter 9
CONCLUSION

We have presented two machine learning methods for defect detection on railway train
wheels. The methods analyse multiple time series of the vertical force of a wheel under
operational speed and output if a wheel has a defect or not. Both methods are trained
automatically on measurements gathered from defective and non-defective wheels. The
first method is based on novel general wavelet features for time series. The second method
employs deep convolutional neural networks to automatically learn features from the time
series directly or from a 2-dimensional representation. We design cyclic shift invariant
artificial neural networks for the detection of wheel flats and non-round wheels that model
the relationship between the measurements inherent to these defects. To evaluate our
methods we collect two data sets from different sources and demonstrate improved
performance for predicting flat spot, shelling and non-roundness.

The methods that were developed for this work are currently being implemented as part of
the SBB wayside train monitoring system. To improve the quality of the training and test
data RFID tags will be deployed to enable perfect association between defect labels and
measurements. Further future work consists of integrating external features into the deep
learning models, optimizing for precision and predicting severity scores for the defects.
For the prediction of severity scores we obtained promising preliminary results on
regressing the flat spot length using support vector regression and the wavelet feature

Department of Mechatronics, NCERC 29

Seminar Report Wheel Defect Detection With Machine Learning

Chapter 10
REFERENCES

[1] R. Müller, D. Leibundgut, B. Stallaert, and L. Pesqueux, “Validation of wheel

maintenance measures on the rolling stock for reduced excitation of ground vibration,”
Trafifikverket, Borlänge, Sweden, Tech. Rep. RIVAS, SCP0-GA-2010-265754, 2013.

[2] P. Huber, B. Nélain, and R. Müller, “RIVAS—Mitigation measures on vehicles

(WP5); experimental analysis of SBB ground vibration measurements and vehicle data,” in
Noise and Vibration Mitigation for Rail Transportation Systems. Berlin, Germany:
Springer, 2015, pp. 531–538.

[3] E. Verheijen and F. Elbers, “Future european noise emission ceilings: Threat or
solution? A review based on Swiss and dutch ceilings,” in Noise and Vibration Mitigation
for Rail Transportation Systems. Berlin, Germany: Springer, 2015, pp. 71–78.

[4] Y. Li and S. Pankanti, “Anomalous tie plate detection for railroad inspection,” in Proc.
21st Int. Conf. Pattern Recognit. (ICPR), 2012, pp. 3017–3020.

[5] Y. Li, H. Trinh, N. Haas, C. Otto, and S. Pankanti, “Rail component detection,
optimization, and assessment for automatic rail track inspection,” IEEE Trans. Intell.
Transp. Syst., vol. 15, no. 2, pp. 760–770, Apr. 2014.

[6] X. Gibert, V. M. Patel, and R. Chellappa. (2015). “Deep multitask learning for railway
track inspection.” [Online]. Available: https://arxiv.org/abs/1509.05267

[7] T. Idé, “Formalizing expert knowledge through machine learning,” in Global

Perspectives on Service Science: Japan. New York, NY, USA: Springer, 2016, pp. 157–
175.

[8] N. Nenov, E. Dimitrov, V. Vasilev, and P. Piskulev, “Sensor system of detecting

defects in wheels of railway vehicles running at operational speed,” in Proc. 34th Int.
Spring Seminar on Electron. Technol. (ISSE), 2011, pp. 577–582.

[9] T. K. Ho et al., “Signature analysis on wheel-rail interaction for rail defect detection,”
in Proc. 4th IET Int. Conf. Railway Condition Monitor., 2008, pp. 1–6.

[10] Y. Jianhai, Q. Zhengding, and C. Boshi, “Application of wavelet transform to defect

detection of wheel flats of railway wheels,” in Proc. 6th Int. Conf. Signal Process., vol. 1.
2002, pp. 29–32.

Department of Mechatronics, NCERC 30