IC 201P – Design Practicum

Machine Learning Based Railway Track

Monitoring System

Shivangi (B21321) Mohit Mahajan (B21055)

Aayush Anupam (B21077) Shantanu Malakolikar (B21131)
Naveen (B21257) Abhishek Meena (B21032)

Under the supervision of

Dr Amit Shukla, amitshukla@iitmandi.ac.in
Dr Narendra Kumar Dhar, narendra@iitmandi.ac.in

Indian Institute of Technology Mandi

 Certificate

This is to certify that the work contained in the project report entitled
“Railway Track Monitoring System”, submitted by Group 8 to the Indian
Institute of Technology Mandi, for the course IC 201P – Design Practicum,
is a record of bonafide research works carried out by him under our direct
supervision and guidance.

Dr Amit Shukla Signature and Date

Dr Narendra Kumar Signature and Date

Dhar

2
 Acknowledgements

We would like to express our heartfelt gratitude to all the individuals who have
contributed for this project.
Firstly, we extend our sincere appreciation to Dr. Amit Shukla and Dr. Narendra
Kumar Dhar for their invaluable guidance, support, and encouragement
throughout this project. Their knowledge, expertise, and feedback have been
instrumental in shaping our ideas and refining our work.
We would also like to thank Mr Naisarga Pandya, for his support and guidance in
helping us with the technical aspects of the project. We appreciate his patience
and willingness to help us with all our queries.
Finally, we would like to express our gratitude to the Indian Institute of
Technology (IIT) Mandi for providing us with the necessary infrastructure and
resources to carry out this project. We thank the institute for creating an
environment that fosters learning and innovation.

3
 Abstract

The Railway Track Monitoring System is an innovative project designed to

enhance the safety and efficiency of railway transportation. The project aims to
develop a drone-based system that can pull along a specialized sensor system to
monitor the railway tracks for any anomalies or defects.
The trolley contains a variety of sensors that can measure several parameters
such as temperature, vibration, and strain. The servo motor-based manipulator
allows the trolley to move and manipulate the sensors to detect any anomalies or
defects in the railway tracks.
The drone is controlled remotely by an operator and is designed to move along
the railway tracks autonomously. The data collected by the sensors is transmitted
wirelessly to a central control station, where it is processed and analyzed to
identify any defects or anomalies in the railway tracks. The YOLO model helps
the mobility of the system by making it autonomous.
The proposed system offers several advantages over conventional methods of
railway track monitoring, such as improved accuracy, reduced cost, and
increased efficiency. It can help detect defects in the railway tracks before they
cause any accidents, thus enhancing the safety of railway transportation.
In conclusion, the Railway Track Monitoring System with a drone-pulled trolley
equipped with sensors and a servo motor-based manipulator is a promising
project that has the potential to revolutionize the railway transportation industry.
The system is expected to be an effective tool for monitoring and maintaining
railway tracks, which will lead to safer and more efficient railway transportation.

4
 Contents

List of Figures Page no. 6

List of Tables Page no. 7

Abbreviations Page no. 8

Chapter 1 Introduction Page no. 9

Chapter 2 Market Research Page no. 11

Chapter 3 Conceptual Design Page no. 14

Chapter 4 Embodiment and Detailed Design Page no. 20

Chapter 5 Fabrication and Assembly Page no. 30

References Page no. 38

5
 List of Figures

Fig 2.1 - Amberg GRP 1000 track inspection trolley

Fig 2.2 - Shenhao SHR-RIIS1005 track inspection robot
Fig 3.1 - Problems Discussed
Fig 3.2 - Conceptual Design
Fig 3.3 - Initial Model
Fig 3.4 - Second Model
Fig 3.5 - Second Model
Fig 3.6 - Universal Joint
Fig 3.7 - Rope
Fig 3.8 - Final Model
Fig 4.1 - Block diagram
Fig 4.2 - Lithium polymer Battery
Fig 4.3 - MG995 Servo motor
Fig 4.4 - HC-SR04 ultrasonic sensor
Fig 4.5 - Carbon fiber tube
Fig 4.6 - Acrylic Sheet
Fig 4.7 - Round caster wheel
Fig 4.8 - NodeMCU ESP8266
Fig 4.9 - Aluminium Sheet
Fig 4.10 - Track detection
Fig 4.11 - 3D View of Model
Fig 4.12 - Orthographic Views of the model
Fig 5.1 - L-Bracket
Fig 5.2 - T-joint
Fig 5.3 - L-joint
Fig 5.4 - Acrylic Sheet
Fig 5.5 - Gantt Chart

6
 List of Tables

Table 3.1 - Decision Matrix

Table 5.1 - Bill of materials

7
 Abbreviations

● PLA - Polylactic Acid

● IEEE - Institute of Electrical and Electronics Engineers
● mAH - Milliampere Hour

8
 Chapter 1
Introduction

The train transport is one of the important methods of transportation in the world
(especially in India) which transfers passengers and goods. Indian railways
carries about 8 billion passengers & 1.42 billion metric tonnes of freight each
year, which is India's 27% of transportation.
The railway track is made up of many important equipment, for example rail,
sleeper, ballast, track bed, etc. But due to the continuous rolling of locomotives
and long-term train loads, the track geometry of the railway line continues to
deteriorate. In worst condition may lead to rollovers or casualties, which in turn
cause irreparable economic losses.
In order to ensure that the train can operate safely, smoothly and uninterrupted at
the specified maximum speed, the railway transportation department can
complete the task of passenger and freight transportation with good quality, the
railway lines must be kept in good condition at the time.

1. Background of the problem :- Lack of efficient and easy to use

mechanism for railway track detection in the world make it a slow and
tedious process. The slow & majorly manual process of inspection is
insufficient to catch the defects at a needed pace causing railway
accidents .

2. Scope of the problem :- Use of technology like machine learning can

make the process automatic and less use of manforce causing it to monitor
the track at a fast pace. We can use the model without causing disturbance
in the railway schedule and can use it at any time.

3. Design philosophy used in this report :- design philosophy of our model

is a bot running on railway track with the help of a drone. The bot will run
on a railway track and data is collected with the help of sensors which will
be examined with the help of a trained ML model. If the defects are found
in the track it will give that data and location to the concerned system along
with the warning. The bot can automatically be removed from track
whenever required.

9
4. Problem Statement :- AI ML based railway track monitoring system to find
defects in the railway tracks with use of modern technology.

5. Beneficiaries (Intended market) :- this concept will modify the existing

technologies used by indian railways and railways over the world to
monitor their tracks. With use of technology we can make track detection
easier and can smoothly do it parallel to the railway schedule. It can
efficiently track the defects leading to minimize the accidents caused by
track defects.

10
 Chapter 2
Market Research

Existing products in the market

Many methods of inspection are present in the world, some are fully manual and
some are a bit automatic.
1. Using track gauge meters :- It is conducted by human inspectors, the
inspectors walk along tracks and conduct measurements manually.

2. GEDO CE track inspection trolley :- widely used in the European railway

industry. It collects line data, obtains the deviation between the design
value and the measured value, and analyzes it through software.

3. Amberg GRP 1000 track inspection trolley :- Made in the German

market , the system consists of a data collection car and a satellite car. It
can measure the horizontal and vertical distances from the track to the
control point.

4. Shenhao SHR-RIIS1005 track inspection robot :- The SHR-RIIS-1005

developed by Shenhao is a fully automatic self propelled cart that can
simultaneously inspect rail profile, track geometry and track clearance, as
shown in Fig. 28. The system contains optional visual sensors which can
automatically detect the surface defect in tunnels, identifying abnormalities
such as water leakage, crack and anomalies. The system has an operating
speed of 5–10 km/h. It has a battery life of over 4 h and a total weight of
115 kg.

11
Fig 2.1 (Amberg GRP 1000 track inspection trolley)

Fig 2.2 (Shenhao SHR-RIIS1005 track inspection robot)

12
Problems associated with these technologies
● For mainly manual techniques like track gauge meters, the working
condition of workers is not ideal as they have to work in adverse conditions
and measurements will be highly biased on proficiency and personal
experience of inspecting staff.
● In the other three inspection processes mentioned above this problem is
there but can be solved. Still the major problem is that they need a human
attendant with the trolley, which may take much labor cost and work at
slow pace as continuous working is not possible.
● Another major problem associated is that these are very heavy trolleys so
it is not easy to remove and fit in track many times. So once fixed it can be
removed at the end of track only.
● Due to this while inspecting the track railway cannot be run on the track
making inspection a slow process. (ex. railway lines in most European
countries are inspected only twice per year).

How our trolley can overcome this problem

● As the trolley we are making is fully operated with the help of a drone the
trolley need not be an attendant moving with the trolley.
● This fully automated trolley can work all the time day and night making the
process of inspection fast and easy. It also consumes very less power.
● This can be removed at any time and not only where the railway track ends
so we need not to block the whole track while inspecting it. The railway can
run on the track in parallel.
● With use of this we can inspect each railway line more number of times
reducing the chances of accidents due to the defect considerably.
● Our trolley is incredibly lightweight. It is more suitable for routine daily
checkups.

13
 CHAPTER 3
Conceptual Design

Problems:

Fig 3.1 (Problems discussed)

Brainstorming and Idea Generation :

We considered developing a bot that can detect cracks in railway tracks using
sensor data. Then we decided to boost efficiency by incorporating AI ML
techniques into our project. Then we researched this problem statement and
discovered that the main issue in this area is how quickly we can do this. Then,
after a few brainstorming sessions, we decided to build a bot that will be pulled
by a drone, and the AI ML will be done on the basis of data collected by the
drone, as well as data collected using the sensors on the bot, which will assist us
in providing confirmation.

Further investigation into this problem statement revealed that one of the major
issues in this area is the speed with which we can detect and respond to track
cracks. Given the sheer size of railway networks and the importance of
minimizing maintenance downtime, it was obvious that any solution we
developed would need to detect and respond to cracks quickly and efficiently. We
made the decision to construct a robot that would be pulled by a drone after
holding several brainstorming sessions. This would enable us to quickly and
effectively cover large tracts of track while also giving us a more thorough and
accurate picture of the track conditions. Data gathered by the drone and data

14
gathered using the sensors on the bot would both be used in the AI/ML process.
This would enable us to confirm any cracks or other issues found and would
allow us to gradually improve our algorithms.

Final Problem Statement: “ Develop a drone-bot system equipped with AI/ML

technology to efficiently detect and respond to cracks and other issues in railway
tracks. ”

Fig 3.2 (Conceptual Design)

Decision Matrix:

Top problems Need for the Alternative Cost of Difficulty level Total
technology availables system
development

Solar panel 4 3 3 2 12
cleaning bot.

Using Al for 3 3 2 3 11
finance
management.

Railway crack 4 4 3 4 15
detection

Pipe cleaning 3 3 2 4 12
bot.

Mind waves 1 2 3 4 10
controlled
drone

Table 3.1 (Decision Matrix)

15
Proposed Solution:
To create the model of the bot we first analyzed the forces involved on the bot.
So the drone will have a horizontal force and a vertical force. The horizontal
force will be responsible for pulling the bot, while the vertical force will lift the
bot upwards.

Here are the images for our first design :

Fig 3.3 (Initial Model)

So, to get stuck onto the railway track we decided to have a model like this in
which the lower part with a caster wheel will be attached to the curved part of
the railway track. We also thought about attaching caster wheels to both curved
faces of the railway track, but to reduce friction we decided to only have two
caster at outer curved surfaces. To handle the jerk due to the varying distance
between the railway track we have used a shock absorber, this inspiration was
taken from the Tie rod of vehicles. These shock absorbing bellows could now
handle basic distance variations.

Our 1st model had the the following problems:

1. Installation on railway tracks was not feasible.

2. Unavailability of bellows in requires size
3. DOF of rod attaching the system and drone was 1
4. Drone of high power was required and stability was a concern.

16
All these problems brought us to the second design

Second Design : Our second design was made keeping into consideration easy
installation of the system in the rail track. We installed servo motors and
designed manipulators which would fit in the railway track and then be lifted by
the drone.

Fig 3.4 (Second Model)

Fig 3.5 (Second Model)

We could solve the first 2 problems by adding servo based manipulators for
installation and spring based system in place of bellows.

17
Designing of universal joint:
Our third problem was partly solved by designing a universal joint. This gave the
connecting rod a DOF of 2.

Fig 3.6 (Universal Joint)

Although universal joint gave us better results, practical runs of your system
showed that a more stable system with better drone connection was still
required. This brought us to our third and final design.

Final design:
In the final cad design we came up with the most simple yet the most impactful
outcome for the connecting rod, which was a rope! (DPF=6)

Fig 3.7 (Rope)

18
The stability issue was also resolved now. We took inspiration from Roller
Coaster rides and made a 3 side covering design. We decided to install rolling
wheels on top and castor wheels on the sides. This ensured stability and easy
frictionless movement.
Since the system now didn't have to be lifted up, less powerful drone could also
do the needful.

Fig 3.8 (Final Model)

Thus we eradicated all listed problems and came up with an efficient model.

Additional Feature:
We realized that use of rope might be an issue while stopping the system. Even if
the drone stopped. the system would now move ahead due to its kinetic energy.
To resolve this problem, we came up with a NodeMCU based automatic braking
system using servo motors similar to the brakes that are found in bicycles. The
wheel arrangement is inspired from the wheels arrangement in roller coasters,

19
 CHAPTER 4
Embodiment and Detailed Design

1. Product Architecture

The bot uses a Lithium Polymer Battery as a power source which powers
the NodeMCU ESP8266 which provides the capability of wireless
transmission as it has a wifi module in it. Ultrasonic sensors are mounted
in the trolley to detect anomalies on the railway track. The sensors send
data to the arduino board which, in turn, sends the information to the
computer in real time using the wifi module. The user can also adjust the
orientation of the servo motors on the IDE for different rail widths. The four
servo motors mounted on the bot to grip the track. Each motor can
generate a torque of 10 kg-cm.

Fig 4.1 (Block Diagram)

20
2. System Level Design

We are using a Lithium polymer rechargeable battery - 30C with a rating of 11.1V
- 2200mAH as the microcontroller that we are using operates safely at voltages
between 7V-12V without a voltage regulator. This battery is responsible for
providing power to the sensors, wifi module and the servo motors. In effect, the
chemical energy stored in the battery changes form into electrical energy and a
part of it subsequently changes into mechanical energy which powers the railway
gripper. It’s weight is 150 grams and its dimensions are 100*20*30 mm.

Fig 4.2 (Lithium polymer Battery)

We have used 4 MG995 servo motors for operating the end effectors. They are
metal geared and provide an average torque of 10 kg-cm. They weigh 55 grams
and as we are using 4 of them therefore a total of 220 grams. Their dimensions
are 40*20*43 mm. They can be controlled using PWM signals. They move the
end effectors.

21
 Fig 4.3 (MG995 Servo motor)

We are using HC-SR04 ultrasonic sensors which can measure distances up to 4

meters. We are using IR sensors also to detect the anomalies. The weights of
these sensors are negligible. The data that they send can be read in real time.

Fig 4.4 (HC-SR04 ultrasonic)

We are also using 3D printed parts made from PLA (Polylactic Acid) and the
whole body frame is made up of carbon fiber tubes as they are very lightweight.
We have tried to make our model as light as possible to make it easy for the
drone to pull it. The carbon fiber rods have a diameters of 25 mm and 18 mm.

22
 Fig 4.5 (Carbon fiber tube)

We are also using 6 mm for mounting the round caster wheels. The caster
wheels have a diameter of about 10 mm. We are planning to use 8 of them.

Fig 4.6 (Acrylic Sheet)

23
 Fig 4.7 (Round caster wheel)

We are using nuts and screws for mechanical fastening and also adhesives at
specific places.
A wifi module named NodeMCU ESP8266 is also to be used for establishing
wireless communication between our system and computer. It is a full fledged
microcontroller in itself.

Fig 4.8 (NodeMCU ESP8266)

The three side roller Coaster inspired locking system has been made using a
3mm Aluminum Sheet of density 2701 kg/m^3.

24
 Fig 4.9 (Aluminium Sheet-3mm)

3. Design Configuration

We have decided that the trolley will have a kind of a gripper mechanism and not
a wheel mechanism to reduce friction to some extent. The caster wheels also
provide an additional advantage of free motion as drones are not very stable.
The wheels have an approximate diameter of 10mm. We are going to install the
sensors near the rails.
The ESP8266 acts as an agent between the whole electronics induced
mechanical system with the computer.
The carbon fiber tubes would be drilled for nut and bolts fittings. We have
decided to keep the caster wheels at a non horizontal angle as the drone
provides an upward pull and this configuration of wheels seems to best fit with
the shape of rails.
The approximate weight of the whole system would not exceed the limit of 1.5
kgs which conforms with our main objective of making a lightweight structure.
The length of the main horizontal is planned to be approximately 500 mm but it is
subject to the availability of railway tracks.

4. Detailed Design

● Electrical/Electronic Aspects: We are using NodeMCU ESP8266 as it

has the wifi module and it enables us to operate the system wirelessly. We
are using jumper wires to connect the sensors and components. A Lithium

25
 polymer battery of 11.1V is used to power the whole system. We use a
buck converter to convert the voltage into small values below 5V as the
operating voltage of the microcontroller is 3.3V. We will connect the Trig
and Echo pins of the HC-SR04 to the digital pins of the arduino. And with
proper use of libraries and code, we can monitor and command wirelessly.

● Software Aspect: We code the NodeMCU in the arduino IDE. We use

several libraries for our job. The list is as follows:
BlynkSimpleEsp8266.h
ESP8266Wifi.h
ESPAsyncTCP.h
ESPAsyncWebServer.h
Servo.h
Ultrasonic.h
We host a web server which shows the output of the ultrasonic
sensors on a graph in real time. We also use the Blynk IoT tool to
control the servo remotely. We have uploaded a HTML file in the
flash memory and use SPIFFS file system to upload it. Anyone who
is connected to the correct wifi hotspot network, can see the data
real time on any device and also control the servo motors.
We also have trained a YOLOv5 model that can detect railway
sleepers in real time. Hence, if we integrate our drone with this
model then the drone can maneuver itself autonomously. The model
is trained for 200 images for 100 epochs at batch-size 2 using
pytorch framework.

Fig 4.10 (Track detection using YOLOv5)

26
● Mechanical Aspects:

Fig 4.11 (3D View of Model)

27
 Fig 4.12 (Orthographic Views of 3D model)

5. Results and Discussion

1. We have tried to keep the setup as light as possible as it is going to

be operated using a drone.
2. Carbon Fiber rods are of very high strength.
3. We will print high density materials to get high strength.
4. The bot is well under the budget.

28
 Chapter 5

Fabrication and Assembly

After making a well versed plan we move towards the final part of our journey.
Fabrication and Assembly. In this part our team uses advanced manufacturing
processes and equipment to fabricate and assemble high-quality components
and products. This chapter describes the final phase of our project and the
report. We present the bill of materials, drawings, material description,
manufacturing process, assembly and integration, limitations and challenges,
and plan, and finally end with our contributions as individuals followed by a
conclusion.

1.Bill of Materials (BOM):

ITEM NAME QUANTITY PRICE PER UNIT TOTAL PRICE

3D printing 1 1200 1200

L joint 1 278 278

Aluminum sheet 1 1600 1600

Ultrasonic sensor 3 220 700

IR SENSOR 3 100 300

CASTOR WHEEL 8 100 800

CARBON FIBER TUBES 6 1001 6006

SPRINGS 6 150 900

WIRES --- -- 200

29
 GLUE GUN 1 460 460

Arduino Mega 2560 1 1600 1600

ESP 8266 3 720 2160

Wifi Module
Data cable 1 149 149

Jumper wires 10 199 199

Lipo 11.1V 1 1500 1500

Servo motors mg995 4 424 1696

Plywood based railway track 1 3200 3200

PLA 1 1600 1600

(3D printing Material)
Nut bolts 1 800 800

Total 25348

Table 5.1 Bill of Materials

2. Mechanical Drawings:
● Carbon fiber to Aluminium connector
Made of PLA through 3D printing
Used for joining carbon fiber tubes with Aluminium.

30
 Fig 5.1 (Connectors)

● Aluminum System
Made of 3mm Aluminum.
Used in the rail lock system.

31
 Fig 5.2 (AluminiumSystem)

T-joint
Made of PLA through 3D printing

Fig 5.3 (T-joint)

32
● L-joint
Made of PLA through 3D printing

Fig 5.4 (L-joint)

● Acrylic Sheet
Used to mount servo motors.
Laser cutting is employed.

33
 Fig 5.5 (Acrylic Sheet)

3. Manufacturing Process description:

Since our cart moves by drone, it has to be lightweight, so we can’t use any
metals, ceramics, or material that would be heavy. This ruled out all the major
manufacturing processes. We are making the main body frame using carbon
fiber tubes. We decided to use PLA as it would be lightweight and comparatively
strong, just perfect for the 3D printing of components. We use a light weight
Acrylic sheet for the mounting of sensors and castor wheels.
We used the following manufacturing process:
● 3D Printing (FDM): It is a type of additive manufacturing process.
We used FDM type of 3D printing for manufacturing our components
like, T joint, L joint, servo mount. We used this method mainly due to
2 reasons, the 1st one being that it is an easy method and secondly
it was a good option to manufacture for light weight components as
our material was PLA. We had to 3D print them several times as
they would just break. We had to try 3D printing them with different
thicknesses to get the perfect one. It took a couple of trials, but we
succeeded in making one that was stable, strong.

34
 ● Laser Cutting: This method was used to cut our Acrylic sheet so
that we can place our sensors and castor wheels.
● Mechanical Fastening: We will use nuts and bolts to join together
the multiple parts.
● Metal Cutting and Bending: The aluminum sheets are bent and cut
to make the appropriate frames.

4. Assembly
After the fabrication and testing of each individual electronic and mechanical part,
we begin with the integration and assembly of our product. We went on to
assemble the product in the following steps:
● Drilling of carbon fiber tubes.
● 3D printing of appropriate components.
● Mechanical Fastening of the parts.
● Laser cutting of acrylic sheet.
● Attaching acrylic sheet to the 3D printed parts using nuts and bolts
● Mounting of caster wheels on the acrylic sheets.
● Mounting of the servo motors.
● Making the electronic circuit using the microcontroller.
● Connecting the jumper wires and arranging the system.

5. . Limitations and Challenges

● Limitations:
o Acrylic sheet is not strong enough.
o Wireless communication
o 3D printed materials strength problem.
o Stability on track.

● Challenges:
o Figuring the best design which is stable on track.
o Designing a lightweight system.

35
 o For placement of the caster wheels.
o Reducing friction.

6.Scheduling Plan
In this section, we present the actual timeline that was decided for the project.

Fig 5.5 (Gantt Chart)

7.Contribution
● Shivangi (B21321) : CAD Design, fabrication, material research, ideas
● Aayush Anupam (B21077): CAD Design, electronics, Machine learning,
ideas
● Mohit Mahajan (B21055) : CAD Design, fabrication, ideas
● Naveen (B21257) : logistics, fabrication, ideas
● Shantanu Malakolikar (B21131) : fabrication, logistics, ideas
● Abhishek Meena (B21032): fabrication, logistics, ideas

8.Conclusion
Here is a system that is immensely lightweight, therefore mobile that can detect
anomalies in railway tracks. This system is not rigorous i.e. due to constraints of
budget this project acts as a starting point of a concept that has a very wide
scope. Much remains to be done, regarding the stability and some other factors.
Its real time data reading capabilities can be harnessed in a very meaningful
manner. It can also maneuver itself using the trained YOLO model.

36
References:

1. Li, Y.; Trinh, H.; Haas, N.; Otto, C.; Pankanti, S. Rail component detection, optimization, and
assessment for automatic rail track inspection. IEEE Trans. Intell. Transp. Syst. 2013, 15, 760–770.

2. Zuwen, L. Overall comments on track technology of high-speed railway. J. Railw. Eng. Soc. 2007, 1,
41–54.

3. Johnson, J.; Karpathy, A.; Li, F.-F. Densecap: Fully convolutional localization networks for dense
captioning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las
Vegas, NV, USA, 27–30 June 2016; pp. 4565–4574.

4. Yang, L.; Tang, K.; Yang, J.; Li, L.J. Dense captioning with joint inference and visual context. In
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, USA,
21–26 July 2017; pp. 2193–2202.

5. Wang, T.J.J.; Tavakoli, H.R.; Sjöberg, M.; Laaksonen, J. Geometry-aware relational exemplar attention
for dense captioning. In Proceedings of the 1st International Workshop on Multimodal Understanding and
Learning for Embodied Applications, Nice, France, 25 October 2019; pp. 3–11

37
38