FACULDADE DE E NGENHARIA DA U NIVERSIDADE DO P ORTO

Sensor System for Predictive

Maintenance in Industrial Environment

André Cunha Enes Gonçalves

Mestrado em Engenharia Eletrotécnica e de Computadores

Supervisor: Vítor Hugo Machado Oliveira Pinto

October 17, 2023

© André Cunha Enes Gonçalves, 2023
Resumo

Com a chegada da Indústria 4.0 e do aumento da acessibilidade de hardware capaz de capturar

e processar grandes quantidades de dados complexos, a manutenção de equipamentos industriais
tem passado por vários avanços tecnológicos, mais especificamente, no que concerne à avali-
ação da condição e performance dos equipamentos. Estes avanços possibilitaram o surgimento
da Manutenção Preditiva, que usa dados em tempo real para prever futuras falhas iminentes
em equipamentos, minimizando tempo de inatividade dos equipamentos e reduzindo custos de
manutenção.
A maioria das operações de manufatura dependem de equipaments tecnologicamente ultra-
passados, ditos legacy, que não possuem os benefícios de sensores integrados que possibilitam a
incorporação de Manutenção Preditiva nas linhas de produção.
Avanços na Internet of Things e no ramo de Machine Learning motivaram o tema e objetivo
desta dissertação: a projeção de um sistema sensorial capaz de integrar Manutenção Preditiva
em sistemas legacy, mais especificamente num Automated Guided Vehicle e nas máquinas envol-
ventes.
O sistema tirará proveito de sensores de vibração, som e temperatura para obter os dados e de
algoritmos de Machine Learning para a deteção de anomalias. Uma análise aprofundada do tipo
de sensor e das características dos mais usados na literatura para medir as grandezas mencionadas
foi realizada, tendo em conta o custo do sistema completo e as limitações físicas, em termos de
espaço, do veículo.
Para testar o conceito e comprovar a metodologia utilizada, o sistema foi utilizado em platafor-
mas de testes, simulando o Automated Guided Vehicle, para detetar comportamento anómalo.
O pré-processamento utilizado baseia-se em características temporais para reduzir a quantidade
de dados enviados à distância. A técnica deu origem a resultados praticamente ideais para os
modelos de Supervised Learning, enquanto os resultados dos modelos de Unsupervised Learning
revelaram-se relativamente fracos. Adicionalmente, uma comparação entre sensores de preços
diferentes foi efetuada.

i
ii
Abstract

In line with the Industry 4.0 Revolution, equipment maintenance is undergoing several advance-
ments which take advantage of the ever-increasing affordability of hardware capable of capturing
and processing high amounts of complex data to analyse equipment health and performance. This
progress has given rise to Predictive Maintenance, a proactive strategy that uses real-time data
analysis to predict equipment failures, which optimises schedules, minimises downtime and re-
duces maintenance costs by replacing components only when necessary.
A substantial portion of manufacturing operations depend heavily on legacy equipment to
meet their production demands, which lack the advantages of contemporary sensor technology to
incorporate Predictive Maintenance in the production line.
By leveraging technology developed for the Internet of Things and state-of-the-art Machine
Learning algorithms, this dissertation’s ultimate goal is to design a wireless sensor system capa-
ble of bringing Predictive Maintenance to legacy machinery, more specifically to an Automated
Guided Vehicle and its surrounding machines.
The system aims to achieve its objective through the use of vibration, sound and temperature
sensors as well as with Machine Learning algorithms. A thorough analysis of the sensor technol-
ogy and characteristics employed in the literature of these physical quantities was made, keeping
in mind the physical space constraints of the vehicle and the ultimate cost of the sensor system.
The system concept was put to the test by detecting anomalous behaviour in testing platforms
meant to replicate an Automated Guided Vehicle. The preprocessing method used is based on
time-domain features to minimise the amount of data to send wirelessly. The technique yielded
nearly ideal predictions for the selected Supervised Learning models, while the performance of the
Unsupervised Learning models was notably poor. Additionally, a comparison between sensors,
which varied in cost, was conducted.

iii
iv
Acknowledgements

It took me some time to realise I had fallen into a black hole of possibilities. Without the help of
many people, I am not sure I would be able to climb out.
First, I would like to thank my supervisor, Professor Vítor Pinto, for giving me the opportunity
to be the first student to take on a new project take will be carried out by other students in the near
future. I would still be worrying about insignificant details without your guidance. I would also
like to thank DIGI2 Laboratory for providing all the extra equipment that I would not have access
to if I was to do my dissertation elsewhere.
My family is responsible for the dedication I put into this project. If not for them, I would not
know how to keep working, even though, sometimes, motivation vanishes. It came back.
I would have gone insane if not for the help and care of all my friends. Why do anything if
you can’t have some fun doing it?
Thank you

André Enes

v
vi
 “Quem não tem cão,
caça com gato.”

Ditado Popular

vii
viii
Contents

1 Introduction 1
1.1 Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Document Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Background 5
2.1 Nomenclature Disambiguation . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Types of Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 Reactive Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Preventive Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.3 Predictive Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.1 Envelope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.2 Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.3 Signal Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3.4 Short-Time Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . 10
2.3.5 Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3.6 Empirical Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.7 Empirical Mode Decomposition . . . . . . . . . . . . . . . . . . . . . . 13
2.3.8 Hilbert Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3.9 Hilbert-Huang Transform . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Related Work 15
3.1 Critical Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.1 Vibration Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2.2 Acoustic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2.3 Acoustic Emission Analysis . . . . . . . . . . . . . . . . . . . . . . . . 19
3.2.4 Temperature Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2.5 Oil Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2.6 Electric Current Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Prediction Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3.1 Prediction Problem Definition . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.2 Statistical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3.3 Supervised Learning Methods . . . . . . . . . . . . . . . . . . . . . . . 24
3.3.4 Semi-Supervised and Unsupervised Learning Methods . . . . . . . . . . 26
3.4 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

ix
x CONTENTS

3.4.1 Outlier Removal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.5 Digital Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.5.1 Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.5.2 Short-Time Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . 31
3.5.3 Time-frequency Techniques . . . . . . . . . . . . . . . . . . . . . . . . 32
3.6 Wireless Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.7 Testing Platforms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34

4 Methodology 37
4.1 Prototype Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.1.1 Vibration Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.1.2 Microphones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1.3 Temperature Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.1.4 Microcontroller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.1.5 Voltage Incompatibility . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.1.6 Analog-to-Digital Conversion . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.7 Hardware Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.8 Local Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.1.9 WiFi Connection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.1.10 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.2 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3 Custom Testbed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.3.1 Tested Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.4 Software Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4.1 Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.4.2 Dataset Creation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.4.3 ML Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5 Experimental Results 67
5.1 Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Test Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.1 Dataset Unbalance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.2 Sensor Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.2.3 ML Automation Script . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.3 Ultrasonic Microphone Measurements . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Temperature Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5 Unsupervised Models Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.6 Supervised Models Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.6.1 Sensor Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.6.2 Unknown Anomalies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

6 Conclusion and Future Work 81

6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

References 85
List of Figures

2.1 Nomenclature Venn Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.2 Envelope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Example STFT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.4 Ricker and Morlet Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Example Wavelet Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.6 Example Empirical Mode Decomposition . . . . . . . . . . . . . . . . . . . . . 13

3.1 Critical Components Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.2 IRT Image of an AC Motor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 Ball Bearing Internal Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4 One-Class SVM Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.5 Autoencoder Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.6 PRONOSTIA testbed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.7 SpectraQuest Simulator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

4.1 ADXL355 Accelerometer Evaluation Board . . . . . . . . . . . . . . . . . . . . 38

4.2 MMA8451 Accelerometer Breakout Board . . . . . . . . . . . . . . . . . . . . . 39
4.3 GRAS 40PH Frequency Response . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.4 Digital Microphone from Adafruit . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.5 Digital Microphone Frequency Response . . . . . . . . . . . . . . . . . . . . . 42
4.6 Ultrasonic Microphone Breakout Board . . . . . . . . . . . . . . . . . . . . . . 43
4.7 Ultrasonic Microphone Frequency Response . . . . . . . . . . . . . . . . . . . . 44
4.8 Microphone Amplifier Circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.9 MLX90514 Infrared Thermometer . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.10 Teensy 4.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.11 Logic Level Converter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.12 Module Connection Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.13 ESP-12E . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.14 FT232RL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.15 ESP-12E Programming Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.16 ESP-12E Serial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.17 Module Overview Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.18 Testbed 1 in a bare state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.19 Testbed 2 in a bare state . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.20 Testbed 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.21 Ball Bearing in normal state . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.22 Ball Bearing in anomalous state . . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.23 Main Finite State Machine - SD Card . . . . . . . . . . . . . . . . . . . . . . . 59

xi
xii LIST OF FIGURES

4.24 Continuous Recording Finite State Machine . . . . . . . . . . . . . . . . . . . . 59

4.25 Read sensor data Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.26 Save data in SD card Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.27 Dataset Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.1 Confusion Matrix Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

5.2 ML Automation Script Flowchart . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3 Ultrasonic Microphone Readings . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.4 Zoomed In Ultrasonic Microphone Readings . . . . . . . . . . . . . . . . . . . . 72
5.5 Temperature Readings - Testbed 2 . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.6 Zoomed In Temperature Readings . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.7 Isolation Forest Confusion Matrix . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.8 XGBoost Confusion Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.9 Confusion Matrix of model in Table 5.7 . . . . . . . . . . . . . . . . . . . . . . 79
List of Tables

3.1 IoT communication protocol comparison . . . . . . . . . . . . . . . . . . . . . . 33

4.1 Accelerometer Feature Comparison . . . . . . . . . . . . . . . . . . . . . . . . 38

4.2 Microphone Feature Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3 SPH0645LM4H’s Main Features . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.4 SPU0410LR5H-QB’s Main Features . . . . . . . . . . . . . . . . . . . . . . . . 44
4.5 Amp Passive Component Values . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.6 MLX90614 Specifications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.7 Communication Protocol Comparison . . . . . . . . . . . . . . . . . . . . . . . 51

5.1 Results Temperature vs. No Temperature . . . . . . . . . . . . . . . . . . . . . . 74

5.2 Unsupervised Models Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.3 Best Results from Supervised Models . . . . . . . . . . . . . . . . . . . . . . . 76
5.4 More results for the same conditions . . . . . . . . . . . . . . . . . . . . . . . . 77
5.5 Sensor Impact on model performance comparison . . . . . . . . . . . . . . . . . 78
5.6 Model metrics combining sensors . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.7 Best results testing with anomalies not present in the dataset . . . . . . . . . . . 79

xiii
xiv LIST OF TABLES
Acronyms and Abbreviations

ADC Analog-to-Digital Converter

ANN Artificial Neural Network
CNN Convolutional Neural Network
DMA Direct Memory Access
DSP Digital Signal Processing
EMD Empirical Mode Decomposition
HHT Hilbert-Huang Transform
HSA Hilbert Spectral Analysis
HTTP Hypertext Transfer Protocol
I2C Inter-Integrated Circuit
I2S Inter-IC Sound
IMF Intrinsic Mode Function
IMF Intrinsic Mode Function
IR Infrared Light
IoT Internet of Things
LOF Local Outlier Factor
MEMS Microelectromechanical Systems
ML Machine Learning
MQTT Message Queuing Telemetry Transport
OSI Open Systems Interconnection
PCB Printed Circuit Board
PM Preventive Maintenance
PdM Predictive Maintenance
RF Random Forest
RM Reactive Maintenance
RTD Resistance Temperature Detector
SPI Serial Peripheral Interface
STFT Short-Time Fourier Transform
SVM Support Vector Machine
TCP Transmission Control Protocol
UART Universal Asynchronous Receiver/Transmitter
UDP User Datagram Protocol
USB Universal Serial Bus
k-NN k-Nearest Neighbours

xv
Chapter 1

Introduction

1.1 Context

In an ever-increasing world of rising industrial productivity, with the advent of Industry 4.0, where
every process is optimised to the fullest, the maintenance of the machines in charge of this surge
cannot be left behind. Traditional maintenance practices often result in excessive downtime, in-
creased repair costs, and inefficient resource allocation. Thankfully, this problem has not gone
unnoticed. A new paradigm in industrial maintenance, leveraging state-of-the-art technology, is
emerging in modern industry: Predictive Maintenance [1]. This new model uses data collected
from machines to estimate when a certain machine part will require either repair or replacement.
Older maintenance systems would either correct the anomaly after it occurred or preventively
replace parts, without confirming their actual necessity.
A significant challenge in implementing Predictive Maintenance (PdM) lies in integrating
modern technologies into older machinery without resorting to replacing the entire equipment
with newer, technologically advanced models. Retrofitting these machines with suitable sensors,
data acquisition systems, and communication interfaces necessitates careful engineering to ensure
seamless compatibility and efficient data integration [2]. A module containing several types of
sensors and capable of processing data autonomously which could be used in different machines
without significant changes would enable legacy systems to benefit from the advantages of data-
driven maintenance techniques while reducing the cost of individually retrofitting older machines.
Recent research developments in the area of Machine Learning (ML) have enabled complex
processes being able to be predicted with a high degree of certainty. These advancements also
include the field of maintenance, with several different methods being explored and evaluated
with favourable results over more traditional methods.

1.2 Motivation

Estimates show that maintenance costs account for 15 to 60 % of the total expenses of manu-
facturing operations [3]. While these costs vary substantially between industries, depending on

1
2 Introduction

how much that production relies on complex and highly specialised equipment, such as the avi-
ation industry, due to the need for specialised technicians, spare parts, and intricate maintenance
procedures, whereas a sector like agriculture may have lesser costs in maintenance on account of
relatively simpler machinery. Nevertheless, all industries can benefit from the implementation of
Predictive Maintenance since it can heavily lower unplanned downtime and lower product defects
due to earlier warnings of necessary repairs [4].
It is also important to highlight the impact of defective equipment on the well-being of factory
workers. Prolonged exposure to such noisy conditions can give rise to various health conditions,
with hearing damage or loss being one of the most glaring concerns. Furthermore, there are more
subtle repercussions, involving a decrease in worker productivity due to annoyance, frustration,
dissatisfaction and increased stress; the risk of accidents also increases [5]. Although most noise
would not be entirely eliminated from a prototypical factory, the deployment of Predictive Main-
tenance in such an environment would alleviate some of the harmful noise.
Most current implementations of PdM in practical applications make use of existing theoret-
ical models in order to monitor and predict faults in the system under evaluation. Nonetheless,
these models require a considerable amount of knowledge about the system and their predictions
can be hindered by high amounts of noise which limit the potential use cases. Employing the
use of ML techniques facilitates the implementation of Predictive Maintenance in more intricate
systems, where there is less knowledge available or in more dynamic conditions where the setting
is constantly affected by noise. Machine Learning methods are capable of extracting features from
a set of data while being agnostic to the problem at hand. This versatility allows for detecting
anomalies in complex data [6].
The domain of the Internet of Things has propelled forward the development of microcon-
trollers, sensor technology, and lightweight communication protocols, unveiling the potential of
energy-efficient devices. The recent interest, in conjunction with the decrease in the cost of elec-
tronics that can be used in a Predictive Maintenance context, such as accelerometers and temper-
ature sensors, makes possible the creation of a module capable of collecting relevant data from
machines, running algorithms that detect faults and sending meaningful information wirelessly to
alert the maintenance department.

1.3 Objective
The objective of this dissertation is to develop a system capable of determining when maintenance
is required on an Autonomous Mobile Robot and its surrounding machines in an industrial sce-
nario. To reach this objective, a sensor module for PdM was designed and tested. The full system
would include software outside the microcontroller to fully process the gathered data to detect
faults. The complete system should be able to fulfil the following criteria:

• Collect data from multiple sensors

• Preprocess the data

1.4 Document Structure 3

• Send the data to a database

• Detect anomalous behaviour present in the data

Considering that the final objective is to achieve the capability to detect anomalies in an Auto-
mated Guided Vehicle whose assignment is to tug cargo around a factory floor, and since intention-
ally damaging the vehicle is not possible, a testing platform must be built to evaluate the system
and to ensure the procedures function as intended. Different Machine Learning algorithms will
also be tested to determine the most effective ones for the given scenario. Although not mandatory,
the overall cost of the module should be taken into account by ensuring cost-effectiveness.

1.4 Document Structure

This dissertation is composed of six chapters, where the entire process of developing the module
is laid out, from researching relevant technologies to the final tests and conclusions. In Chapter
2, some theoretical concepts are discussed, which will be pertinent for the rest of the disserta-
tion. Chapter 3 reviews articles containing work related to the objectives, which components are
the most critical for machines, the types of measurements used in the field, signal processing
techniques, machine learning algorithms, preprocessing methods and wireless technologies. In
Chapter 4, the methodology for developing the module is detailed, analysing how the components
were chosen, how the software to run the system works and the method behind the testing proce-
dures. Chapter 5 discusses the outcomes while explaining the metrics utilised and exploring the
results. Chapter 6 brings forth the conclusion and emphasises further future improvements.
4 Introduction
Chapter 2

Background

In this chapter, relevant concepts to the dissertation’s objective will be reviewed. Ranging from
the existing types of maintenance to signal-processing theoretical base concepts. Additionally, a
dedicated section will address and clarify certain nomenclature used in pertinent articles where a
definitive standard or consensus is lacking regarding specific definitions.

2.1 Nomenclature Disambiguation

Condition Monitoring refers to the continuous assessment and tracking of various parameters or
indicators to evaluate the health and performance of a system or equipment [7]. It aims to detect
any deviations from normal operating conditions and provides early warnings of potential failures
or abnormalities. It is not generally used in other contexts apart from Predictive Maintenance.
The Remaining Useful Life of an asset or system is the duration remaining from the present
moment until the conclusion of its productive operational span [8]. Predicting the Remaining Use-
ful Life involves analysing historical data, condition monitoring data, and possibly other factors
to estimate when a system might fail or become inefficient. Other definitions are used in other
research fields, but the previous one remains the predominant choice when considering system
health monitoring.
Fault Detection is defined as the determination of the occurrence of faults in the monitored
system. It entails the detection of faults in the processes, actuators and sensors by using depen-
dencies between different measurable signals [9]. A fault is defined as an unpermitted deviation
of at least one feature of the system from the standard and acceptable condition [10]. A failure is a
permanent interruption of a system’s ability to perform a required function under a set of specific
operating conditions [11]. The difference between these definitions clarifies that articles referring
to Fault Detection tend to focus on a specific part of a bigger machine or structure, as the case
with [12], which is a review on bearing fault detection, [13] which is a review on induction motor
fault detection and [14] which is about fault detection in vibrating structures. The term Failure
Detection is rarely used and normally appears in articles in the Distributed Computing field [15].

5
6 Background

Fault Diagnostics delves into the identification and understanding of the root cause behind a
detected fault or anomaly. It involves a comprehensive analysis of available data to pinpoint the
exact source of the problem. Fault diagnostics aim to answer questions like "What is causing the
fault?" and are crucial for effective troubleshooting and targeted repair strategies. It is a more
in-depth step following fault detection.
Fault Prognostics take a step beyond fault detection and diagnostic processes. Prognostics in-
volve predicting the future behaviour of a system based on its current state and historical data. This
includes estimating the time remaining before a system experiences a specific failure, considering
factors like wear and tear, usage patterns, and environmental conditions.
Anomaly Detection is the process of identifying deviations or outliers from normal patterns
within a given dataset or system behaviour [16]. Techniques referring to Anomaly Detection can
be applied in several different fields, including fraud detection [17], cyber-intrusion detection [18]
and image processing [19]. The most relevant articles on Anomaly Detection for this dissertation
are those whose focus is on industrial damage detection, which will be further explored. The term
is usually meant for moving parts analysis, as with bearings and gearboxes, and not referring to
structural or static parts analysis
Figure 2.1 shows how the research terms relate to each other. It is important to note that
the circles’ proportions do not represent the frequency of keyword occurrence; rather, they serve
as a visual aid to enhance comprehension. Condition Monitoring, Anomaly Detection and Fault
Detection were the target terms for the research part of this dissertation since the project focus
area is on detecting faults in the system under analysis. The scope of this project, as outlined in
Figure 2.1, does not encompass tasks such as predicting the Remaining Useful Life of specific
components or engaging in fault diagnosis and prognostics. Although the previous keywords
were not the priority during the research phase, some techniques that are often employed in these
research fields were analysed due to the comparable goals of the articles and of this dissertation.
Another ambiguity found while researching for this dissertation is that some articles refer to
Anomaly Detection as Abnormality Detection; however, both terms can be used interchangeably
when referring to the same goal: to identify deviations from normal patterns, as previously defined.
In PdM, Anomaly Detection is more common, with Abnormality being more common in the
medical field, [20] [21].

2.2 Types of Maintenance

The impact of maintenance represents a total of 15 to 60% of the total costs of operating of all
manufacturing [22]. As such, it is essential to compare the existing maintenance plan strategies to
select the one which minimises these costs.

2.2.1 Reactive Maintenance

Reactive Maintenance (RM) (or Corrective Maintenance) is the simplest maintenance strategy as
it only happens when a piece of equipment stops working and consists of stopping production to
2.2 Types of Maintenance 7

Figure 2.1: Venn Diagram of research terms

repair the failure [23]. RM can lead to increased downtime, production delays, and higher repair
costs. Moreover, the unpredictability of breakdowns can impact operational efficiency and hinder
long-term asset reliability. As a result, many industries seek to transition from reactive to more
proactive maintenance approaches to optimise equipment performance and minimise unplanned
disruptions.

2.2.2 Preventive Maintenance

Preventive Maintenance (PM) is a technique involving periodic maintenance as to prevent sudden

failure. This approach relies on scheduled inspections and repairs to ensure that equipment is kept
in good working condition and that any potential problems are identified and addressed before they
lead to major breakdowns or downtime, thus saving maintenance costs [23]. It is more effective in
reducing costs than RM but repairs and machine parts are replaced regardless of the deterioration
state of the piece of equipment, leading to unnecessary corrective actions.

2.2.3 Predictive Maintenance

PdM as a concept has long existed, but only recently have emerging technologies become seem-
ingly capable and inexpensive enough to make PdM widely accessible [24]. It has been adopted
by many sectors, especially by the ones where reliability is paramount, such as power plants, util-
ities, transportation systems, communication systems, and emergency services [25]. In essence, it
uses data analysis to predict when equipment failure is likely to occur, allowing maintenance to be
performed proactively and optimising the evaluation of the monitored system’s deterioration state
[23].
8 Background

As mentioned above, PdM uses data to predict future failures by leveraging data analytics,
sensor technology, and algorithms to predict and prevent equipment failures before they occur.
Unlike traditional maintenance approaches, RM and PM, which rely on fixed schedules or reactive
responses to breakdowns, PdM uses real-time data from sensors and historical performance infor-
mation to forecast potential issues and identify early warning signs of equipment deterioration.
By continuously monitoring the health and performance of machinery, PdM enables proactive
interventions, such as timely repairs or component replacements, extending the lifespan of equip-
ment and minimising costly downtime and unplanned disruptions to operations. This approach has
gained significant traction across industries, including manufacturing, energy, transportation, and
healthcare, as it offers increased operational efficiency, reduced maintenance costs, and improved
safety, ultimately leading to enhanced productivity and customer satisfaction [24].

2.3 Signal Processing

Signal processing is the field that deals with analysing and manipulating signals to extract useful
information. It involves techniques like filtering, compression, and analysis to enhance signals
and remove noise. It has applications in telecommunications, audio and video processing, medical
imaging, and more. Signal processing plays a pivotal role in condition monitoring by extracting
meaningful insights from diverse signals like vibration, temperature, and pressure. Through tech-
niques like feature extraction, pattern recognition, noise filtering, and time-frequency analysis,
signal processing enables the identification of anomalies, patterns, and changes in signals, which
are indicative of system health and potential faults. It aids in enhancing data quality, detecting
deviations from normal behaviour, and enabling accurate predictions, thus facilitating informed
decision-making for maintenance and operational optimisation.

2.3.1 Envelope

In the domain of signal processing, an envelope represents a curve that traces the changes in
amplitude or intensity of a signal over time. Figure 2.2 shows a signal and its envelope, generated
using the SciPy 1 Python package.
Envelopes play a role in time-frequency analysis, offering insights into how a signal’s fre-
quency content evolves over time. This is particularly valuable in audio processing, enabling the
identification of time-varying features in audio signals. Additionally, envelopes serve as a visu-
alisation tool, providing a concise representation of a signal’s overall behaviour. They simplify
waveform analysis, facilitating tasks like noise reduction and feature extraction [26].

2.3.2 Discrete Fourier Transform

The Fourier Transform converts a function into a form that describes the frequencies present in the
original function [26]. It is a fundamental technique that is used in many different fields, ranging
1 https://scipy.org/
2.3 Signal Processing 9

Figure 2.2: Envelope of a signal generated through SciPy

from mathematics and physics to image analysis and audio compression. This project’s most
relevant use case is its use in Signal Processing, which deals with the analysis, manipulation, and
interpretation of signals. Signals are defined as a function that conveys information, such as audio,
video, images, sensor data, or any other form of measurable data. Signal processing techniques
aim to extract useful information from signals, enhance their quality, remove noise, compress them
for efficient storage or transmission, and perform various other operations [27]. Signal Processing
can be subdivided into two sub-fields: Analog Signal Processing and Digital Signal Processing
(DSP); the latter handles signals that need to be analysed using digital hardware, as the case with
this dissertation.
The Fourier Transform in the DSP world must be adapted for discrete signals since analog
signals cannot be sampled at an infinite rate. Thankfully, the Discrete Fourier Transform (DFT)
exists and is widely used in DSP. It enables the analysis of discrete-time signals in the frequency
spectrum, and the equation that describes it for a periodic signal is displayed in Equation 2.1 [26].

1 N−1 2π
X(k) = ∑ x(n)e j( N )kn (2.1)
N k=0

The most challenging obstacle of using this equation lies on the fact that its computation is
of O(N 2 ) complexity, using the big O notation [28]. The most well-known and used algorithm to
compute the DFT faster and more efficiently is the Fast Fourier Transform (FFT) [29]. The main
advantage of using the FFT algorithm is that its complexity is O(Nlog(N)), and therefore, it is
significantly faster for larger sequences. This speed advantage is crucial for real-time signal pro-
cessing applications, such as audio and video processing, communication systems, image analysis,
and many other domains where processing large amounts of data is required.

2.3.3 Signal Filtering

In the realm of Signal Processing, noise poses a substantial challenge, as it can impact the desired
outcome at every stage, ranging from the amplification of electrical signals to the subsequent
sampling of these signals. Most electrical noise manifests itself as unwanted frequencies in the
signal. These undesirable frequencies can be entirely or partly removed from the signal using
10 Background

analog and/or digital filters. The greatest difference between analog and digital filters lies in the
fact that analog filters deal with continuous signals and digital filters with discrete signals. Another
divergence is that analog filters are deployed using physical components, e.g. capacitors and
inductors, while digital filters are mostly set up using programs processed using a microprocessor
after the signal has been sampled with an Analog-to-Digital Converter (ADC) [26]. Different types
of filters, which vary in complexity, affect the original signal differently in its frequency response
and phase delay, and some of the most common designs will be explored further along.
Analog and digital filters are essential signal processing components used to modify or extract
specific frequency components from signals. Analog filters are implemented using analog elec-
tronic components such as resistors, capacitors, and inductors. They operate on continuous-time
signals and are typically found in analog circuits, audio systems, and radio frequency applications.
Analog filters are characterised by their frequency response, which determines how the filter at-
tenuates or amplifies different frequency components in the signal. The design of analog filters
involves selecting the appropriate filter type (e.g., low-pass, high-pass, band-pass) and adjusting
parameters to achieve the desired frequency response [26].
On the other hand, digital filters process discrete-time signals represented in digital form, such
as those found in digital audio processing, image processing, and communication systems. Digital
filters are implemented using digital signal processing (DSP) algorithms and are often realised
using software or specialised hardware. The design of digital filters is based on mathematical al-
gorithms that manipulate the signal’s discrete samples. Digital filters can offer greater flexibility
and precision compared to analog filters due to the inherent advantages of digital processing, such
as easy reconfigurability, precise control over filter characteristics, and the ability to implement
complex filtering operations. Moreover, digital filters can be easily replicated, stored, and trans-
mitted, making them ideal for modern digital systems where data is processed and transmitted
in digital format. Both analog and digital filters play critical roles in various engineering appli-
cations, ensuring efficient signal processing and signal conditioning in a wide range of systems
[26].

2.3.4 Short-Time Fourier Transform

The Short-Time Fourier Transform (STFT) is an algorithm meant to increase clarity in the time
domain when analysing signals in the frequency spectrum. As previously seen in Subsection 2.3.2,
the Fourier Transform enables the transformation of a signal from the time domain to the frequency
domain, yet all information about how certain frequencies change throughout time is lost. The
STFT tries to solve this issue by dividing the signal into smaller parts and then computing the FFT
for every section. This way, it is possible to analyse how the frequencies of the signal change over
time [26]. Figure 2.3, taken from [5], displays a signal with environmental noise and a significant
event at the 4-second mark. The DFT of this signal would not be able to pinpoint the time at which
the increase of amplitude of the signal occurred. By using the STFT, information about time can
be extracted from the signal, as well as the different frequencies and their amplitudes, which in
2.3 Signal Processing 11

this case, is able to detect an increase in amplitude at 1 kHz at the 4-second mark, which is shown
in yellow, meaning a stronger amplitude, according to the scale on the right of the graph.

Figure 2.3: Example of a noisy signal and its STFT [5]

2.3.5 Wavelet Transform

The Wavelet Transform shares similarities with the Fourier Transform as both methodologies
strive to express a given function using known functions. In the case of the Fourier Transform,
the known functions are sines and cosines, which are non-local and stretch out to infinity. This
fact causes the Fourier Transform to approximate sharp spikes very poorly. Wavelet analysis uses
functions that are contained in finite domains, which are better suited for approximating data with
sharp discontinuities [30]. In Figure 2.4, two functions are displayed, also referred to simply as
wavelets, namely the Ricker and Morlet Wavelets. These examples were generated using the SciPy
Python package with the scipy.signal.morlet and the scipy.signal.ricker method. Using different
Wavelets to calculate the Wavelet Transform produces different results, so which type is selected
influences the subsequent analysis.
12 Background

Figure 2.4: Ricker and Morlet Wavelets generated through SciPy

Using the Wavelet Transform, it is possible to determine how the frequency of a signal changes
over time, while the Fourier Transform provides the prevalent frequencies of a signal, indepen-
dently of its duration. In Figure 2.5, a signal that increases its frequency over time was generated
through SciPy and the Wavelet Transform was computed.

Figure 2.5: Signal with increasing frequency over time and its Wavelet Transform

2.3.6 Empirical Wavelet Transform

The Empirical Wavelet Transform is a recent advancement in signal processing and data analysis
that builds upon the traditional Wavelet Transform. It is a data-driven signal processing tech-
nique that derives wavelet functions directly from the data itself, making it adaptive to the signal’s
specific characteristics. In contrast, the traditional Wavelet Transform uses predefined wavelets,
which may not be as well-suited for non-stationary signals [31]. It outperforms other methods,
such as the Empirical Mode Decomposition (which will be explored in Subsection 2.3.7) and
the methods that derive from it, in theoretical reliability, meaning it has a solid mathematical
foundation and no mode mixing phenomenon (also explored in Subsection 2.3.7). However, the
Empirical Wavelet Transform has some disadvantages, including being considerably affected by
high amounts of noise, being sensitive to modulation effects, requiring a great amount of parame-
ter tuning and being computationally intensive, especially for larger datasets due to its data-driven
nature [31].
2.3 Signal Processing 13

2.3.7 Empirical Mode Decomposition

Empirical Mode Decomposition (EMD) is a signal processing technique that provides a data-
driven way to decompose a complex signal into a set of simpler components known as intrinsic
mode functions (IMFs). EMD perceives a signal as consisting of a fast oscillation superimposed
on a slower one. Once the fast oscillation is extracted, the process treats the remaining slower
component as the new signal and again regards it as a fast oscillation superimposed on a slower
one. This iterative cycle continues until a predefined stopping criterion is met. While EMD
does not employ fixed functions, as seen in wavelets, its fundamental approach shares conceptual
similarities with the wavelet method. Both techniques involve breaking down the signal into its
constituent details and approximations, followed by further decomposition of the approximation
into details and an approximation [32].
Figure 2.6, taken from [33], displays a non-stationary signal, labelled Data, and its decom-
position into several IMFs, which enables the extraction of information about the signal’s time
and frequency characteristics. This method is especially useful for non-linear and non-stationary
data since it is self-adapting according to the analysed signal. Nonetheless, EMD possesses a few
downsides, which include a lack of a sound theoretical basis and mode mixing (when an IMF
obtained through the EMD decomposition exhibits components with varying frequencies) [31].
Other methods have been created which try to supplement EMD, Ensemble Empirical Mode De-
composition, Local Mean Decomposition and Empirical Wavelet Transform (in Subsection 2.3.6),
to name a few.

Figure 2.6: Example Signal and its decomposition into other functions through Empirical Mode
Decomposition, from [33]
14 Background

2.3.8 Hilbert Transform

The Hilbert Transform is a mathematical operation used in signal processing and mathematics
to analyse and manipulate signals and functions, particularly in the context of time-domain and
frequency-domain analysis. The key feature of the Hilbert Transform is its ability to create a
complex-valued signal from a real-valued one, where the real part corresponds to the original
signal, and the imaginary part represents a 90-degree phase-shifted version of the signal. This
complex-valued signal is known as the analytic signal [26]. Equation 2.2 describes the Hilbert
1
Transform, xh (t), which can be thought as the convolution between x(t) and h(t) = tπ .

1 x(τ)
Z ∞
H[x(t)] = dτ (2.2)
π −∞ t −τ

It can be used to extract the envelope and instantaneous frequency of a signal, which are
useful for analysing non-stationary and nonlinear signals. The Hilbert transform can also be used
for time-frequency analysis, phase analysis, and amplitude modulation detection [34].

2.3.9 Hilbert-Huang Transform

The Hilbert-Huang Transform (HHT) is a signal processing method that includes Empirical Mode
Decomposition (EMD) and the Hilbert Transform. It is based on the local characteristic time
scales of a signal and can decompose a complicated signal into a number of IMFs. The IMFs are
extracted in a data-driven manner, which means they adapt to the signal’s inherent dynamics [35].
The uniqueness of HHT lies in its ability to capture both time and frequency domain informa-
tion simultaneously. Each IMF encapsulates a particular frequency component of the signal, and
these frequency components are not fixed to the sampling frequency but rather adapt to changes
in the signal itself. This adaptability makes HHT well-suited for analysing nonlinear and non-
stationary signals [36].
The HHT exhibits similarities to the previous time-frequency analysis methods, as they share
a common objective. Nevertheless, the HHT tends to be employed in scenarios where no previous
knowledge of the signal exists [36]. When utilising the HHT, the other methods are also typically
used and compared to check which approach is best for the situation at hand.
Chapter 3

Related Work

This chapter provides a comprehensive review and analysis of existing literature, research, and
studies relevant to the dissertation’s subject matter. The main research topics revolve around what
components are most often analysed, the types of data used in PdM, Machine Learning methods
typically employed in the field, preprocessing techniques, signal processing methods and hardware
components and solutions matching the dissertation’s goal.

3.1 Critical Components

Critical components encompass the essential elements within a system or process that are crucial
to its proper functioning, efficiency, reliability, and success. In many cases, the breakdown of a
single critical component can have severe consequences, leading to system failures, downtime, or
even catastrophic events [1].
These components can be identified by analysing the frequency of faults in each compo-
nent relative to the downtime caused by those faults [37]. When dividing components using this
method, four quadrants can be described that display the most relevant information for the iden-
tification of the critical components of a production line. Figure 3.1 was adapted from [37] and
illustrates the existence of these four regions using a case study from a production line that used
wheeled mobile robots which recorded the number of faults of each type of component over one
year and the average downtime each fault caused in the production line. For the case study, the
boundaries for the quadrants were defined at 10 minutes of average downtime and 2000 faults in
one year.
The four quadrants characterise the behaviour of components in the context of failure fre-
quency and downtime, offering an insight into their economic efficiency and sustainability con-
siderations.

• 1st Quadrant: components that fail often and cause a high amount of downtime. These
represent the most wasteful components, from an economic and sustainability standpoint,
and, therefore, should be avoided at the machine’s design stage.

15
16 Related Work

Figure 3.1: Number of faults in a production line in 1 year and average downtime caused by such
faults, adapted from [37]

• 2nd Quadrant: components that fail often but require a relatively low amount of time to
be replaced. According to [1], can be maintained using PM without significant production
efficiency losses.

• 3rd Quadrant: components with infrequent failures, resulting in minimal downtime. Their
impact on production efficiency is relatively low, making them suitable for RM strategies.

• 4th Quadrant: components that seldom fail but demand extended downtime when faults
occur. Components that meet these criteria are considered critical components. The rare
faults do not justify replacing these components in the design stage, meaning their impact
has the biggest impact in terms of potential downtime of all parts if third quadrant compo-
nents were avoided at the design stage.

Components that belong to the fourth quadrant tend to be of high mechanical complexity,
such as gearboxes [38] and induction motors [39], but low mechanical complexity critical com-
ponents also exist in machinery, e.g. bearings [40] and gears [41]. Critical components are the
best candidates for PdM at this moment because this type of maintenance strategy is still under
development, and predicting faults in these components brings the biggest benefit to production
lines. This is reflected in articles whose subject matter is condition monitoring, fault detection or
anomaly detection, which tend to focus on monitoring the same kinds of components. [1] reviews
articles regarding machine health and lists the monitored components, and its findings show that
3.2 Data Collection 17

the most frequently monitored components are the ones referring to rotary components and ma-
chining tools and processes, comprising 78% of a total of 93 analysed articles. Gears, bearings,
and shafts represent the most frequently analysed single parts for machine health assessment in the
rotary components group. Bearing faults account for 40% of rotating machinery fault events [31],
which coincides with the literature, as most articles in the dissertation’s area of interest account for
fault detection of bearings. However, it is worth noting that induction motors and gearboxes are
integral components of virtually all industrial machines and hold critical importance. As a result,
they are consistently subjected to condition monitoring and fault detection, as evidenced by [42]
and [43].

3.2 Data Collection

Data collection plays a critical role in condition monitoring and PdM, processes that involve con-
tinuously monitoring the health and performance of machinery and equipment. Effective data
collection is essential for accuracy, enabling the analysis of various parameters and signals to
identify patterns, trends, and anomalies. It represents the foundation for developing predictive
models and algorithms that provide early warnings, enhance equipment reliability, and ultimately
drive operational efficiency.

3.2.1 Vibration Analysis

Vibration analysis is the most widely used approach for PdM. It is primarily applied to rotating
machinery, either continuously or at predetermined intervals, to gather insights into the system’s
current condition [25]. Numerous vibrations are closely associated with periodic events in the
machine’s operation, such as rotating shafts, meshing gear teeth, rotating electric fields, and more.
The frequency at which these events repeat serves as a direct indication of their source, making
frequency analysis the foundation of many robust diagnostic techniques [44].
This form of analysis for machine condition monitoring has a number of advantages com-
pared with other methods. In contrast to temperature or oil analysis, which will be reviewed
further along, it reacts immediately to change, which means it can be used for permanent as well
as intermittent monitoring. This approach allows maintenance to be performed more rapidly on
a critical component, leading to a decrease in cost as there will be less downtime. Certain com-
ponents can be repaired before reaching a state of irreparable damage, extending their useful life.
Another advantage lies in the ability to extract subtle fault indications from background noise and
other masking signals in vibration signals through the application of signal processing techniques
[44], a topic that will be explored further ahead.
For measuring vibration, two types of measurement exist: absolute and relative vibration. Ab-
solute vibration refers to measuring vibration levels with respect to an absolute reference point
outside of the system itself. Transducers that capture this type of vibration are either velocity, e.g.
seismometers, or acceleration transducers, e.g. accelerometers. Relative vibration is, as the name
implies, the measurement of vibration relative to a point within the system under measurement,
18 Related Work

which can captured through the use of proximity transducers [44]. Accelerometers have emerged
as the predominant transducer in condition monitoring research. This preference arises from their
ease of installation in comparison to proximity probes, which necessitate closer proximity to the
specific machine part under inspection, such as a rotating shaft. Proximity probes may require
drilling into certain components or other costly operations. In contrast, accelerometers can be
conveniently placed in more accessible locations while still delivering precise and reliable read-
ings. Proximity probes are also susceptible to other outside factors such as dust, dirt and thermal
expansion [40]. This does not mean that accelerometers are immune to noise or impervious to
environmental factors. The accuracy of their readings can be influenced by the mounting method
used, meaning that careful consideration of the installation is crucial to ensure reliable measure-
ments [45].
Researchers exploring novel signal processing or Machine Learning techniques for fault detec-
tion with vibration data benefit from the availability of reliable datasets that have been extensively
employed in the field. Among these datasets, the Case Western Reserve University Dataset [46]
holds a prominent position, to the extent that there are even reviews dedicated to the algorithms
applied to this dataset, such as [47]. This review focuses on how the dataset was produced, while
comparing it to other well-known ones, e.g. the Paderborn University Bearing Dataset [48] and
the PRONOSTIA dataset [49], and reviewing Deep Learning algorithms that employ these.
In [50], a testing platform consisting of a motor, shaft and load is created and three accelerom-
eters are used to record vibration data for detecting unbalance of the shaft. The data is used as
input for a Convolutional Neural Network while testing different preprocessing techniques. The
best classification model obtained 98.6% accuracy in identifying four different degrees of mis-
alignment. [51] relies on thresholds of vibration for classifying faults according to their severity
of a conveyor belt motor, based on a theoretical model built specifically for the motor’s conditions,
including power and rotation speed.
Additional articles that utilise vibration as its source data to either detect or diagnose faults
will be reviewed throughout the rest of this chapter, but the focus of the review will shift to the
techniques used to process the data.

3.2.2 Acoustic Analysis

Under typical operating settings, most machines produce consistent sound patterns, which can be
recorded as reference. Changes in the reference pattern indicate wear or other types of component
deterioration [25]. At a low-frequency range (which is often considered up to 400 Hz), unbalance,
misalignment and other mechanical construction shifts can be detected [52]. Other types of failures
can be identified at higher frequencies, such as gearbox faults and bearing faults [53] [54].
Using sound data as a means of condition monitoring has two distinct advantages over vi-
bration analysis: fewer costs associated with sensor installation in preexisting machinery [5] and
the lack of physical interaction with the test subject, which does not affect the system’s vibration
dynamics [55].
3.2 Data Collection 19

According to [56], the most popular datasets used in acoustic condition monitoring up until
its publishing were: ToyADMOS [57] and MIMII [58]. The ToyADMOS dataset comprises three
subsets of equipment audio samples in normal and anomalous operating conditions. The subsets
were meant to replicate different scenarios: product inspection, fault diagnosis for a fixed machine
and fault diagnosis for a moving machine. The recordings were of a toy car, toy conveyor and toy
train, supplemented with controlled environmental noise, to reproduce real scenarios more accu-
rately. The microphones used were the SHURE SM11-CN, whose frequency response is charted
between 50 Hz and 15 kHz 1 . Another iteration of this dataset, ToyADMOS2 [59], has been since
published. It complements the first version by possessing controlled domain shifts and higher
variation (machine models and operating speeds) in the same operating conditions (normal and
anomalous). Domain shifts refers to a change between the training dataset used for an algorithm
and the dataset it encounters during deployment. The article also adds a different microphone
model (TOMOCA EM-700) whose frequency response is charted from 20 Hz to 20 kHz 2 . The
MIMII dataset consists of audio recordings sampled in normal and anomalous working conditions
of four types of machines: valves, pumps, fans and slide rails. Each machine’s anomalous record-
ings included different root causes, e.g. for the pump, a leakage state was recorded separately
from a clogged state. The recordings were made using a TAMAGO-03 (with a frequency response
charted up to 16 kHz 3 ) microphone array, which enables multi-channel-based analysis as well as
the more common single-channel-based analysis.
Most articles which use sound data typically use microphones with a charted frequency re-
sponse charted up to 20 kHz, as is the case with all articles which use the aforementioned datasets,
e.g. [60] [61]. Other implementations that exemplify the previous statement are [62], using the
ReSpeaker 4-Mic Array with SPU0414HR5H-SB microphones (up to 10 kHz), and [63], using the
GRAS 40PH-10 microphone (up to 20 kHz). However, acoustic analysis can go beyond human
hearing and into ultrasonic analysis. Ultrasonic sensors can be used to detect other types of faults,
e.g. leakage from water pipes [64] and railway inspection [65]. Ultrasonic analysis is commonly
referred to as Acoustic Emission analysis, which is further explored in the next Subsection.

3.2.3 Acoustic Emission Analysis

Acoustic emission is a phenomenon defined as the propagation of transient stress waves as a result
of localised internal energy, such as a micro-fracture in elastic material [66]. Acoustic emission
waves are not sonic waves but elastic waves in a solid whose frequency range goes beyond the
human audible range (usually up to 20 kHz [67]). Sonic waves higher than the audible range are
defined as ultrasonic waves. Acoustic waves, strictly speaking, are neither ultrasonic nor acoustic
and are commonly detected up to 1 MHz [68].
These emissions can be monitored to detect micro-fractures in many different scenarios, more
commonly applied to structural health monitoring by analysing defects in concrete [69] [70]. It
1 https://pubs.shure.com/guide/SM11/en-US
2 https://www.tomoca.co.jp/brand/tomoca/etc_mic/em-700/
3 https://www.sifi.co.jp/en/product/microphone-array/
20 Related Work

is also used to monitor seismic activity in scenarios such as earthquake detection [71]. The most
relevant use cases are those applied to machine health monitoring, such as bearing condition mon-
itoring [72] [73] and rotating machinery monitoring [74], including gearboxes [75].

3.2.4 Temperature Analysis

Thermography refers to the use of a non-invasive imaging technique that utilises infrared technol-
ogy to capture and visualise the temperature patterns of objects and surfaces. By detecting and
measuring the emitted heat radiation, thermography produces thermal images that depict varia-
tions in temperature across a specific area. It has been used in several different research fields,
including medical imaging, vulcanology and chemical process control, but the most relevant use
case scenario for this project is in condition monitoring.
Any surface can release energy as electromagnetic waves through the radiation process, which
is the foundation of this technology. Objects emit electromagnetic radiation in the infrared region
of the electromagnetic spectrum when their temperature exceeds 0 K [76].
In condition monitoring, this technique has been extensively adopted in electrical equipment
inspection, material fatigue evaluation, weld inspection and printed circuit board (PCB) inspec-
tion, among other varied uses [76]. The biggest advantages of using this technology are the fact
that it eliminates the need for direct physical contact between the Thermography sensor and the
examined component and that several different points of a surface and objects can be measured
simultaneously. These features enable checking a large surface area to check for hotspots or other
temperature-related anomalies [77].
Faults in mechanical and electrical components can lead to a temperature increase in the af-
fected component [76], which makes it a feature suitable for condition monitoring and, therefore,
PdM. The most common use of temperature analysis is through Infrared Thermography (IRT) [1],
which utilises the principle that objects emit infrared (IR) when their temperature exceeds absolute
zero and that the wavelength of the light grows as the object’s temperature increases. This can be
used to visualise temperature changes by displaying different temperatures as different colours, as
displayed in Figure 3.2, but it must be noted that in addition to temperature, surface characteristics,
object shape and view angle also affect the amount of radiation that reaches the thermography de-
vice [25]. Since thermography merely relies on detecting radiation produced from the component
under examination, direct contact with the component is not necessary [77], which reduces the
impact the sensor has on the health signals of the examined component. As opposed to manual
inspection, automatic IR-based condition monitoring does not require a human expert to inter-
pret the thermal images [78], meaning models can be made to automatically detect faults or even
predict them.

3.2.5 Oil Analysis

Oil Analysis is a non-intrusive method that involves periodically sampling and analysing lubricat-
ing oil or hydraulic fluid to detect potential signs of wear, contamination, and degradation within
3.3 Prediction Algorithms 21

Figure 3.2: IRT Image of an AC Motor from [76]

the machinery. By examining the oil’s physical and chemical properties, as well as the presence
of metal particles, contaminants, and additives, oil analysis provides valuable insights into the
condition of critical components such as bearings, gears, and hydraulic systems [25]. Anoma-
lies detected through oil analysis can help identify early signs of equipment deterioration, abnor-
mal wear patterns, and potential failures, enabling proactive maintenance interventions to prevent
costly breakdowns and unplanned downtime.

3.2.6 Electric Current Analysis

Electric Current Analysis relies on monitoring the electrical current consumption of different ele-
ments of a machine to check if a fault has occurred. It is especially helpful in identifying electrical
faults in components such as motor drivers, windings and contactors. Faulty components produce
high currents and winding overheating, which may lead to irreversible damage, making Electric
Current Analysis a valuable tool for preventing such deterioration [79]. Electric Current Analysis
is a frequently employed technique, especially in the context of electric motors, due to its versa-
tility in monitoring multiple components, particularly through the precise examination of motor
current signals. This specialised approach is exemplified in [80], where the current signal of a
permanent magnet synchronous machine is harnessed to successfully identify three distinct types
of faults within the system.

3.3 Prediction Algorithms

Prediction algorithms play a crucial role in the field of PdM, where the goal is to anticipate equip-
ment failures and perform maintenance activities proactively, reducing downtime and optimising
resource allocation.
22 Related Work

3.3.1 Prediction Problem Definition

This project’s ultimate objective is detecting looming faults in the AGV, possibly discerning which
components are about to experience faults. However, due to time constraints, this dissertation only
plans on determining if a fault is about to take place without discriminating if it will affect the
motors or roller bearings or the electrical wiring. In the field of data analysis, this is a problem
known as Anomaly Detection defined as "finding patterns in data that do not conform to expected
behaviour", according to [16]. This definition fits the problem at hand since impending faults can
be predicted from sensor measurements that deviate from the "normal" readings, e.g. the increase
of a roller bearing’s vibration frequency and the temperature increase from a motor driver’s heat
sinks. A noteworthy characteristic feature of Anomaly Detection is that using it enables detecting
faults for components that are not modelled or that do not possess previously researched fault
diagnosis methods. As the case with planetary gearboxes, whose behaviours differ significantly
from fixed axis gearboxes, which invalidates using functioning fault diagnosis methods from the
latter to be used with the former, as noted by [43].
Anomaly Detection can be approached through two primary methodologies: statistical meth-
ods and Machine Learning algorithms. Statistical methods rely on theoretical insights into the
data or utilise simulated models with known behaviour to identify anomalies. In contrast, Machine
Learning (ML) methods do not rely on pre-existing knowledge and can autonomously determine
anomalies within the dataset. Given the absence of a comprehensive model for the AGV based
on theoretical principles and analysis, the utilisation of ML techniques emerges as the most fitting
approach within the context of this dissertation.
It is important to note that Anomaly Detection represents a classification problem since its
primary objective is to determine whether a set of measurements at a particular sample time falls
into one of two categories: anomaly or non-anomaly. Before selecting what ML algorithms to
implement, there is a need to better define the problem at hand, taking into account the lack of
a similar public dataset that encompasses all features of the planned sensor module: vibration,
temperature and sound data. A custom dataset must be created for evaluating the performance of
different algorithms.
ML algorithms for Anomaly Detection can be broadly categorised into three groups, each
of which aligns with the availability of labelled data. These groups, as outlined by [81], are
Supervised Learning, Semi-Supervised Learning, and Unsupervised Learning:

• Supervised Learning involves a dataset in which all classes are explicitly represented. For
instance, in the case of temperature data, a label might indicate that 75.6°C corresponds to
the anomaly class, while 60°C belongs to the non-anomaly class.

• Semi-Supervised Learning, on the other hand, entails a dataset where not all classes are
clearly delineated or represented.

• Unsupervised Learning operates with data that lacks predefined classes, and the algo-
rithm’s task is to autonomously identify patterns and anomalies without explicit labels.
3.3 Prediction Algorithms 23

Considering the project’s ultimate goal of deployment in an industrial setting, noise is a sig-
nificant problem. It cannot be considered white noise since distinctive patterns can emerge from
characteristic machinery parts, such as fans. As previously seen, in the MIMII dataset [58], the
equipment sound recordings were mixed with background noise from multiple factories to sim-
ulate a real environment. In the spectrograms available in the article, it is possible to observe
patterns that do not correspond to either the equipment subject to recording or white noise. The
fact that industrial noise is not Gaussian is rather intuitive as multiple concurrent activities occur
during working hours, e.g. material transportation from and to different stations and changes of
the cycles of production. One way to combat this is to use other types of measurements that are
not affected by background noise to reinforce the built model to better resist noise. An example
of this type of measurement is temperature, considering that a component’s temperature is not
influenced by background acoustic noise in a meaningful way. More sophisticated techniques to
reduce noise will be explored further along.

3.3.2 Statistical Methods

Statistical methods in this context combine signal processing techniques with theoretical knowl-
edge for fault detection. Such is the case with [82], whose objective is to identify faults in roller
bearings using a Microelectromechanical Systems (MEMS) microphone contained in a small
battery-powered module. It records sound in 1-second snippets and converts the audio from the
time domain into the frequency domain using the FFT algorithm combined with a windowing
function to limit the range of the FFT.

 
Z fr d
BPFO = 1 − cosα (3.1)
2 D

In Equation 3.1, BPFO refers to the Ball Pass Frequency of the Outer Raceway of a rolling
element bearing 4 , which describes the resonant frequency of the outer shell of a rolling element
bearing based on its physical characteristics. Z is the number of rolling elements, fr is the inner
rotation frequency, d is the rolling element diameter, D is the pitch diameter and α is the contact
angle. Figure 3.3 displays a schematic drawing of the basic structure of a ball bearing. It is
worth noting that some variables have a different name from Equation 3.1 since the drawing was
taken from a different source. Other resonant frequencies are used, such as the one for the Inner
Raceway and for the Ball Spin Frequency, which have their own equations based on the same
characteristics. [82] calculates these frequencies for the tested bearings and sets a small frequency
window due to possible deviations from the ideal model. The proposed model, as previously
detailed, turns the audio sample into the frequency domain and checks the amplitude of the signal
in the previously calculated frequency windows, and if the amplitude surpasses a certain threshold,
there is a mechanical fault in that specific part of the bearing, such as the outer race. It is important
4A rolling element can be either a sphere or a cylinder
24 Related Work

to note that the algorithm needs calibration to set the minimum threshold since it is highly affected
by background noise.

Figure 3.3: Ball Bearing Internal Structure, from [83]

3.3.3 Supervised Learning Methods

Supervised Learning for classification tasks is based on the premise that the training data contains
instances of all categories, in this case, normal behaviour and all expected anomalies. Many arti-
cles which focus on fault diagnostics tend to use Supervised Learning Methods due to potentially
producing more accurate predictions provided that enough data for training is available [16].

3.3.3.1 Decision Trees

Decision trees are a widely used and intuitive machine learning algorithm for classification and
regression tasks, presenting a flowchart-like structure where internal nodes represent features and
leaf nodes represent class labels or predicted values. The goal is to create a model that predicts the
target variable by learning simple decision rules from the data. Each tree can be seen as a "piece-
wise constant approximation," partitioning the feature space into regions with unique predictions.
The interpretability of decision trees allows for better understanding and insights into the model’s
decision-making process [23]. Due to the latter feature, it is a popular choice of algorithm for
Anomaly Detection. However, it is prone to overfitting and more efficient algorithms that achieve
better results with a similar structure, e.g. Random Forest has overtaken Decision Trees as a choice
in more recent works [81].

3.3.3.2 Random Forest

The Random Forest (RF) algorithm combines multiple Decision Trees in order to increase predic-
tive accuracy and reduce overfitting [24]. By aggregating the predictions of individual Decision
Trees, the RF algorithm mitigates the tendency of any single tree to make biased or erroneous
predictions. Additionally, the term "Random" refers to the fact that this algorithm introduces an
3.3 Prediction Algorithms 25

element of randomness during both tree construction and prediction, which enhances the model’s
robustness.
In [84], the condition of a friction stir welding tool is monitored using vibration signals, where
several Decision Tree-like algorithms are compared, including the RF algorithm. Multiple states
of the tool were recorded, in normal condition, in partially dysfunctional condition and in broken
condition, at two different speeds. When comparing four different ML algorithms, RF obtained
the highest accuracy for classifying the condition of the tool based on the testing data, obtaining a
73.16 % score for the 1400 rpm test and a 93.51 % score for the 1800 rpm.

3.3.3.3 Gradient Boosted Decision Trees

Boosting algorithms, in general, combine weak learners, as in models slightly better than random,
into stronger learners in an iterative way. In the case of Gradient Boosted Decision Trees, the
weak learners are Decision Trees [85]. Gradient Boosting refers to the way the loss function is
computed, relying on gradient descent optimisation. The most popular Gradient Boosted Decision
Trees frameworks for this variety of ML algorithm are XGBoost [86], LightGBM [87] and Cat-
Boost [88]. These are widely spread across multiple fields of study since they can be used in both
classification and regression tasks. According to [85], LightGBM provides the fastest training
times, while CatBoost had the most accuracy for classification tasks. XGBoost produced the most
balanced results. All tests considered hyperparameter tuning, as the results might have been fairly
compared without adjusting each model for the best results possible.
In [89], a wind turbine is simulated, as well as several fault states, including faults in the rotor,
controller and pitch actuator. A model using XGBoost is created in order to diagnose the simulated
faults after a feature analysis was performed to denote which features were the most relevant to
use as input. The article concludes that the created model was adequate for diagnosing faults in
wind turbines, including offshore ones, in several different working conditions. It was also shown
that the model performed better than Random Forest and an SVM which were also tested.

3.3.3.4 Artificial Neural Networks

Artificial Neural Networks (ANN) are a very popular ML model archetype, based on the structure
of biological neural networks. It consists of a network with a large number of simple processors
with many interconnections in parallel. Rather than adhering to rules defined by human experts,
ANNs acquire fundamental principles by learning from a set of symbolic situations provided in
examples given in the training data. Since ANNs are able to learn solely from examples, ideal
for random data, fuzzy data, and nonlinear data, variations of it are widely used in the field of
Condition Monitoring [90], due to some systems’ high complexity and nonlinear behaviour. In
[50], which as been previously reviewed, a Convolutional Neural Network (CNN) is used to detect
if a shaft is unbalanced through vibration data. The article compares using raw sensor data through
the CNN and preprocessing the data before using it as input for the CNN. Although the CNN using
preprocessed data performed significantly better than the raw data CNN, the latter still achieved
26 Related Work

an average of 79% accuracy across all tests, which demonstrates that ANNs and their variations
can obtain satisfactory results from crude data. However, it is demonstrated that preprocessing the
data before inputting it into a CNN can lead to substantial improvements in the performance of the
model, given that the accuracy raised to 97.5 %.

3.3.3.5 Support Vector Machines

The Support Vector Machine (SVM) algorithm aims to find an optimal hyperplane that best sep-
arates different classes, maximising the margin between them. SVMs have been successfully
applied to multiple applications ranging from speech recognition, and face detection to fault di-
agnosis. By using a Kernel Trick, SVM can effectively handle non-linearly separable data, which
is the reason for being able to yield suitable results in such a varied span of research fields [24].
The objective in [91] is to diagnose faults in roller bearings, for which an SVM, in conjunction
with Principal Component Analysis (PCA), is used. This method was chosen due to its non-linear
regression abilities since non-linear data was used, vibration signals. When classifying different
faults in the tested bearings, the lowest accuracy obtained was 96.6%.

3.3.4 Semi-Supervised and Unsupervised Learning Methods

Semi-supervised Learning offers a promising avenue for Anomaly Detection, particularly in sce-
narios where obtaining a large number of labelled anomalies is challenging. This approach lever-
ages a combination of labelled normal samples and a more extensive set of unlabelled data, en-
abling the chosen model to learn meaningful representations of normal patterns [92]. Unsuper-
vised Learning techniques in the realm of Anomaly Detection possess the broadest applicability
due to the aforementioned fact of the rarity of availability of anomalous data while assuming that
the testing data contains a significantly higher frequency of normal instances compared to anoma-
lies [16].

3.3.4.1 Isolation Forest

The Isolation Forest algorithm was specifically developed for Anomaly Detection. It explores a
different approach from those typically used by other techniques to try to solve the Anomaly De-
tection problem: isolating the anomalies explicitly instead of profiling the normal instances [93].
The article that first proposes the algorithm, [93], considers isolation as "separating an instance
from the rest of the instances" and anomalies as "few and different", meaning that anomalies are
prone to isolation. It has successfully been used in the fault detection field, as in [94], where
Isolation Forest was employed to detect faults in an aero-engine. For aeronautical scenarios, it is
difficult to obtain actual fault data, so the model was trained in a Semi-Supervised manner, using
only normal data. For testing purposes, anomalous data was simulated through a State Variable
Model of the engine to mimic most fault states, e.g. faults related to the gas path (an aero-engine
specific part), and through adding noise to the measurements to replicate the impact of vibration.
Four different Anomaly Detection algorithms were used: Angle-based Outlier Detection, Local
3.3 Prediction Algorithms 27

Outlier Factor, One-class Support Vector Machine and Isolation Forest. Out of these, the most
accurate and with the lowest running times was Isolation Forest.

3.3.4.2 One-Class Support Vector Machine

One-Class SVMs differ from SVMs as their objective is to build a boundary between normal
and anomalous data. The fundamental approach involves assessing each test instance to ascertain
whether it falls within the learned region. If a test instance is found to lie within the learned region,
it is classified as normal; otherwise, it is classified as anomalous [16]. This makes its primary use
Anomaly or Outlier Detection. Figure 3.4 shows a One-Class SVM in a two-dimensional space,
where the red line represents the boundary between normal and anomalous data points.
In [95], a One-Class SVM is built using 300 samples of a simulated system, which takes into
account noise, instrument failure, formatting errors and non-representative samples, i.e. misread-
ings. The results of the implementation show that the One-class SVM was robust to outliers but
still sensitive enough to determine if a test instance was anomalous or normal, demonstrating that
this algorithm is adequate for Fault Detection.

Figure 3.4: One-Class SVM Example in 2 Dimensions, from [95]

3.3.4.3 Local Outlier Factor

The Local Outlier Factor is an unsupervised anomaly detection algorithm that assesses the ab-
normality of data points within their local neighbourhood. It calculates the degree of isolation of
each data point by comparing its density to that of its neighbouring points. If a data point has a
significantly lower density compared to its neighbours, it is considered an outlier. It is widely used
in Anomaly Detection and as a comparison metric in articles comparing their proposed method
to more traditional methods, such as the Local Outlier Factor. It is based on a concept of a local
density, where locality is given by x nearest neighbours, whose distance is used to estimate the
density [16].
28 Related Work

The Local Outlier Factor (LOF) is frequently employed as a benchmark when evaluating new
algorithms. For instance, in [94], it falls into a category of traditional methods for anomaly detec-
tion, particularly within the realm of fault detection in engines.

3.3.4.4 Autoencoders

Autoencoders are a type of neural network intended for feature extraction and data compression
and are trained to attempt to copy their input to its output. The algorithm reconstructs its input data
by learning a compressed representation of the input in a lower-dimensional space [92]. Figure 3.5
shows the basic structure of an Autoencoder: the input layer followed by encoding layers to reduce
the dimensionality of the input, the layer containing the least amount of information that the model
needs to classify the input, and decoding layers leading to the output which will try to replicate the
input. During training, the Autoencoder is exposed only to normal data, and it learns to capture
the essential patterns and features of the normal samples. Anomalies, being dissimilar to the
learned normal patterns, result in higher reconstruction errors. This makes Autoencoders sensitive
to anomalies and can effectively distinguish them from normal data. By setting a threshold on
the reconstruction error, data points with higher errors are flagged as anomalies. In [96], a hybrid
model is created that deploys an Autoencoder in conjunction with the STFT to diagnose faults
in rotating machinery, namely a hydro generator rotor, using vibration signals, for which only
normal functioning data was provided. The proposed method outperformed two other models
used for comparison, a modified Autoencoder that uses Softmax Regression and a CNN, in terms
of accuracy. In addition, the models were compared when trained using 20 up to 2000 labelled
samples, with the hybrid method beating the other models at every stage.

Figure 3.5: Autoencoder structure from [97]

3.4 Preprocessing
As with all statistical models, noise, missing values, and other types of disturbances affect the
intended prediction. Certain techniques are employed to reduce the influence these characteristics
3.4 Preprocessing 29

have on the expected results. The set of procedures to make the available data better suited for
the intended use is called Data Preprocessing. Considering the given definition for the problem
in Subsection 3.3.1, Anomaly Detection is the chosen approach to reach this project’s objective.
However, this technique bears an inherent uncertainty: the difference between anomalies and out-
liers [16]. Some articles do not bother to set these apart [16], but these tend to refer to the field
of cybersecurity, where network packets that deviate from the norm might represent suspicious
activity. In this context, separating anomalous behaviour and outliers is essential since the data
under scrutiny is obtained from sensors that occasionally fail or provide a misreading, so a method
to discern these cases is needed to identify actual anomalous behaviour. In the following Sub-
sections, several data preprocessing techniques will be discussed, including noise reduction and
outlier removal.
One aspect of preprocessing that will not be analysed is missing value imputation since one
of the goals of this dissertation lies in building a testbed, meaning the datasets will be created
specifically for this project and not sourced from other sources. This way, missing values can be
avoided.

3.4.1 Outlier Removal

As previously stated, outliers in this context are regarded as misreadings from the sensors, so this
view was taken into account when reviewing relevant articles for adequate techniques.
The Kalman Filter and its variations, e.g. Extended Kalman Filters, are a well-known method
for removing noise from sensor readings. However, these rely on some presumptions that might
not be guaranteed in the dissertation context: Gaussian sensor noise, a control signal and the
system’s dynamics [98]. Since one of the goals of this project is for the model to be agnostic to
the system’s theoretical dynamics, there is no available model and therefore, no control signal or
system dynamics are known to be applied to a Kalman-like Filter. It is also not guaranteed that
the sensor noise follows a Gaussian distribution since the sensors in the use-case scenario will be
exposed to industrial environmental noise, meaning random occurrences might happen from the
sensor’s perspective.
Non-parametric filters, such as the Particle Filter, do not rely on Gaussian distributions for
their sensor readings and can use other probability distributions, including using random sample
points sampled at the time of deployment [98]. This type of filtering method is also not adequate
for the problem at hand due to also relying on the system’s dynamics.
Vibration and sound signals present similar characteristics, as already demonstrated, so anal-
ogous approaches can be applied to remove outliers. In [50], vibration data is used to detect
imbalance in a rotating shaft through the use of ML algorithms. It compares the performance of
a Convolutional Neural Network using the raw sensor data and preprocessed data and also with
a Random Forest on Automatically Extracted Time-series Features. The most relevant aspect of
this article, in the context of this Section, is the first appointed comparison since the preprocessed
data achieved a better result. The method for preprocessing described in the article is the fol-
lowing: sampling 4096 values of vibration data equalling 1 second of data; converting the sensor
30 Related Work

data to the frequency spectrum using the FFT; removing the median component from each Fourier
coefficient and dividing the remaining values by the interquartile values. By analysing the avail-
able code 5 , it is possible to verify that the former operations are executed using the Scikit-Learn
method sklearn.preprocessing.RobustScaler 6 . In [99], the goal is to diagnose faults in a gearbox
using the motor’s (the one attached to the gearbox) current values and a CNN. It is important to
note that these values are for Alternate Current, which share the oscillatory properties of vibration
signals. Before using the current values, a second-order Butterworth low-pass filter with a cut-off
frequency of 500 Hz is used to remove noise from the signal. However, this filter is only applied
after removing outliers which are defined as more than three times the Scaled Median Absolute
Deviations (MADS ) characterised in Equation 3.2, where A is the input data and C is defined in
Equation 3.3.

MADS = C · median(|A − median(A)|) (3.2)

1
C= √ (3.3)
2 · er f cinv( 32 )

In Equation 3.3, the term er f cinv is the inverse of the Gauss error function, er f , which is
described in Equation 3.4.

Z x
2 2
er f (x) = e−t dt (3.4)
π 0

Other types of data used in PdM, such as temperature and direct current, are not waves, and
their measurement readings do not oscillate significantly between samples taken milliseconds apart
as in the case of sound measurements. This means that more standard outlier removal techniques
can be applied to these data types. Assuming the sensor readings follow a Gaussian Distribution, as
they are not subject to as many noise sources as their vibration counterparts, parametric techniques
can be used [16]. One of the simplest is to declare that all measurements that are more than 3σ
(standard deviations) away from the distribution mean µ are outliers since the µ ± 3σ region
contains 99.7% of the data instances. More sophisticated algorithms are used, such as the Local
Outlier Factor [16], which has been reviewed in Subsection 3.3.4.

3.5 Digital Signal Processing

The Signal Processing techniques explored in Chapter 2 are exhaustively used throughout the
literature. Additionally, articles that have already been reviewed make use of them to enhance the
5 https://github.com/deepinsights-analytica/ieee-etfa2020-paper
6 https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.RobustScaler.html
3.5 Digital Signal Processing 31

performance of fault-detecting or diagnosing algorithms. The following Section will focus on the
application of Signal Processing methods and their diverse use cases in augmenting performance.

3.5.1 Discrete Fourier Transform

As seen in the previous chapter, the DFT is widely used in the current technology landscape,
and the field of condition monitoring is no exception. Frequency analysis is essential to several
different model-based and data-driven fault detection approaches, whether in terms of vibration,
acoustic emission or sound analysis, since these physical quantities are generally harder to analyse
using time-series data [29].

3.5.1.1 DFT in Model-Based Analysis

Frequency analysis is at the basis of most articles about condition monitoring that use vibration and
sound as measurements. [100] uses the FFT algorithm to create a model to monitor ceramic ball
bearings in a starved lubrication state based on simple thresholds. Empirical Mode Decomposition
(EMD) is also extensively used in Model-Based methods, such as the case with [101] where EMD
is used in fault detection for roller bearings. In [102], EMD is adopted to de-noise vibration
measurements before using them to train a Probabilistic Neural Network. [31] implements the use
of the Empirical Wavelet Transform (EWT) to build the frequency spectrum of bearing vibration
data and uses a threshold to diagnose the nature of the fault of the affected bearing.

3.5.1.2 DFT in Data-Driven Analysis

[54] uses the common FFT algorithm in order to calculate the Power Density Spectrum, the Root
Mean Square Frequency and the Frequency Center to analyse vibration and sound data to detect
faults in a worm gearbox using an ANN. In [103], vibration measurements are put into the fre-
quency realm using FFT and fed into a predictive maintenance model with a Convolutional Neural
Network to diagnose faults in rotating machinery.

3.5.2 Short-Time Fourier Transform

The STFT is often used in combination with machine learning algorithms to create models capable
of distinguishing anomalous from behaviour. These models learn the typical time-frequency pat-
terns of the signal during normal operating conditions and can then identify deviations or unusual
patterns that may indicate a potential fault or anomaly in the system. In [5], the STFT is used to
convert audio data from the MIMII dataset to the frequency spectrum in intervals of 0.1 seconds
before using the output of the STFT as the input of an Autoencoder.
In [96], the STFT is used as a 2D input for an Autoencoder whose function is to diagnose
faults in roller bearing through vibration data, more specifically using the Case Western Univer-
sity Bearing Dataset [46]. The STFT was favoured over the Wavelet Transform because it demands
less prior knowledge compared to the manual selection of a wavelet that best matches the signal’s
32 Related Work

features, which is required by the Wavelet Transform. The article considers the STFT as a prepro-
cessing step since its output is normalised and only after the latter step is the data fed into the ML
model.

3.5.3 Time-frequency Techniques

In Section 2.3, state-of-the-art time-frequency analysis methods have been explored, apart from
the STFT. This Subsection is dedicated to briefing the use of these techniques in the field of fault
detection. [104] mentions the following methods as being extensively used in the research field of
interest, more specifically using vibration data:

• Empirical Wavelet Transform

• Empirical Mode Decomposition

• Hilbert-Huang Transform

In [105], the EMD decomposes a vibration signal captured from a machine used in a mining
facility into IMFs and faults are detected from the subsequent spectrum analysis created from the
IMFs. [106] simulates vibration signals from defective roller bearings and decomposes the signals
using the EWT, which allowed for detecting early faults through kurtosis parameters and envelope
spectrum of the signals. In [107], a method is developed to improve the performance of the HHT
when diagnosing faults in simulated vibration data from rotating machinery.

3.6 Wireless Connection

In order to fulfil one of the objectives of this dissertation of developing a wireless sensor module
connected to a database for storing its readings. The research fields with the most similarities
to this issue are the IoT field and the Wireless Sensor Networks field. IoT was first defined as
any physical object that sends data and communicates with a network [108]. Wireless Sensor
Networks complement IoT since it is defined as networks of interconnected sensor nodes that
wirelessly communicate with each other to collect and exchange data [109].
IoT devices encompass several different use cases, which means many different wireless com-
munication protocols have been used and developed for multiple scenarios. These protocols are
distinguished primarily by criteria such as range, data rate or throughput, power consumption and
bandwidth [110]. In the dissertation’s scenario, the most important features are low range (less
than 100 m), the highest data rate possible and the lowest power consumption possible. According
to [111], the most used protocols that are considered "low-range" are ZigBee, Bluetooth and WiFi.
Table 3.1 compares the listed relevant features. By analysing the table, WiFi is the most suitable
candidate for its higher data rate and range. Power consumption was not taken into consideration
because the project is not a final implementation, and as such, ease of use was also taken into
account, and WiFi fulfilled all the needed criteria.
3.7 Testing Platforms 33

Table 3.1: IoT communication protocol comparison

ZigBee [112] Bluetooth [113] WiFi [114]

Frequency Band 2.4 GHz 2.4 GHz 2.4 GHz, 5 GHz
Data rate (maximum) 250 kbps 3 Mbps 7 Gbps
Range (maximum) 100 m 30 m 100 m
Power consumption 50 mA 35 mA 245 mA

The application layer, according to the Open Systems Interconnection (OSI) model, is the clos-
est layer to the end user. It serves as the interface between the user applications and the underlying
network infrastructure. For this project, as previously stated, technologies and protocols used for
IoT devices fit the necessary characteristics. Concerning application layer protocols for IoT, the
most popular ones are Message Queuing Telemetry Transport (MQTT), Hypertext Transfer Pro-
tocol (HTTP), User Datagram Protocol (UDP) and Transmission Control Protocol (TCP) [110].
While several other protocols exist, these were the reviewed ones due to their much higher imple-
mentation abundance, which simplifies the research process, as this is not one of the main points of
the dissertation. MQTT is the lesser-known protocol, so a small explanation proceeds: it operates
on a publish-subscribe messaging paradigm, facilitating efficient communication between devices
in the IoT and other distributed systems. The protocol involves two key entities: the MQTT broker
and the MQTT clients. The broker acts as an intermediary, receiving messages published by the
clients and distributing them to subscribed clients based on predefined topics. MQTT clients can
either publish data to specific topics or subscribe to topics of interest. When a client publishes a
message to a topic, the broker receives it and forwards it to all subscribed clients interested in that
topic. This lightweight and efficient approach minimises network overhead and conserves power,
making MQTT ideal for IoT applications with limited resources and intermittent connectivity 7 .

3.7 Testing Platforms

All of the methods for detecting and diagnosing faults in machinery previously introduced require
a platform to validate them. While a significant portion of the articles in the literature rely on data
gathered from external sources, such as public datasets (e.g., [46]), some researchers construct
their own customised testbeds tailored to the specific requirements of their proposed methods.
The majority of articles that employ their own testbeds have very similar characteristics, with the
main difference being what component is being monitored, mostly consisting of either bearings or
gearboxes. [49] encapsulates the trends of bearing testing equipment and procedures, as PRONOS-
TIA was developed as a platform for testing PdM algorithms for the IEEE PHM 2012 Prognostic
Challenge for the 2012 Prognostics and Health Management Conference 8 . The testbed, displayed
in Figure 3.6 contains an AC motor, a gearbox, a bearing, accelerometers, a temperature sensor, a
data acquisition card as well other components to couple the mechanical elements.
7 https://mqtt.org/
8 https://phmsociety.org/first-european-conference-of-the-prognostics-and-health-management-society-2012/
34 Related Work

Figure 3.6: PRONOSTIA testbed from [49]

In the PRONOSTIA, the AC motor has a power of 250 W and a rated speed of 2830 rpm. The
gearbox converts the rated speed into less than 2000 rpm while maintaining the rated torque. The
setup allows the radial force exerted on the bearing to reach its dynamic load limit of 4000 N. The
rotational speed is sampled at 100 Hz, the bearing’s vibration at 25.6 kHz and the bearing’s external
ring temperature at 10 Hz. Two one-axis accelerometers are used and placed over the bearing’s
outer race, one on the horizontal axis and the other on the vertical axis. The accelerometers’ model
is not specified in the article, but it can be concluded that its bandwidth reaches 12.8 kHz due to the
Nyquist Sampling Theorem. For the temperature measurements, an RTD (Resistance Temperature
Detector) is used, yet it is also unknown the exact model.
The PRONOSTIA testbed and subsequent dataset is one of the most popular in the literature,
however, some articles either custom testbeds are used or opt for a commercial solution. The most
prevalent commercial solution is made by SpectraQuest [115]. Multiple models are available; the
basic one, named Machinery Fault Simulator, is displayed in Figure 3.7.
The basic simulator includes a 750 W motor and all parts associated with its control, a re-
configurable shaft mount for different sizes and positions, and vibration isolation mounts, among
other features to test different elements, such as roller bearings and gearboxes. This simulator
has been successfully utilised in several articles, among them [31], [99] and [36], and is most
commonly employed for testing heavy machinery components.

3.8 Summary

This chapter delved into the core concepts employed in the relevant literature. The exploration
encompassed a wide spectrum, ranging from identifying the most suitable components for PdM to
exploring the latest advancements in ML algorithms within the domains of Condition Monitoring
and Anomaly Detection. The realm of preprocessing techniques was also examined, which play
3.8 Summary 35

Figure 3.7: SpectraQuest Machinery Fault Simulator [115]

a pivotal role in enhancing model performance, especially in noisy environments. Furthermore,

prevalent data types were scrutinised and an in-depth analysis of IoT protocols for wireless com-
munication warranting comprehensive coverage. Finally, the most used testing platforms were
analysed to determine the common elements between them.
36 Related Work
Chapter 4

Methodology

This chapter focuses on detailing the approach and techniques employed to implement the pro-
posed objective. As in the previous Chapter 3, a careful review of which sensors fit the proposed
objective while keeping in mind the cost and measurement fidelity. Furthermore, this chapter
delves into the communication protocols established for data exchange between the microcon-
troller and the sensors, as well as between the microcontroller and the database. It offers an
overview of the software designed to run on the microcontroller and for the analysis of the cap-
tured data. Lastly, it outlines the preprocessing steps applied to the collected data and introduces
the models employed.

4.1 Prototype Design

This section will dwell on the hardware design choices for the sensor module. More specifically,
by analysing different sensor specifications and other details concerning the interface between the
microcontroller and the sensors.

4.1.1 Vibration Sensors

Vibration can be expressed in three parameters: translational displacement, velocity, and acceler-
ation [44]. Translational acceleration provides information on the rate of change in translational
velocity, which describes the speed of displacement. Displacements are larger at lower frequen-
cies and considering that the vibration frequency intervals that better indicate deterioration in the
most common machinery fault cause, bearings, are above 150 Hz [116], measuring acceleration is
the best method to measure a body’s vibration [117].
In order to measure acceleration, displacement transducers or accelerometers can be used.
Displacement transducers, such as proximity probes, measure the distance a body has moved in
a straight line in relation to the transducer which means that the sensor must be placed in a static
position in relation to the ground [44]. This is not possible in the context of the project’s objective
making accelerometers the only viable option. Accelerometers typically comprise of a suspended
mass attached to a spring and damper mechanism housed inside a casing. The accelerometer is

37
38 Methodology

firmly affixed to the object experiencing acceleration and its readings rely on Fa = M ẍ, being Fa
the force on the body caused by the acceleration, ẍ, on the mass, M [117].
The most common accelerometers for vibration measurement in the project context are piezo-
electric and MEMS accelerometers [118]. Piezoelectric accelerometers use the piezoelectric effect
to output a voltage proportional to the acceleration, thus providing a way to measure it [44]. This
category of accelerometer typically provides a measurement resolution of 0.1% of full-scale, with
an accuracy of approximately ±1%. These devices are available in a range of configurations to
accommodate a broad spectrum of measurements and frequencies, spanning from 0.03g full-scale
up to 1000g full-scale and from 1 Hz up to 30 kHz, with some reaching up to 90 kHz [118]. Mi-
croelectromechanical systems (MEMS) accelerometers are comprised of a micro-machined proof
mass that is suspended between two parallel plates from which acceleration is obtained [119].
This type of accelerometer is becoming more popular in the field of condition monitoring as this
technology had relatively high internal noise [118], but low-noise MEMS accelerometers are be-
coming available. The frequency response of these accelerometers are usually much narrower than
those of piezoelectric accelerometers, but those of comparable characteristics are much lower in
cost.
Taking into consideration the previous statements, the vibration transducers selected were
MEMS accelerometers, namely the ADXL355 [120], using the EVAL-ADXL355Z evaluation board
in Figure 4.1, and the MMA8451 [121], using the Adafruit breakout board in Figure 4.2. Table 4.1
compares the chosen sensors using the most relevant parameters, particularly sensitivity, noise
density and bandwidth.

Table 4.1: Feature comparison between selected accelerometers

Accelerometer MMA8451 ADXL355

Num. of Axis 3 3
Sensitivity 4096 counts/g 256 000 LSB/g
√
Noise Density (µg/ Hz) 99 22,5
Bandwidth (Hz) 400 2100
Price 1 7.39 C [122] 43.46 C [123]

Figure 4.1: ADXL355 Evaluation Board - EVAL-ADXL355Z

1 At the time of writing

4.1 Prototype Design 39

Figure 4.2: MMA8451 Breakout Board from Adafruit

The Sensivity parameter in Table 4.1 has different scales for each accelerometer since it was di-
rectly transcribed from their respective datasheets. By further analysing the datasheets, the scales
express the same unit of measurement, the number of digits used to divide each g (referring to
the force per unit mass due to gravity at the Earth’s surface, 9.80665 ms−2 ). With this informa-
tion, it is possible to state that the ADXL355 is several times more sensitive than the MMA8451.
Concerning the other parameters in the Table, the ADXL355 has lower noise density and higher
bandwidth. These distinctions are justifiable, given the price differential between the two devices.
The MMA8451 was selected due to the objective of creating a cost-effective sensor module and
would serve as a comparison to an accelerometer with more adequate mechanical characteristics,
such as the ADXL355. For simpler use of the accelerometers, breakout boards were chosen since
they remove the need to solder Surface Mount Technology components and the need to build the
interaction circuits. However, another advantage that is often overlooked is the fact that the evalu-
ation boards are built with the mechanical structural integrity required for accurate accelerometer
readings per the manufacturer’s specifications, while a custom solution would have to follow the
compulsory guidelines which would delay transducer deployment.

4.1.2 Microphones

As seen in Chapter 3, sound analysis is a well-explored method to monitor the condition of me-
chanical parts that are included in the critical components group. The microphones that are usually
chosen appear to be ones that are readily available without much thought given to how the sonic
transducer’s physical characteristics might affect the measurements and, consequently, the chosen
algorithm’s results. Nonetheless, it is important to note that this strategy has produced effective
results, as remarked previously in the cited articles, which means there is little incentive to switch
approaches. As the ultimate goal of this project is the deployment of the prototype in an environ-
ment with very high levels of noise (an industrial setting), more thought was put into choosing the
right microphones for the situation.

4.1.2.1 Electret vs. MEMS

The most adequate microphone types for this prototype are of the Electret and MEMS varieties
due to their relatively small dimensions when compared to other technologies, since Electret mi-
40 Methodology

crophones do not exceed 8 mm in diameter and 4 mm in length and MEMS microphones do not
have any size measurement over 4 mm. Although Electret microphones are the most popular
choice, MEMS microphones are the better option for the use case scenario, due to their compact-
ness, ruggedness, and resistance to environmental factors and lower power consumption. MEMS
microphones also posses better noise performance and a higher signal to noise ratio (SNR) in re-
lation to Electret ones [124]. Additionally, it is noteworthy that MEMS microphones can be used
as ultrasonic transducers, as some models present its frequency response over the human hear-
ing range of 20 kHz. [125] used MEMS microphones to build an ultrasonic microphone array to
inspect concrete for faults using ultrasonic wavefield imaging.

4.1.2.2 Microphone Selection Criteria

With hundreds to thousands of MEMS microphone models to choose from, guidelines need to be
established in order to select the one with the most appropriate features for the use case scenario.
Two considerations other than environmental resistance have already been determined: noise per-
formance and SNR. One way to create guidelines is to check the specifications of the microphones
used in the already reviewed literature. As many of the previously cited articles use the same three
datasets, MIMII [58], ToyADMOS [57] and ToyADMOS2 [59], Table 4.2 compares features from
each microphone used. The model used in [63] was added to the table since a specialised measure-
ment microphone was chosen while the other articles used general-purpose microphones for video
recording usage. The MIMII dataset uses the TAMAGO-03 microphone array but the datasheet
was not available, so all values are missing from Table 4.2. However, it is possible to infer that the
frequency range is probably up to 20 kHz since the product page 2 lists the product as a device to
aid video calls.
A quick examination of Table 4.2 reveals that most information is not available to the con-
sumer, including SNR and noise performance. These two parameters tend to be only displayed
in the datasheet for the sound transducer itself, for which consumer microphones do not list, so it
is not possible to know these unless directly contacting the manufacturer. This is the justification
for the prior statement relating to the lack of criteria for selecting microphones, given that certain
characteristics that greatly affect measurements are not available and those microphones are still
used. The Shure SM-11 and the GRAS 40PH are the ones with the most available information, so
a deeper analysis can be performed.
Regarding sensitivity, the GRAS 40PH has much higher sensitivity, since the higher the output
voltage for the same pressure, the higher the sensitivity and the GRAS has an output voltage over
150 times higher than the SM-11. Microphone output impedance affects the amount of noise and
signal quality that the microphone’s output endures and the lower the output impedance, the lower
the noise and higher the signal quality. Frequency response details how the output varies across
the frequency spectrum. The relative superiority of the GRAS 40PH over the SM-11 is to be
expected, because the first one was developed to be a measurement microphone, meaning its goal
2 https://www.sifi.co.jp/en/case/hylable/
4.1 Prototype Design 41

is to capture sound in the most accurate way, while the SM-11’s objective is to capture the human
voice in the clearest way, minimising unwanted environmental noise.

Table 4.2: Feature comparison between microphones used in cited datasets

Dataset ToyADMOS ToyADMOS2 MIMII Milling tool

Microphone Model Shure SM11 EM-700 TAMAGO-03 GRAS 40PH
Sensitivity 3 0.6 Not available Not available 100
Output Impedance (Ω) 150 400 Not available >50
Frequency Range (Hz) 50 - 15 000 20 - 20 000 Not available 50 - 20 000
Frequency Response Not available Not available Not available Figure 4.3
SNR Not available Not available Not available Not available
Noise Performance Not available Not available Not available Not available

Figure 4.3: GRAS 40PH Microphone Frequency Response from [126]

The final overview for the selection criteria is that these should include the frequency response,
sensitivity, frequency range and SNR, if available. These features describe how the microphone’s
characteristics influence readings and the considerations that should be taken in order to retrieve
measurements with the highest amount of fidelity.

4.1.2.3 Chosen Microphones

One of the objectives of this dissertation is to make use of cost-effective sensors in the final so-
lution. With this goal in consideration, the SPH0645LM4H digital microphone [127] (using a
breakout board from Adafruit 4 ) was selected. Even though certain criteria were delineated in
SubSubSection 4.1.2.2, the selection of this microphone was predominantly driven by its cost of
6,46 C [128], at the time of writing. Another significant factor contributing to this decision was the
absence of established criteria in the literature for selecting a suitable microphone while consis-
tently achieving accuracies surpassing 90% in the final models. This gap in guidance prompted the
exploration of potentially employing a microphone originally designed for smartphone integration
in the context of detecting anomalies within mechanical systems.
3 (mV/Pa @ 1 kHz)
4 https://www.adafruit.com/product/3421
42 Methodology

Figure 4.4: SPH0645LM4H Microphone Breakout Board from Adafruit

The SPH0645LM4H uses the Inter-IC Sound (I2S) communication protocol. Figure 4.5 dis-
plays the microphone’s frequency response plotted up to 10 kHz. The labels on the graph refer
to the frequency of the microphone’s internal clock, which is set by the BCLCK frequency of the
I2S protocol that the controller device provides, according to the datasheet [127]. The standard
frequency is 3.072 MHz, meaning the frequency response is characterised by the light blue line,
according to the label. Table 4.3 exhibits the other main characteristics of the microphone.

Figure 4.5: SPH0645LM4H Microphone Frequency Response, from [127]

Table 4.3: SPH0645LM4H’s Main Features

Microphone SPH0645LM4H
Sensitivity (dBFS @ 1 kHz) - 25
Bandwidth (Hz) 50 - 15 000
SNR (94 dB SPL @ 1 kHz) (dB(A)) 65

As previously mentioned, especially in Subsection 3.2.3, some articles have found better re-
sults regarding anomaly detection when analysing vibration frequency at higher frequency ranges,
which is the basis for Acoustic Emission. However, all of the cited articles that use sound anal-
ysis, tend to adhere to utilising consumer microphones, made for recording human perceptible
sounds, up to 20 kHz. This might be due to the relatively higher prices of ultrasonic microphones
4.1 Prototype Design 43

and limited available selection. Moreover, the name "ultrasonic microphone" is hardly used, with
the term "ultrasonic transducer" being much more popular. Furthermore, "ultrasonic transducer"
is usually associated with devices for measuring distance between the sensor and the object, e.g.
the ever-popular HC-SR04 5 . These sensors could be used for detecting ultrasonic waves created
by faulty parts, but their frequency response is only typically charted between a small interval of
frequencies, as with the CUSA-R75 whose frequency response is plotted between 37 and 43 kHz.
This fact is detrimental to the experiment since the exact frequency of the faulty part sound signal
is not known and a relatively small interval of captured signal might not acquire the sought-after
signal. It has also

At the time of writing, MEMS technologies have enabled cheaper microphones to be mass-
produced, with very few of the models having a charted frequency response above 20 kHz. How-
ever, these do exist and are available at a low enough price to be included in the project’s objec-
tives, which include using low-cost sensors. The selected ultrasonic MEMS microphone was the
SPU0410LR5H-QB [129]. Figure 4.6 shows the selected microphone in a breakout board since
MEMS components are very small when compared to "regular" components. The SPU0410LR5H-
QB’s length is 3.76 mm, which makes it very difficult to solder and a breakout board is needed for
easier prototyping. Another important feature to note is that the SPU0410LR5H-QB’s output is
an analog voltage, with a maximum peak-to-peak voltage of, so an amplifier circuit is needed to
fully utilise the Analog-to-Digital Converter’s (ADC) resolution capabilities.

Figure 4.6: SPU0410LR5H-QB Microphone Breakout Board

Table 4.4 lists the main characteristics of the ultrasonic microphone. These were not exten-
sively selected due to the very low amount of MEMS ultrasonic microphone models available. In
Figure 4.7, the frequency response is plotted, from 10 to 80 kHz and a cursory look at the chart
reveals that at these frequencies, the output is not linear throughout the interval and this fact must
be taken into account when amplifying the analog output voltage and when processing the data.

5 https://cdn.sparkfun.com/datasheets/Sensors/Proximity/HCSR04.pdf
44 Methodology

Table 4.4: SPU0410LR5H-QB’s Main Features

Microphone SPU0410LR5H-QB
Sensitivity (mV/Pa @ 1 kHz) 12.6
Output Impedance (Ω) 400
Frequency range (Hz) 100 - 80 000
SNR (94 dB SPL @ 1 kHz) (dB(A)) 63
Price 6 14.76 C [130]

Figure 4.7: SPU0410LR5H-QB Microphone Frequency Response over 10 kHz [129]

4.1.2.4 Analog Signal Amplification

Analog signal amplification is a crucial process in the context of MEMS microphones, which are
highly sensitive transducers used to convert sound waves into electrical signals. Due to the small
size of MEMS microphones, the generated electrical signals are typically very weak. To ensure
accurate and reliable signal detection, the analog signals from the microphone need to be amplified
before further processing.
The SPU0410LR5H-QB microphone outputs an analog voltage, so two different steps are re-
quired before having audio data capable of being analysed by an anomaly detection algorithm:
amplifying the output and sound digitisation. In this Subsection, the first step’s proposal is delin-
eated.
The first step in designing an amplifier circuit is to check the requirements. Further ahead will
be made clear the choice of ADC, which has a voltage range between 0 and 3.3 volts. The output
of a microphone is a symmetrical sinusoidal signal centred around 0 V, meaning the output must
be given an offset of half the maximum input value of the ADC in order to take advantage of the
full scale of the ADC. The next step is to check the microphone’s characteristics, available in its
datasheet [129]. The DC output with a supply voltage of 1.5 V is typically 0.73 V. The output
6 At the time of writing
4.1 Prototype Design 45

sensitivity at a sound pressure level of 94 dB at 1 kHz has a maximum of -35 dBV/Pa which is
equivalent to 17 mV/Pa. Since the selected microcontroller ADC’s maximum input is 3.3 V and
considering the 1.65 V offset, the gain the amplifier needs to provide is 97. It is necessary to
point out that the output is not constant, as remarked by the frequency response graph, present
in Figure 4.7, but the gain must take into account the maximum output so as not to oversaturate
the audio, which causes loss of information, also known as clipping. Amplifier circuits for MEMS
microphones are not generally not publicly accessible, but in [131], an open-source acoustic device
for monitoring biodiversity is detailed, including schematics for an amplifier circuit. One of the
objectives of the device is to detect ultrasonic sounds produced by bats, so it uses a previous
iteration of the model of the microphone chosen for this dissertation, the SPM0408LE5H-TB.
The final amplifier circuit design for this project is presented in Figure 4.8 and follows the
same configuration of two inverting operational amplifier circuits, using the OPA350 [132], as the
√
aforementioned one present in [131]. The OPA350 was selected due to its noise of 5 nV/ Hz.
Although the use of a single operational amplifier appears to be intuitive, it was observed that
that configuration would introduce an amount of electrical noise in the circuit’s output that would
completely obscure the sound signal’s voltage.
A further inspection of the frequency response chart in Figure 4.7 reveals that the sensitivity is
very uneven over the charted frequency values, and a peak exists around 25 kHz. Since the goal of
using this specific microphone is to capture ultrasonic waves, a passive band-pass filter was made
to ensure the readings are not overly influenced by signals in the 10 to 40 kHz range. The region
of interest and of lower variation is between 45 and 80 kHz, which was the band chosen for the
signal to pass through. A passive band-pass filter can be made using a passive low-pass filter and
a passive high-pass filter directly connected to each other. The filter’s resistance and capacitance
were calculated using Equation 4.1, where fc represents cut-off frequency. This calculation was
performed twice: once for the low-pass component and once for the high-pass component of
the filter. The filter is connected to the output of the microphone, as seen in Figure 4.8, and its
components are identified by either bp_h (band-pass high) or by bp_l (band-pass low).

1
fc = (4.1)
2πRC

Table 4.5 contains all the values of the passive components used in the amplifier circuit in
Figure 4.8. The SPU0410LR5H-QB’s datasheet mentions that capacitors placed close to the mi-
crophone should not contain Class 2 dielectrics, which are a type of dielectric used in certain
ceramic capacitors. This recommendation is due to the Piezoelectric Effect in this type of capac-
itor, as current passes through the capacitors, these expand and contract, which causes vibration
and affects the signal captured by the microphone [133]. The designed circuit will be built on
a breadboard, so this recommendation was ignored since the placing of the components is much
more unstable and susceptible to other sources of vibration than the Piezoelectric Effect affecting
the capacitors.
46 Methodology

Figure 4.8: Microphone Amplifier Circuit

Table 4.5: Amplifier Circuit Passive Component Values

Passive Component Value

Rmic_1 12 k Ω
Rmic_2 10 k Ω
Rbp_h 1k8 Ω
Rbp_l 2k2 Ω
R1 1kΩ
R2 10 k Ω
R3 1kΩ
R4 10 k Ω
R5 22 k Ω
Cbp_h 1 nF
Cbp_l 2.2 nF
C1 100 nF
C2 100 nF
C3 100 µF

4.1.3 Temperature Sensors

Certain applications necessitate the use of non-contact temperature sensors for measuring com-
ponent temperatures. One such scenario arises when assessing the inner raceway of a bearing.
Utilising a contact sensor in this context can disrupt the system’s dynamics, leading to undesirable
inaccuracies that fail to represent the system’s true performance accurately. IR thermometers are
used as non-contact sensors since they possess the ability to detect the IR radiation from objects,
taking into account the object’s emissivity, as seen in Subsection 3.2.4.
4.1 Prototype Design 47

Keeping in line with one of the goals of this dissertation, an IR thermometer was chosen,
the MLX90614 [134], displayed in Figure 4.9. Table 4.6 presents the key specifications of the
thermometer. The "Sensor Temperature Range" specifies the ambient temperature range within
which the sensor is guaranteed by the manufacturer to operate within its stated specifications. The
chosen breakout board is sourced from seeed studio7 , selected primarily for its ready availability
and lower cost.
Table 4.6: MLX90614 Specifications [134]

Sensor Temperature Range -40 to +125 ºC

Object Measurement Range -70 to +380 ºC
Accuracy ± 0.5 ºC
Resolution 0.02 ºC
Field of View 35º
Refresh Rate 0.5 to 64 Hz
Digital Communication Protocol I2C
Price 8 25.14 C [135]

Figure 4.9: MLX90614 IR Thermometer

4.1.4 Microcontroller
According to the chosen sensors, the prototype’s microcontroller needs to fulfil certain require-
ments, namely:

• Support for Inter-Integrated Circuit (I2C)

• Support for Serial Peripheral Interface (SPI)

• A CPU capable of processing data with sampling rates over 1 MHz

• A high-resolution integrated ADC is not mandatory but welcomed

The selected microcontroller is the Teensy 4.1 9 , displayed in Figure 4.10 from PJRC. It has an
ARM Cortex-M7 CPU [136] with a maximum clock speed of 600 MHz, 2.14 DMIPS/MHz and
7 https://www.seeedstudio.com/Grove-Thermal-Imaging-Camera-MLX90614-DCC-IR-Array-with-35-FOV-p-

4657.html
8 At the time of writing
9 https://www.pjrc.com/store/teensy41.html
48 Methodology

includes a Digital Signal Processing hardware accelerator. It also possesses three units capable
of communicating over I2C and two others capable over SPI, although either protocol is able to
communicate using only one bus for multiple devices [137]. The last feature worthy to point out
is the fact that the Arduino framework 10 is available, which greatly simplifies the programming
of the microcontroller as the Arduino libraries are available.

Figure 4.10: Teensy 4.1 Microcontroller

4.1.5 Voltage Incompatibility

Some of the chosen sensors, such as the MLX90614, operate at 5 V, whereas the Teensy 4.1
microcontroller is designed to interface with 3.3 V compatible devices. Attempting to directly
connect these sensors to the Teensy without proper voltage level translation may result in potential
issues such as voltage mismatch and could lead to malfunction or damage to both the sensors and
the microcontroller.

Figure 4.11: Sparkfun Bi-directional Logic Level Converter

Given that the digital communication protocols for the mentioned sensors are all based on
I2C, this must be taken into account for choosing a method for converting the 5 V signals into
3.3 V and vice-versa. While alternative methods may come to mind, utilising a dedicated logic
level converter is the less complex and more practical solution. These devices are specifically
designed to ensure reliable voltage translation, minimising the risk of communication errors or
10 https://www.arduino.cc/
4.1 Prototype Design 49

damage to the components. The SparkFun Bi-directional Logic Level Converter [138], in Figure
4.11 provides these advantages while also being priced at 2,74C.

4.1.6 Analog-to-Digital Conversion

Considering the choice of the SPU0410LR5H-QB ultrasonic microphone, it becomes necessary
to utilize an Analog-to-Digital Converter (ADC) to convert the analog output signal from the
amplifier circuit into a digital format for analysis alongside other sensors. The Teensy 4.1 micro-
controller is equipped with a 12-bit ADC that can achieve a sampling rate of 1 million samples per
second, as per its datasheet.
While external ADCs with higher resolutions (e.g., 16-bit and 20-bit) and faster sampling
rates are available, the decision to opt for the integrated Teensy 4.1 ADC was made due to the use
of a 12-bit ADC in [131], and the incorporation of an external ADC would introduce additional
implementation complexities.
It’s worth noting that while the Teensy 4.1 is supported by the Arduino Framework, the stan-
dard library does not encompass all of the microcontroller’s functionalities. Therefore, the integra-
tion of an external library is necessary. Fortunately, an open-source library recommended directly
by the manufacturer, known as the ADC library 11 , proves to be a valuable solution in this context.

4.1.7 Hardware Connections

Figure 4.12 contains a diagram of the hardware connections between the Teensy 4.1 and the se-
lected sensors. It is worth noting that the diagram does not represent the physical location of the
pins on the microcontroller; the labels are meant to aid in analysing the connections.

4.1.8 Local Processing

By selecting the Teensy 4.1 microcontroller, it is possible to take advantage of the ARM Cortex-
M7 processor’s floating point unit, which enables the use of the CMSIS-DSP 12 library from ARM
Software. This library is an optimized compute library for embedded systems which takes advan-
tage of the available hardware and improves speed and efficiency for certain algorithms, such as
complex and real FFT, matrix functions (e.g. inverse matrix and matrix multiplication), quaternion
functions and statistical functions (e.g. root mean square). It also has support for SVM run-time
optimization, if the training stage is performed using scikit-learn.
Other libraries were also considered, namely the ArduinoFFT 13 and the FFTW [139] library.
The first one due to its ease of use, wide range of examples and documentation and the fact that it
is meant for the Arduino Framework which makes it more intuitive while using this framework for
the rest of the main program. The second one was considered by virtue of its efficient performance
14 , but its implementation is not optimized for embedded CPUs and the CMSIS-DSP FFT set of
11 https://github.com/pedvide/ADC
12 https://github.com/ARM-software/CMSIS-DSP
13 https://github.com/kosme/arduinoFFT
14 https://www.fftw.org/speed/
50 Methodology

Figure 4.12: Sensor Module Hardware Connection Diagram - Without ESP-12E

functions uses the Cortex-M7’s capabilities to their most efficient operation.

4.1.9 WiFi Connection

One of the primary goals of this dissertation is to establish a wireless sensor module capable
of connecting to a database for storing its readings. The chosen physical layer protocol for this
purpose is WiFi, owing to its suitable range and data transfer rate, which align well with the
project’s environmental conditions and use case requirements. The selected microcontroller, the
Teensy 4.1, does not have WiFi capabilities, so a separate microcontroller is needed in order to
transmit data wirelessly. One of the most popular WiFi chipsets available is the ESP8266 [140],
and the selected microcontroller which contains the chipset was the ESP-12E [141] (in Figure
4.13) due to its immediate availability. At the time of writing, newer and more capable models
have been launched, e.g. the ESP32-C6 [142], but the ESP-12E is supported by the Arduino
Framework, which simplifies programming and enables access to the existing libraries made for
that framework. It is also noteworthy the amount of documentation which is available for this
microcontroller, which simplifies troubleshooting and implementation of programs.
The ESP-12E is a microcontroller, however, it is not suitable for the project’s requirements,
since it lacks processing power, enough SPI and I2C connections, ADC speed and accuracy and
digital ports for all the selected sensors. So, the best possible combination of hardware for the
project’s needs is to merge both microcontrollers, using the Teensy 4.1’s processing and I/O char-
acteristics and the ESP-12E WiFi connectivity. A way to communicate between the Teensy 4.1
and the ESP-12E is needed, for which three possible approaches exist: UART, I2C or SPI, as both
4.1 Prototype Design 51

Figure 4.13: ESP-12E Microcontroller

microcontrollers share these communication methods. Table 4.7 compares the most relevant char-
acteristics in relation to each other, namely, speed, ease of use and the number of wires needed
between the microcontroller boards.

Table 4.7: Serial Communication Protocol Comparison

Protocol UART I2C SPI

Relative Speed Slowest Intermediate Fastest
Relative Ease of Use Easiest Hardest Intermediate
Number of Wires 2 2 4

After analysing Table 4.7, the most suitable protocol is either UART, due to its ease of im-
plementation and minimal connection usage, or SPI, which offers relatively higher data transfer
speeds. In this context, UART has been selected as the preferred protocol, considering that this
project primarily serves as a proof of concept and does not demand the utmost performance. Thus,
opting for simpler implementation over faster communication is a valid and pragmatic decision.

The ESP-12E lacks a straightforward means of connection to a computer in order to be pro-

grammed, so external hardware is needed. It is also important to note that the microcontroller has a
preloaded program which enables simple actions via a Universal Asynchronous Receiver/Transmitter
(UART) connection using AT Commands to control the ESP8266 chipset, e.g. connecting to an
access point with AT+CWJAP or sending data with AT+CIPSENDEX. This method of interac-
tion is more restrictive and less user-friendly than simply programming it, which can be achieved
via its UART pins and setting some pins to HIGH or LOW to put it in programming mode. The
FT232RL [143], in Figure 4.14 is a Universal Serial Bus (USB) to Serial adapter which enables
the programming of the ESP-12E using a computer with a USB port.

Figures 4.15 and 4.16 display the hardware connections in order to program the ESP-12E
and to connect to the Teensy 4.1, respectively. It is important to point out that although these
connection schemes function for interacting with the ESP-12E, the connection is unstable and
requires extra capacitors to negate instability in the power supply.
52 Methodology

Figure 4.14: FT232RL USB to Serial

Figure 4.15: ESP-12E Programming Mode connection diagram

Figure 4.16: ESP-12E Serial connection with the Teensy 4.1

The Teensy 4.1 is equipped with eight serial ports, and to transmit data to the ESP-12E through
one of these ports, the function Serial1.print is employed. To verify the successful transmission
of data, the ESP-12E can utilise the Serial.read function. For error checking and to ensure all
4.2 Database 53

information has been sent and correctly received, an application layer must be created, which has
not been accomplished, but small tests allowed for verifying that this method is valid.

4.1.10 Overview
Figure 4.17 contains a simplified diagram with the selected sensors and the connection method
from the microcontroller to the sensors. The sample rate of the sensors is also displayed as it is
the feature with the most impact on the program’s structure due to the time restrictions it imposes.

Figure 4.17: Sensor Module Hardware Overview

4.2 Database
As with any sensor system, a countless amount of data is generated, especially considering the
accelerometers and ultrasonic microphones for which data must be captured at least 160 thousand
times each second, a way to store these measurements must be thoroughly considered. The Teensy
4.1 has support for a micro SD card 15 for which storage at the time of writing can reach up to
1.5 TB 16 , but this would not be a practical solution, since it would require an operator to retrieve
the data from the SD card manually. The most practical solution for dealing with data storage
management is to employ a database, analogous to how IoT devices store data. Using a database
would also enable interconnectivity between similar sensor modules or other hardware to run the
developed statistical and ML algorithms.
Time-series databases are a specialised category of databases designed specifically for effi-
ciently storing, managing, and analysing time-stamped data. Time series data, commonly gen-
erated by IoT devices, sensors, or monitoring systems, consists of measurements or events that
occur over time. Unlike traditional relational databases such as PostgreSQL, which can handle
15 https://www.sdcard.org/
16 https://www.micron.com/products/memory-cards
54 Methodology

time-series data to some extent, time-series databases are optimized for storing time-stamped data
that of high-frequency ingestion.
The most popular time-series databases, at the time of writing, are: InfluxDB, Prometheus and
DolphinDB [144]. The chosen one was, ultimately, InfluxDB due to being the most popular one,
meaning it is the one with the most documentation and readily available examples. The cloud-
based servers can be used for free, which alleviates the amount of data storage needed for the
present iteration of this project’s prototype. Several different web client APIs are also available,
including one for the Arduino Framework which uses either the ESP32 or ESP8266 WiFi chipsets
which can be used since the ESP8266 was the chosen WiFi connection hardware method chosen
for the prototype.
While MQTT is a widely used and efficient protocol for communication between multiple de-
vices in an IoT (Internet of Things) system, it may not be the best fit for a project that involves
only one device. MQTT is designed to facilitate communication between multiple devices, en-
abling efficient data exchange and real-time updates in a network of interconnected nodes. In a
scenario where there is only one device, using MQTT may introduce unnecessary complexity and
overhead. Alternative communication protocols, such as HTTP or simple serial communication,
may be more straightforward and suitable for a single-device project. InfluxDB has an Arduino
Client API for ESP32 and ESP8266 which allows for a simpler implementation route.

4.3 Custom Testbed

Before testing the final prototype in an industrial setting, a controlled environment is required to
fully test the prototype to ensure no unexpected behaviours occur in the final test. The process of
guaranteeing the prototype’s viability goes through a few steps that need to be checked:

• Sensors output correct measurements according to their datasheets;

• Measurements are sent to the database;

• Communication error handling;

• Prototype’s soldered connections do not break due to excessive vibration;

• Sensor position in relation to the target component;

The previous steps only refer to the prototype’s physical integrity and reliability, nonetheless,
using a testbed also brings another obvious advantage: being able to test the anomaly detection
algorithm with a controlled measurement noise level. Noise in this context refers to spurious
vibrations and sounds that, in an industrial setting, might be caused by several different sources.
These are not limited to environmental causes, e.g. other machines operating, but also by the
target machine, as is the case with structural resonances. This noise source should be taken into
account, as exemplified in [145], where a simulation model of a bearing was created, and structural
resonance is part of the equation that characterises the degradation process.
4.3 Custom Testbed 55

The first and most logical step for knowing which hardware to use in order to create a testbed
suitable for the dissertation’s objective is to compare testbeds in articles with similar goals of
condition monitoring and anomaly detection, which have already been explored in Section 3.7.
The testbeds previously explored in Section 3.7 are noticeably guided towards emulating heavy
machinery, however, the goal of this dissertation consists of detecting anomalies in an AGV and
accurately replicating the generic setup in the literature does not represent the critical components
of an AGV due to the differences in size and power of the elements, such as the motors and
shaft diameters. Most characteristics of the analysed testbeds can be adapted to fit the objective,
namely the structure elements of a motor and gearbox, a mechanical coupler, a shaft, a bearing
and a mechanical load. This structure can be used to represent the traction system of the AGVas it
comprises motors, gearboxes, shafts, bearings and wheels. Environmental noise is a big concern
in the ultimate scenario, and this factor should be taken into account while testing the prototype,
however, it is mostly ignored in the literature. In [82], a Bluetooth device is created, and sound
is used to identify faults in rolling bearings and to simulate real conditions, a speaker playing
prerecorded sounds from a similar environment as the planned deployment space is added to the
testing procedures to evaluate the prototype’s performance with and without extra noise. For this
dissertation, the same concept can be used to add noise to the overall system, however, parasitic
vibration is more difficult to simulate due to the need for costly equipment, such as a controllable
vibrating table.
To validate that the final sensor module is capable of detecting anomalous behaviour in the face
of noise and different scenarios, three different testbeds were built. In order to more accurately
simulate the AGV, components that are within the required scale were selected. The chosen basis
components were:

• A 12 V DC motor

• A mechanical decoupler connecting the motor shaft to the main shaft

• Two ball bearings for securing the shaft onto the platform

• An aluminium shaft

• A central ball bearing for testing purposes

• A wooden base to secure all the components

All testbeds use the same fixating ball bearing setup while varying the motor and main shaft
length, in Figures 4.18 to 4.20. The first testbed, in Figure 4.18, is comprised of a 1000 rpm DC
motor and a 40 cm shaft, the second testbed, in Figure 4.19, includes a 1000 rpm DC motor and
a 10 cm shaft and the third testbed, Figure 4.20, a 2640 rpm DC motor and a 10 cm shaft. The
black pieces visible in all testbeds were 3D printed and served as a support for the motor to rest
on, otherwise, the motor would hang and would only be supported by the decoupler.
56 Methodology

Figure 4.18: Testbed 1 without any sensors

Figure 4.19: Testbed 2 without any sensors

Figure 4.20: Testbed 3 with the SPH0645LM4H microphone visible

4.3 Custom Testbed 57

On the reviewed testbeds, the ones using accelerometers usually place them either on top of
the bearing support, on top of the motor or gearbox under testing or on the main shaft. However,
in this case, due to the physical dimensions of the AGV and, thus, of the testbed, it is not possible
to fixate the accelerometers on any of these locations. Since the wooden base is attached to the
testbed, the vibrations of the setup also propagate to the base, making it a viable option to place
the sensors. This is analogous to placing the sensors in the AGV since space inside is limited, and
the only feasible option is to put the sensors as close as possible to the critical components. The
placement of the temperature sensors and microphones are not as carefully placed as these can
take measurements without being physically connected to the object.

4.3.1 Tested Components

The most analysed component in the literature is the roller bearing, since it is a critical component
for rotating machinery, as explored in Section 3.1. Due to this fact, more information is available
pertaining to simulating or replicating faults in this type of component. This led to the choice
of replicating faults in a ball bearing in order to collect samples of anomalous behaviour of the
system. Figure 4.21 displays the bearing in normal operating conditions, the top picture shows the
bearing with the protective caps, which prevent debris from entering the inside structure and also
prevent loss of lubricant, while the bottom picture reveals the inside of the bearing.

Figure 4.21: Ball Bearing under testing in normal operating conditions

Figure 4.22 features the alterations done to the normal condition ball bearing: removal of most
of the lubricant and mechanical indentation. These methods are one of the most common for
creating defects on bearings [40].
58 Methodology

Figure 4.22: Ball Bearing under testing with lubricant partially removed

4.4 Software Structure

In this section, the software structure for all parts of this project will be detailed, from the data
collection portion, reviewing the algorithms used for the Teensy 4.1, passing through the dataset
creation, up to the ML models implementation.

4.4.1 Data Collection

To recapitulate from Section 4.1, the chosen sensors were the following:

• Accelerometers: ADXL355 and MMA8451

• Microphones: SPH0645LM4H and SPU0410LR5H-QB

• IR Thermometer: MLX90614

All the digital sensors except for the microphone possess libraries for the Arduino framework,
which greatly facilitates the software implementation since it minimises the direct interaction with
the SPI and I2C protocols as well as with direct register manipulation. Such low-level coding ap-
proaches are inherently error-prone and less human-readable when compared to higher-level code,
making the availability of Arduino libraries a significant advantage in ensuring robust software.
The digital microphone uses a different protocol, I2S. Thankfully, the Teensy platform provides a
library for interacting with I2S devices using Teensy microcontrollers, the Teensy Audio Library17 ,
so it is possible to abstract the final code from directly managing I2S operations.
Before utilising the ESP-12E for sending data wirelessly, it was decided that it was better
to record every sample and to save onto an SD card, made possible by an integrated SD card
reader available in the Teensy 4.1. This enables saving the raw samples and testing different pre-
processing techniques. It also removes the need to implement an application layer, redirecting the
effort on analysing the data.
The objective of the microcontroller’s program is to read the data from the sensors and to
store the measurements in the provided SD card. For convenience reasons, the data from each
sensor should be stored in a separate file or folder. For running the higher order operations of the
data collection algorithm, a finite state machine was developed, in Figure 4.23. Two push buttons
17 https://www.pjrc.com/teensy/td_libs_Audio.html
4.4 Software Structure 59

were connected to the microcontroller to interact with the program: one to record a single one-
second sample and a different one to start a loop of recording and saving samples. These buttons,
combined with the Serial port, enabled debugging each stage. The finite state machine in Figure
4.24 controls the looping of the program.

Figure 4.23: Main Finite State Machine - SD card version

Figure 4.24: Continuous Recording Finite State Machine

The finite state machines control the higher-order operation of the microcontroller, but the
state of recording and saving the data requires more detail since it is necessary to maintain a fixed
and rigorous sampling frequency for each sensor. Figure 4.25 shows a flowchart of the reading
stage. For reading from the digital I2C and SPI sensors, the following libraries were used:

• https://github.com/DFRobot/DFRobot_MLX90614

• https://github.com/adafruit/Adafruit_MMA8451_Library

• https://github.com/plasmapper/adxl355-arduino/

These libraries simplify the interaction with the sensors as the library’s user only needs to
call one or two functions to get the values. However, the libraries do not guarantee the necessary
60 Methodology

sampling rates, so an extra library was utilised to control the hardware timers available in the
Teensy 4.1: https://github.com/luni64/TeensyTimerTool.

Figure 4.25: Read sensor data Algorithm Flowchart

The final digital sensor, the SPH0645LM4H microphone, uses the I2S protocol. It requires a
specific hardware controller, which is present in the Teensy 4.1, and an integrated library can be
used: Audio.h 18 . The library uses the Teensy Audio System Design Tool 19 to simplify the setup
of different audio devices. The tool uses a block diagram to represent input sources, audio filters,
amplifier stages, audio synthesizers, etc., as the Teensy platform is often used for audio projects,
and turns the diagram into code, automatically. The blocks used for setting the I2S microphone as
an input device and to record the output to a buffer:

• i2s1 - I2S input

• filter1 - filter stage

• amp1 - amplifier stage

• queue1 - output

The code to record audio to an SD card was based on the example provided by the Audio.h
library, Recorder.ino. It records 512 bytes of audio before writing to an open file in the SD card
until 1 second has passed. The microphone outputs a binary file format, .raw, which has to be later
encoded into an audio format if playback is desired.
18 https://github.com/PaulStoffregen/Audio
19 https://www.pjrc.com/teensy/gui/index.html
4.4 Software Structure 61

Concerning the analog microphone, to achieve the needed sampling rate of twice the band-
width, according to the Nyquist Sampling Theorem, meaning a sampling frequency of 160 kHz,
the normal usage of the ADC with the ADC.h library 20 is not possible. The analogRead.ino exam-
ple that is provided, in the fastest setting, would take about 13 µs to take each sample, while a 160
kHz sampling frequency requires a sample every 6.25 µs, which makes it impossible to use the
normal ADC mode in this use case scenario. Thankfully, the library enables accessing the ADC
using Direct Memory Access (DMA), which can drastically decrease the time of retrieving the
samples. The final implementation of the analog sampling algorithm was based on the adc_timer
_dma.ino example, which continuously fills a buffer with the samples taken from the signal and
when the timer set frequency is met, an interrupt is made, and the values from the buffer are passed
to the memory of the microcontroller. Using DMA, each operation to retrieve the samples from
the buffer, which contained 1600 samples, takes less than 1 µs, enabling the required sampling
frequency to be achieved.
For saving the data into the SD card, the SD.h library in the Arduino Framework is used. It
utilises SPI for interfacing with the card and its storage. Figure 4.26 contains a flowchart show-
ing the algorithm employed to store the measurements (contained in buffers). Character strings
containing the file names are modified after each sample is taken to reflect the number of the
sample so that it is possible to join each sensor’s second worth of data together. Apart from the
SPH0645LM4H microphone, the readings are stored in .csv files making it easier to separate the
measurements of different axes, in the case of the accelerometers. Writing .csv files does not need
a different method from writing .txt files, as it is a text format type file and the only needed step is
to separate each axis with a comma and each instance of the buffer with a paragraph.

Figure 4.26: Save data in SD card Algorithm Flowchart

20 https://github.com/pedvide/ADC
62 Methodology

4.4.2 Dataset Creation

In the last Subsection, the methods required to collect the relevant data for the project’s objective
were reviewed, but steps remain before using the data to train ML models to detect anomalies.
This Subsection addresses these procedures, which include Preprocessing and the merging of the
different types of data, vibration, sound and temperature.

4.4.2.1 Preprocessing Steps

As discussed in Section 3.4, different types of data require different processes for extracting the
most relevant information from them. In this dissertation, particular emphasis has been placed on
vibration and sound data. This is due to these data types’ high variability in the context of me-
chanical maintenance. Temperature rises in much longer time frames and more consistently when
compared to vibration or sound. The chosen method is composed of two steps: the scaling of the
data and time domain feature calculation. The first step was taken from [50], where acceleration
data is passed through the RobustScaler function from sklearn21 . This scaler method removes
the median and scales the data according to the quantile range, which was chosen to be between
quantiles 5 and 95, as per the cited article. For the temperature, the StandardScaler 22 was used,
since this type of data is not prone to such high variance as vibration and sound. The second
step is described in [83], where central tendency-based time domain features are computed before
using them as the input for an ML model. The utilised characteristics are defined in the following
Equations 4.2 to 4.9, where X represents a sample.

! 12
1 N
RMS = ∑ [X(i)2 ]
N i=1
(4.2)

1 N
Variance(σ 2 ) = ∑ (X(i) − µ)2 (4.3)
N i=1

1 N X(i) − µ 3
 
Skewness = ∑ (4.4)
N i=1 σ

1 N X(i) − µ 4
 
Kurtosis = ∑ (4.5)
N i=1 σ

21 https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.RobustScaler.html
22 https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.StandardScaler.html
4.4 Software Structure 63

max(|X|)
CrestFactor = (4.6)
RMS

max(|X|)
ImpulseFactor = 1
(4.7)
N ∑Ni=1 |X(i)|

RMS
ShapeFactor = 1
(4.8)
N ∑Ni=1 |X(i)|

Range = max(X) − min(X) (4.9)

This method was chosen over more traditional steps, e.g. the FFT, since it is simpler to imple-
ment and these calculations could be easily transferred to the microcontroller, which could reduce
the amount of data to be sent wirelessly. Using the FFT implies selecting a certain number of
frequency bins, which can be a very large number when compared to the selected method. For
example, in [50], 2048 per 1-second sample is the chosen value, based on the approximate ac-
celerometer bandwidth of 2 kHz. Since, in this project, 3-axis accelerometers were chosen, this
number would be multiplied by three, while using the previous Equations, the total number of
values to be sent is 24 for each accelerometer. Another possible advantage lies in potentially being
less reliant on the bandwidth of the sensors, as sensors with lower bandwidth tend to be of lower
cost. Although in [83], these preprocessing calculations in exclusively in vibration data, due to the
similar nature of vibration data and sound data, which was previously explored, Equations 4.2 to
4.9 where also applied to the sound data.

4.4.2.2 Dataset Merging

To employ a single ML model with all the data types, merging the datasets is necessary to create the
training data. The primary challenge lies in that each sensor captures different sample quantities of
data for the same 1-second sample. For example, the ADXL355, according to its datasheet [120],
has a bandwidth of 1 kHz, meaning at least two thousand data points (due to the Nyquist Theorem),
are recorded, for three different axes, while only one temperature measurement is made by the
MLX90614. The method that was found to be the most suitable is the one used in [83], which is
previously detailed in Subsection 3.4.1. The strategy lies in using time-domain characteristics of
the vibration and sound signals, such as the variance and kurtosis. By utilising these features, it
is possible to describe the measured signal through variables that represent 1 second’s worth of
data. The selection of sample duration aligns with the methodology presented in [83], wherein
10-second samples are initially acquired and subsequently partitioned into 0.25-second segments.
64 Methodology

The decision to divide the samples by a factor of 40 is rooted on the fact that the motor used in the
testbed rotates the main shaft at 29 Hz, meaning each 0.25-second excerpt contains more than 7
revolutions of the shaft which was considered to be sufficient for computing trustworthy statistical
features. In the context of this dissertation, the exact rotation speed of the shaft is unknown, while
it is over 7 Hz, so 1 second is a suitable time frame to capture credible measurements within the
scenario.
Figure 4.27 shows some features and the structure of the final dataset. The suffix _x, _y or
_z represents the axis of the feature. The prefix 355 refers to the ADXL355 accelerometer, mma
to the MMA8451 accelerometer, digital_mic to the SPH0645LM4H microphone and object to the
object’s measured temperature using the MLX90614 IR thermometer.

Figure 4.27: Dataset Structure

4.4.3 ML Models

As previously described, the problem to be solved is Anomaly Detection, where the objective lies
in finding data that deviates from the majority. In this context, an anomaly constitutes a mechanical
fault that causes the system to behave differently from the normal state. This behaviour manifests
through different vibration patterns, which leads to an increase in temperature of the damaged
component. As the AGV comprises several critical components, it’s important to acknowledge that
the built testbed can only simulate a subset of potential anomalies. Consequently, it’s not feasible
to capture data representing every possible anomaly scenario. This limitation suggests that the
most suitable type of ML model to employ would be a Semi-Supervised model, which theoretically
possesses the capability to classify a new sample that deviates from the characteristics present in
the training data, labeling it as an anomaly. However, a significant portion of successful Semi-
Supervised Learning algorithms are rooted in Deep Learning, which typically demands substantial
datasets (comprising tens of thousands of data points) and can be more complex to implement. Due
to time constraints, the project focused exclusively on Supervised and Unsupervised ML models.
4.4 Software Structure 65

4.4.3.1 Supervised Models

In the selection of Supervised ML models, our choices were primarily guided by prior research
findings and their availability within the sklearn package or their ease of implementation. The
final set of models includes:

• k-Nearest Neighbours (k-NN) 23

• Support Vector Machine 24

• Random Forest 25

• AdaBoost 26

• XGBoost 27

It is worth noting that all models, except XGBoost, are readily available through sklearn. Some
models are not covered in Chapter 3 because no relevant articles were found focusing on them.

4.4.3.2 Unsupervised Models

Following the selection criteria for the Supervised ML models, the Unsupervised ML model set:

• One-Class SVM 28

• Isolation Forest 29

• Local Outlier Factor (LOF) 30

It is important to highlight that the selection of Unsupervised models is intentionally narrower

compared to the Supervised models. This decision stems from the belief that, in this use case, it
is crucial to have the capability to classify data as normal behaviour. While Unsupervised models
were still tested to account for unpredictable outcomes.

23 https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.KNeighborsClassifier.html
24 https://scikit-learn.org/stable/modules/generated/sklearn.svm.SVC.html
25 https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html
26 https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.AdaBoostClassifier.html
27 https://xgboost.readthedocs.io/en/stable/python/index.html
28 https://scikit-learn.org/stable/modules/generated/sklearn.svm.OneClassSVM.html
29 https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.IsolationForest.html
30 https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.LocalOutlierFactor.html
66 Methodology
Chapter 5

Experimental Results

In continuation of the processes outlined in the preceding chapter, the subsequent chapter out-
lines the techniques employed to assess the chosen models, while considering various scenarios,
including sampling restrictions.

5.1 Metrics

For evaluating the effectiveness of the ML models in detecting anomalies in the datasets, it is
necessary to choose adequate metrics that discern the most pertinent aspects of the model’s per-
formance for the associated scenario. Recapitulating, the ML problem is a binary classification
question: if the sample is of anomalous behaviour, then the anomaly feature is a 1; if it is normal
behaviour, the anomaly feature is a 0.
For classification problems, one of the most popular methods for evaluation of the model is
by using a confusion matrix. This N · N matrix, where N represents the number of target classes,
displays the actual target values with those predicted by the model. Figure 5.1 1 shows an example
of a confusion matrix: the top left corner is the case where the true label for the data point was
0 and the model correctly predicted that label; the top right corner represents the mislabelling
of the 0 class; the bottom left corner is the mislabelling of the 1 class; the bottom corner is the
correct labelling of the 1 class. From these four values, it is possible to extract the number of true
positives, false positives, true negatives and false negatives.
Other metrics have emerged from these four base numbers whose objective is to highlight
certain aspects of the model. In the context of this dissertation, the focus lies on minimising false
negatives, where an anomalous sample is incorrectly classified as normal, as this misclassification
could potentially lead to an unscheduled maintenance stop for the AGV and a halt of production,
which is less favourable than an unnecessary scheduled maintenance stop. Keeping this goal in
mind, the following metrics were employed:

• Recall
1 Taken from https://scikit-learn.org/stable/modules/generated/sklearn.metrics.ConfusionMatrixDisplay.html

67
68 Experimental Results

Figure 5.1: Confusion Matrix Example

• Precision

• F1-Score

• Accuracy

To underscore the importance of reducing false negatives, recall (as defined in Equation 5.1
2) was picked. A recall value closer to 1 signifies a higher rate of true positives relative to false
negatives, highlighting the model’s ability to correctly identify anomalies. Precision, defined in
Equation 5.2, highlights the fraction of relevant instances among the retrieved instances, meaning
a value closer to 1 is a better score than a value closer to 0. F1 Score is defined in Equation 5.3 and
takes into account Recall and Precision and can be understood as the harmonic mean of the two.
The F1 Score reaches its best value at 1. The final metric is accuracy and encompasses predicted
labels that exactly match the true labels, calculated using Equation 5.4.

TP
Recall = (5.1)
T P + FN

TP
Precision = (5.2)
T P + FP

Precision · Recall
F1 Score = 2 · (5.3)
Precision + Recall
2 TP - True Positives; TN - True Negatives; FP - False Positives; FN - False Negatives
5.2 Test Structure 69

TP+TN
Accuracy = (5.4)
T P + T N + FP + FN

5.2 Test Structure

In this section, the structure for the evaluation of the models will be overviewed. The goal of
producing a custom dataset was to validate the approach taken, concerning the chosen sensors,
pre-processing steps and ML models, while addressing a setting that is not frequently researched
in the field. The decision to build three different testbeds allows for comparing the same models
trained with different data and to control what anomalies are recorded, the environmental noise
and the amount of samples to obtain.

5.2.1 Dataset Unbalance

As previously stated, one of the challenges of Anomaly Detection is the potential lack of samples
of anomalous data due to anomalies representing rare events which can be difficult to capture. A
benefit of having a custom dataset is the ability to control the amount of anomalies present in the
training dataset in relation to the normal samples. Testing the models with different percentages
of anomalies in the dataset allows for simulating a scenario where anomalous behaviour is hard to
capture and comparing the model’s performance when being provided with an unbalanced dataset.
It is important to note that some of the selected models, namely Isolation Forest and Local Outlier
Factor expect that the dataset is unbalanced and one of the input parameters is the fraction of
anomalous data.
Different amounts of samples were recorded from each testbed:

• Testbed 1: 3018 samples of normal behaviour and 3058 samples of anomalous behaviour;

• Testbed 2: 4049 samples of normal behaviour and 4323 samples of anomalous behaviour;

• Testbed 3: 4161 samples of normal behaviour and 3966 samples of anomalous behaviour;

The disparity between the testbeds occurred due to initial software issues. These manifested
as the finite state machine becoming stuck in the recording function of the SPH0645LM4H I2S
microphone after approximately 100 samples had already been saved. The software was adjusted
to mitigate the issues, reducing the number of stops to restart the program, after extracting the
samples from the SD card.
For creating unbalanced datasets, the following percentages of anomalies were chosen: 50%,
43%, 33%, 20% and 9%. These values were selected a posteriori, following manually entered
amounts and an analysis of the impact on model performance using these thresholds. These per-
centages were subsequently employed to automate the model testing procedures, which will be
explored ahead.
70 Experimental Results

5.2.2 Sensor Comparison

One of the objectives of this dissertation is for the sensor module to be cost-effective, which is the
justification for integrating two different accelerometers in the testing procedures: the ADXL355
presents more adequate characteristics for the subsequent analysis, according to the research done
for the project, but the MMA8451 is, at the time of writing, more than six times cheaper. The
custom datasets can combine the readings from all the integrated sensors, combine certain sensors
or use the readings from a single sensor. This enables the comparison of multiple combinations
which adapt to the problem at hand: comparing the two accelerometers, comparing the most
expensive sensors with a junction of the cheapest and comparing the impact of each sensor on
the model’s performance, which was also taken into account in the automation of the testing
procedures.

5.2.3 ML Automation Script

Python was the language of choice for creating and testing the ML models. One advantage of using
it is the ability to create a list or array element containing any type of variable, including sklearn
objects and any other object from different packages. This enables iterating through models and
retraining them using different percentages of anomalies in the datasets while saving the relevant
metrics and images of each confusion matrix for each scenario.
Figure 5.2 displays the process, through a flowchart, used in the script for automating the
training and testing of each models while modifying the dataset provided with different numbers
of anomalies, as described in SubSection 5.2.1, and computing and saving the metrics discussed
in Section 5.1.
The script possesses two different parts, one for the Unsupervised models and another for
the Supervised ones, since these two kinds of models require slightly different procedures, as the
Unsupervised models do not require a training procedure. For using the script with different sensor
data, it is only required to alter the .csv file containing the data. All Supervised models were trained
and tested with a Train Test Ratio of 80/20, meaning 80% of the data was used for training the
models and 20% for testing. It is also important to note that it is assumed that the preprocessing
steps have already been performed and that the provided file contains the preprocessed data.

5.3 Ultrasonic Microphone Measurements

One of the primary challenges encountered during the development of the sensor module revolved
around the SPU0410LR5H-QB ultrasonic microphone, particularly in terms of amplifying its sig-
nal and converting it into the digital domain. Unfortunately, the measurements proved to be un-
usable, as evidenced by the complete saturation of the ADC’s input, as illustrated in Figure 5.3.
During testing, the input was not saturated, and the values were centred around 2048 (in line with
the offset of 1.65 V given to the output of the amplifier).
5.3 Ultrasonic Microphone Measurements 71

Figure 5.2: ML Automation Script Flowchart

Figure 5.3: Ultrasonic Microphone Readings - 0.5 second sample

72 Experimental Results

Figure 5.4 shows a more zoomed-in perspective of the sampled signal, where it is possible to
observe that not all values are either 0 or close to 4000 (the limits of the input range).

Figure 5.4: Zoomed In Ultrasonic Microphone Readings

Comparing the difference between the microphone’s testing setup and the sensor module’s
setup, the difference lies in the number of sensors connected to the power lines of the microcon-
troller, so the signal might be contaminated by noise in the power supply. The amplifier’s design
was based on the one in [131], but the one employed in this dissertation does not have an electrical
noise isolation stage nor an external voltage reference chip.
The observation of this error was almost immediate, and the ultrasonic microphone circuit
was removed from the module and its measurements were not included in the training of the ML
models.

5.4 Temperature Analysis

The data was sampled sequentially, starting from Testbed 1. Then, the data was analysed and
the Python scripts for preprocessing and for concatenating were made and tested on the data.
This sequence, as opposed to recording all the Testbeds and then make the scripts, was chosen
in order to optimise the process of acquiring the samples and to check what parts of the whole
process would take up the most time, building the final workflow. The easiest variable to process
both computationally and visually is temperature. Figure 5.5 displays the temperature readings
from Testbed 2 of the normal bearing, from sample 1450 up to sample 4030. In orange, the
temperature read from the IR thermometer and, in blue, the ambient temperature. A constant rise
in the temperature up to a steady state is noticeable and explained by the fact that recordings began
while the components were at ambient temperature.
5.4 Temperature Analysis 73

Figure 5.5: Temperature Readings from Testbed 2 - Sample 1450 to 4030

A different aspect is also noteworthy and emerges by comparing the difference between two
consecutive readings of the bearing’s temperature (Figure 5.6 shows a zoomed-in perspective of
the readings): the temperature fluctuates a relatively high amount within seconds apart, while the
ambient temperature’s fluctuation is much smaller. It can be deduced that the IR thermometer
was not measuring the same point on the bearing surface, probably due to being attached to the
wooden platform, which vibrated due to the components being tested. This error in the method
of measurement could possibly have been solved by using a different board for the sensor with a
smaller field of view since the one used from seeed studio has a field of view of 35º or by using a
more secure attachment method.

Figure 5.6: Zoomed In Temperature Readings from Testbed 2

An analysis perspective that contrasts the initial reasoning point for temperature as a measure-
ment unit to use to detect anomalies is that it is not useful for the objective. Temperature changes
74 Experimental Results

much more slowly than vibration or sound, and it would require for the AGV to continuously run
for its critical components’ temperature to increase significantly before a fault is detected, while
vibration and sound signals are always affected by damaged components. This is perhaps the
reason for the difference in the number of articles which use vibration or sound for PdM, while
temperature is not as often utilised.

The previous speculation is demonstrated in the results of the models when removing temper-
ature from the dataset. Table 5.1 compares metrics of 3 models trained (k-NN, SVM and Random
Forest) with all sensors and with all sensor data apart from temperature. In these examples, the
exclusion of temperature has a negligible impact on the results and the differences of less than 1%
could be attributed to the limited dataset size or the inherent randomness inherent in ML models.

Table 5.1: Comparison of models using all sensors and without temperature data - Testbed 1

Model Sensors Anomaly (%) Recall Precision F1-Score Accuracy

All 0.99 0.99 0.99 0.99
k-NN
No temperature 0.99 0.99 0.99 0.99
50
All 0.98 0.99 0.98 0.98
SVM
No temperature 0.98 0.98 0.98 0.98
All 1 1 0.99 0.99
RF 33
No temperature 1 1 1 1

Since adjustments were made between recordings and due to this lack of impact of temperature
in the results, temperature was not recorded for Testbed 3.

5.5 Unsupervised Models Results

In this Section, the metrics of the Unsupervised Learning Models will be analysed. To recapitulate,
three models were selected: One-Class SVM, Isolation Forest and Local Outlier Factor. In Table
5.2, the best results for all combinations are displayed, while excluding F1-Score. Many of the
models exhibited performance akin to that of the Testbed 1 version of the One-Class SVM, where
their accuracy was only marginally superior to that of random guessing. However, some models
(e.g. Testbed 2 Isolation Forest with 9% anomalies) possess relatively high accuracy and recall
values, but the precision values are very low. Figure 5.7 shows the confusion matrix of the previous
example and demonstrates the statement, as the model classifies almost every sample as normal,
which explains high accuracy and recall since only 9% of all samples to classify are anomalous.
This displays the benefit of using multiple metrics to evaluate a model’s performance.
5.6 Supervised Models Results 75

Figure 5.7: Isolation Forest Confusion Matrix - 9% anomalies

The poor performance of the Unsupervised models could potentially be attributed to the choice
of pre-processing steps employed on the data. To enhance their performance, alternative ap-
proaches could be explored, such as frequency-based or time-frequency-based methods, as dis-
cussed in Chapters 2 and 3.

5.6 Supervised Models Results

This section will go further in the exploration of the results of the Supervised models than in
Section 5.5 since the metrics are objectively superior. Table 5.3 details the best models in the most
challenging conditions, using all features except temperature: 9% of anomalies in the dataset. The
results are nearly ideal, while it is necessary to consider the small sample size, within the case of
Testbed 2, 109 anomalous samples, as shown in Figure 5.8.

Table 5.2: Unsupervised Models Metrics

Model Testbed Sensors Anomaly (%) Recall Precision Accuracy

1 All 0.59 0.59 0.59
OC SVM 2 All 50 0.46 0.45 0.46
3 No temperature 0.46 0.45 0.45
1 All 20 0.87 0.34 0.81
I-Forest 2 All 0.91 0.17 0.85
9
3 No temperature 0.90 0.060 0.83
1 All 0.91 0.073 0.87
LOF 2 All 9 0.91 0.14 0.85
3 No temperature 0.93 0.34 0.89
76 Experimental Results

Table 5.3: Best Results from Supervised Models

Model Testbed Sensors Anomaly (%) Recall Precision F1-Score Accuracy

1 No temperature 1 1 0.99 0.99
XGBoost 2 No temperature 9 1 1 0.99 0.99
3 No temperature 1 1 1 1

Figure 5.8: Testbed 2 - XGBoost Confusion Matrix - 9% anomalies

It is necessary to take into account that Table 5.3 shows the model with the best results, namely
XGBoost. Table 5.4 shows the performance of the other models under the same conditions. Us-
ing four different metrics proves to be a valuable approach, given that three of these metrics are
consistent across all models, with the exception of the F1-Score, which demonstrates markedly
inferior results in the SVM and Random Forest models. This multi-metric evaluation structure
allows for a more comprehensive assessment of each model and its respective performance.

5.6.1 Sensor Comparison

As emphasised throughout this dissertation, a key goal is to ensure the cost-effectiveness of the
module. Therefore, it becomes essential to evaluate the performance of various sensors while con-
sidering their cost. The ML Automation Script allows for simple testing of multiple alternatives:
using only the accelerometers or only the microphone or by combining different sensors. Table
5.5 presents a selection of models along with the corresponding training data. The models chosen
were the ones were the most variation occurred between evaluation metrics.
Table 5.5 shows that the better-performing models only present slight variations when chang-
ing the sensors used in the training data. However, the ones with previous inferior performance
5.6 Supervised Models Results 77

Table 5.4: More results for the same conditions

Model Testbed Sensors Anomaly (%) Recall Precision F1-Score Accuracy

1 1 1 0.97 0.99
k-NN 2 0.98 0.99 0.98 0.99
3 1 1 0.81 0.97
1 1 1 0.62 0.95
SVM 2 1 1 0.083 0.90
3 0.95 0.99 0.39 0.93
No temperature 9
1 1 1 0.97 0.99
RF 2 1 1 0.55 0.94
3 1 1 0.74 0.96
1 1 1 1 1
AdaBoost 2 0.99 0.95 0.95 0.99
3 1 1 1 1

display much higher variation, especially by only using either the MMA8451 accelerometer or the
SPH0645LM4H microphone, e.g. SVM’s accuracy goes down to 0.57 and 0.78, respectively. One
possible way to improve the model performance using these sensors is to combine the data from
both. Table 5.6 displays metrics for the worst performing models in Table 5.5 being retrained with
data from both sensors, while maintaining the conditions of the testbed and anomaly percentage.
Employing only the ADXL355 accelerometer for the training data produces marginally worse
performance for almost every scenario than using all sensor data. The combination of the cheaper
sensors (MMA8451 and SPH0645LM4H), which total in price of 13,85 C [128] [122] 3 , also
produces near-ideal results.

5.6.2 Unknown Anomalies

While the previous results have shown that some Supervised models have produced near-perfect
predictions in detecting anomalous behaviour based on 1-second samples. However, the recorded
anomalies do not represent every possible anomaly that might occur in a real scenario. Since three
different testbeds were used, it is possible to join datasets to simulate new anomalies by using
data from a different testbed than the one used for training to test the models and evaluate their
performance when facing anomalies that were not represented in the dataset. Table 5.7 presents the
best model to have met these conditions, it was trained using data from Testbed 3 and tested using
anomalies from Testbed 2. Figure 5.9 shows the confusion matrix of the same model. Automating
tests would provide more insight into this possibility of testing unseen anomalies, but the built
script is not capable of taking two different datasets.
The results obtained from this small experiment are promising but further exploration is re-
quired to figure out if this method of detecting faults is achievable in a real scenario, particularly
when confronted with unforeseen anomalies.
3 At the time of writing
78 Experimental Results

Table 5.5: Sensor Impact on model performance comparison

Model Testbed Sensors Anomaly (%) Recall Precision F1-Score Accuracy

No temperature 0.97 0.99 0.98 0.99
ADXL355 0.99 0.99 0.99 0.99
1
MMA8451 0.96 0.99 0.97 0.99
SPH0645LM4H 1 1 0.97 0.99
XGBoost 20
No temperature 1 1 0.99 0.99
ADXL355 1 1 0.99 0.99
2
MMA8451 0.93 0.98 0.93 0.97
SPH0645LM4H 1 1 0.97 0.99
No temperature 0.98 0.99 0.98 0.99
ADXL355 1 1 0.99 0.99
K-NN 2 33
MMA8451 0.87 0.76 0.82 0.87
SPH0645LM4H 0.89 0.95 0.77 0.86
No temperature 0.94 0.95 0.91 0.92
ADXL355 0.9 0.91 0.79 0.81
SVM 2 50
MMA8451 0.74 0.75 0.79 0.78
SPH0645LM4H 0.65 0.8 0.45 0.57
No temperature 1 1 0.98 0.99
ADXL355 1 1 0.99 0.99
RF 3 33
MMA8451 0.99 0.98 0.86 0.92
SPH0645LM4H 0.97 0.99 0.88 0.92
No temperature 1 1 0.99 0.99
ADXL355 1 1 1 1
AdaBoost 1 20
MMA8451 0.9 0.98 0.9 0.97
SPH0645LM4H 0.93 0.98 0.91 0.97

5.7 Summary
This Chapter delved into the methods and metrics employed to comprehensively evaluate the pre-
viously chosen ML models and the operating conditions. One of the main issues in Anomaly
Detection, the lack of anomalous data, was addressed by testing the models with different levels
of dataset unbalance. More in line with the context of this dissertation, the models were also
trained using different sensors to compare performance.
The Unsupervised models performed very poorly, possibly due to the preprocessing steps,
while the Supervised models obtained near-ideal results, even with datasets containing 91% of
normal data. Having data from three different testbeds allowed for experimenting with classifying
previously unseen anomalies. However, more research must be conducted before it is determined
the applicability of these tests in a real use case scenario.
5.7 Summary 79

Table 5.6: Model metrics combining MMA8451 and SPH0645LM4H

Model Sensors Anomaly (%) Recall Precision F1-Score Accuracy

SVM 50 0.74 0.72 0.79 0.79
RF MMA8451 and SPH0645LM4H 33 0.98 0.99 0.93 0.96
K-NN 33 0.87 0.93 0.87 0.91

Table 5.7: Best results testing with anomalies not present in the dataset

Model Sensors Anomaly (%) Recall Precision F1-Score Accuracy

RF No temperature 33 0.59 1 0.74 0.86

Figure 5.9: Confusion Matrix of model in Table 5.7

80 Experimental Results
Chapter 6

Conclusion and Future Work

6.1 Conclusion

The objective of this dissertation was to create a cost-effective sensor module for detecting anoma-
lies in an AGV and its surrounding machines. In order to validate the created system, a custom
platform was necessary to replicate the AGV, as the existing platforms and public datasets are
typically meant for studying industrial machinery, whose physical characteristics are not identical
to the AGV.
Three different custom testbeds were built, which allowed for diversifying the collected data
and further validating the system, including the hardware and software. The custom platforms can
also be used to simulate other types of machines, apart from AGVs, such as machines with smaller
traction systems.
A study of the most important features of the sensors employed in public datasets and in the
literature enabled a mindful choice of which sensors were the most adequate. This knowledge can
be utilised to streamline the sensor selection process, aligning them with a researcher’s specific
requirements.
The selected accelerometers and the digital microphones proved to be fit for the task at hand,
meeting physical space restrictions and providing reliable measurements, which enabled the ac-
curacy of the Supervised models. The IR thermometer and ultrasonic microphone require more
preparation and study to determine their possible contribution to the objective.
Regarding the preprocessing of the data and the results, while the preprocessing method em-
ployed yielded nearly ideal predictions for the Supervised ML models, the performance of the
Unsupervised ML models was notably poor. Therefore, it is advisable to explore alternative tech-
niques, starting from the ones described in Chapter 3. An experiment was made to simulate anoma-
lies not present in the training dataset in order to further test the Supervised models and to evaluate
their performance in a possible deployment scenario. Although, in this scenario, the models were
less accurate than in the conventional scenario, the results proved promising and further research
can validate the strategy. It is important to note that the number of collected samples is not suffi-
cient for experimenting with Deep Learning techniques, which are thoroughly reviewed and used

81
82 Conclusion and Future Work

in the literature.
Consideration for an industrial environment was taken throughout the document and thought
process, including the environmental noise inherent to a factory, the constricting space limitations
of the inside of an AGV and the inability to record every possible fault that might occur in the
system.
The ultimate goal of this dissertation was not met, which involved the detection of anoma-
lies in an AGV and in its surrounding machines. However, a strategy for constructing a sensor
module capable of achieving the aforementioned objective was delineated and validated, all while
considering the constraints inherent to the environmental and spatial conditions.

6.2 Future Work

The following section is meant to discuss ideas for improving the work developed in this disserta-
tion. As noted throughout the chapters, PdM is a complex field that encompasses a wide array of
research domains, offering ample opportunities for improvement in various aspects of the sensor
system, including both hardware and software.
While building the testing platforms, insufficient planning was made for attaching the sensors
to the mechanical system, even though it had been recognised that this aspect had a discernible
impact on the measurements. For further testing using the AGV itself, more preparation must be
done to ensure that the sensors are keeping fidelity in their readings and to try to isolate them
from noise as much as possible. Therefore, an enhancement would involve transitioning to either
a PCB or soldering the relevant connections. Additionally, using long cables for the sensors would
provide greater flexibility for positioning them in optimal locations inside the AGV.
Furthermore, if it is decided to build a more accurate testbed for experimenting with other
preprocessing techniques or to fine-tune the sensor module, it is advisable to select components
that align more closely with the AGV’s specifications, as the one selected for the custom testbeds
were only approximately matched.
Using multiple microphones to create a microphone array could potentially enhance the per-
formance of the Anomaly Detection algorithms since it makes it possible to pinpoint the faulty
sound signal. This concept has been applied in the field of PdM, namely in [62]. Considering the
limited space inside the AGV, this idea requires more exploration in order to name it viable in the
target use case scenario.
In this dissertation, calibration of the sensors was not taken into account and relied on the
manufacturer’s specifications for labelling the measurements as accurate. For future iterations,
considering calibration tests would be advisable due to the noise levels present in industrial envi-
ronments.
As previously discussed, the ideal model type for the problem at hand is Semi-Supervised
Learning due to a lack of data of all possible faults of the system. In order to obtain the needed
amount of data, changing the recording process would be beneficial since using the one employed
in this dissertation takes approximately two hours per 4000 samples captured. A possible approach
6.2 Future Work 83

is to adapt the algorithm used in [83], where vibration is sampled continuously for 10 seconds and
then the data is divided into segments, which would accelerate the recording process. One possible
limitation of using this method is the available memory of the microcontroller, which limits the
amount of data that can be stored at once before transferring either wirelessly or to external storage.
TinyML 1 is a field of study that brings ML to embedded devices. Implementing a TinyML
solution in the prototype would enable the usage of more complex ML models for detecting
anomalous behaviour without needing to send data wirelessly, apart from a warning signal that
an anomaly has been detected. Furthermore, it would reduce dependence on cloud infrastructure
for deploying complex models.

1 https://www.tinyml.org/
84 Conclusion and Future Work
References

[1] Gil-Yong Lee, Mincheol Kim, Ying-Jun Quan, Min-Sik Kim, Thomas Joon, Young Kim,
Hae-Sung Yoon, Sangkee Min, Dong-Hyeon Kim, Jeong-Wook Mun, Jin Woo Oh, In Gyu
Choi, Chung-Soo Kim, Won-Shik Chu, Jinkyu Yang, Binayak Bhandari, Choon-Man Lee,
Jeong-Beom Ihn, Sung-Hoon Ahn, and William E Boeing. Machine health management
in smart factory: A review †. Journal of Mechanical Science and Technology, 32(3):987–
1009, 2018. ISBN: 1220601802011. doi:10.1007/s12206-018-0201-1.

[2] Patrick Strauß, Markus Schmitz, René Wöstmann, and Jochen Deuse. Enabling of Predic-
tive Maintenance in the Brownfield through Low-Cost Sensors, an IIoT-Architecture and
Machine Learning. In Proceedings - 2018 IEEE International Conference on Big Data, Big
Data 2018, pages 1474–1483. Institute of Electrical and Electronics Engineers Inc., January
2019. doi:10.1109/BigData.2018.8622076.

[3] Tiago Zonta, Cristiano André da Costa, Rodrigo da Rosa Righi, Miromar José de Lima,
Eduardo Silveira da Trindade, and Guann Pyng Li. Predictive maintenance in the industry
4.0: A systematic literature review. Computers and Industrial Engineering, 150, 12 2020.
doi:10.1016/j.cie.2020.106889.

[4] Douglas Thomas and Brian Weiss. Maintenance costs and advanced maintenance tech-
niques in manufacturing machinery: Survey and analysis. International journal of prog-
nostics and health management, 12(1), 2021.

[5] Yuki Tagawa, Rytis Maskeliūnas, and Robertas Damaševičius. Acoustic anomaly detection
of mechanical failures in noisy real-life factory environments. Electronics (Switzerland),
10(19), October 2021. doi:10.3390/electronics10192329.

[6] Enrico Zio. Prognostics and Health Management (PHM): Where are we and where do we
(need to) go in theory and practice. Reliability Engineering & System Safety, 218:108119,
February 2022. doi:10.1016/j.ress.2021.108119.

[7] ISO Central Secretariat. Condition monitoring and diagnostics of machines — General
guidelines. Standard, International Organization for Standardization, Geneva, CH, 2018.

[8] Xiao-Sheng Si, Wenbin Wang, Chang-Hua Hu, and Dong-Hua Zhou. Remaining useful
life estimation–a review on the statistical data driven approaches. European journal of
operational research, 213(1):1–14, 2011.

[9] Dubravko Miljković. Fault detection methods: A literature survey. In 2011 Proceedings of
the 34th international convention MIPRO, pages 750–755. IEEE, 2011.

[10] Rolf Isermann. Fault-Diagnosis Systems. Springer Berlin Heidelberg, Berlin, Heidelberg,
2006. doi:10.1007/3-540-30368-5.

85
86 REFERENCES

[11] Rolf Isermann. Model-based fault-detection and diagnosis – status and applications. Annual
Reviews in Control, 29(1):71–85, January 2005. doi:10.1016/j.arcontrol.2004.
12.002.

[12] Dhiraj Neupane and Jongwon Seok. Bearing fault detection and diagnosis using case west-
ern reserve university dataset with deep learning approaches: A review. IEEE Access,
8:93155–93178, 2020.

[13] Purushottam Gangsar and Rajiv Tiwari. Signal based condition monitoring techniques for
fault detection and diagnosis of induction motors: A state-of-the-art review. Mechanical
systems and signal processing, 144:106908, 2020.

[14] Spilios D Fassois and John S Sakellariou. Time-series methods for fault detection and
identification in vibrating structures. Philosophical Transactions of the Royal Society A:
Mathematical, Physical and Engineering Sciences, 365(1851):411–448, 2007.

[15] Felix C Freiling, Rachid Guerraoui, and Petr Kuznetsov. The failure detector abstraction.
ACM Computing Surveys (CSUR), 43(2):1–40, 2011.

[16] Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomaly detection: A survey. ACM
computing surveys (CSUR), 41(3):1–58, 2009.

[17] Tahereh Pourhabibi, Kok-Leong Ong, Booi H Kam, and Yee Ling Boo. Fraud detection:
A systematic literature review of graph-based anomaly detection approaches. Decision
Support Systems, 133:113303, 2020.

[18] Mohiuddin Ahmed, Abdun Naser Mahmood, and Jiankun Hu. A survey of network
anomaly detection techniques. Journal of Network and Computer Applications, 60:19–31,
2016.

[19] Stefania Matteoli, Marco Diani, and Giovanni Corsini. A tutorial overview of anomaly
detection in hyperspectral images. IEEE Aerospace and Electronic Systems Magazine,
25(7):5–28, 2010.

[20] Ashik Mostafa Alvi, Siuly Siuly, and Hua Wang. Neurological abnormality detection from
electroencephalography data: a review. Artificial Intelligence Review, 55(3):2275–2312,
2022.

[21] Feifei Liu, Chengyu Liu, Lina Zhao, Xiangyu Zhang, Xiaoling Wu, Xiaoyan Xu, Yulin Liu,
Caiyun Ma, Shoushui Wei, Zhiqiang He, et al. An open access database for evaluating the
algorithms of electrocardiogram rhythm and morphology abnormality detection. Journal of
Medical Imaging and Health Informatics, 8(7):1368–1373, 2018.

[22] R. Keith Mobley. An introduction to predictive maintenance. Butterworth-Heinemann,

2002.

[23] Thyago P. Carvalho, Fabrízzio A.A.M.N. Soares, Roberto Vita, Roberto da P. Francisco,
João P. Basto, and Symone G.S. Alcalá. A systematic literature review of machine learning
methods applied to predictive maintenance. Computers and Industrial Engineering, 137,
2019. doi:10.1016/j.cie.2019.106024.

[24] Yongyi Ran, Xin Zhou, Pengfeng Lin, Yonggang Wen, and Ruilong Deng. A survey of
predictive maintenance: Systems, purposes and approaches, 2019. arXiv:1912.07383.
REFERENCES 87

[25] Sule Selcuk. Predictive maintenance, its implementation and latest trends. Proceedings
of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture,
231:1670–1679, 2017. doi:10.1177/0954405415601640.

[26] Alan V. Oppenheim and Ronald W. Schafer. Discrete-time signal processing. Pearson
Education India, 1999.

[27] Sophocles J Orfanidis. Introduction to signal processing. Prentice-Hall, Inc., 1995.

[28] Donald E Knuth. Big omicron and big omega and big theta. ACM Sigact News, 8(2):18–24,
1976.

[29] E. Oran Brigham. The fast Fourier transform and its applications. Prentice-Hall signal
processing series. Prentice Hall, Englewood Cliffs, N.J, 1988.

[30] A. Graps. An introduction to wavelets. IEEE Computational Science and Engineering,

2:50–61, 1995. URL: http://ieeexplore.ieee.org/document/388960/, doi:
10.1109/99.388960.

[31] Yonggang Xu, Yunjie Deng, Jiyuan Zhao, Weikang Tian, and Chaoyong Ma. A novel
rolling bearing fault diagnosis method based on empirical wavelet transform and spectral
trend. IEEE Transactions on Instrumentation and Measurement, 69(6):2891–2904, 2019.

[32] Norden E. Huang, Zheng Shen, Steven R. Long, Manli C. Wu, Hsing H. Shih, Quanan
Zheng, Nai-Chyuan Yen, Chi Chao Tung, and Henry H. Liu. The empirical mode decom-
position and the Hilbert spectrum for nonlinear and non-stationary time series analysis.
Proceedings of the Royal Society of London. Series A: Mathematical, Physical and En-
gineering Sciences, 454, March 1998. URL: https://royalsocietypublishing.
org/doi/10.1098/rspa.1998.0193, doi:10.1098/rspa.1998.0193.

[33] Practical Introduction to Multiresolution Analysis - MATLAB & Simulink Example. Ac-
cessed on 08-06-2023. URL: https://www.mathworks.com/help/wavelet/ug/
practical-introduction-to-multiresolution-analysis.html.

[34] Michael Feldman. Hilbert transform in vibration analysis. Mechanical systems and signal
processing, 25(3):735–802, 2011.

[35] Norden Eh Huang. Hilbert-Huang transform and its applications, volume 16. World Sci-
entific, 2014.

[36] Gang Cheng, Yu-long Cheng, Li-hua Shen, Jin-bo Qiu, and Shuai Zhang. Gear fault
identification based on hilbert–huang transform and som neural network. Measurement,
46(3):1137–1146, 2013.

[37] Jay Lee, Fangji Wu, Wenyu Zhao, Masoud Ghaffari, Linxia Liao, and David Siegel. Prog-
nostics and health management design for rotary machinery systems—Reviews, method-
ology and applications. Mechanical Systems and Signal Processing, 42:314–334, 2014.
doi:10.1016/j.ymssp.2013.06.004.

[38] U Bhardwaj, AP Teixeira, and C Guedes Soares. Reliability prediction of an offshore wind
turbine gearbox. Renewable Energy, 141:693–706, 2019.
88 REFERENCES

[39] Gustavo Scalabrini Sampaio, Arnaldo Rabello de Aguiar Vallim Filho, Leilton Santos da
Silva, and Leandro Augusto da Silva. Prediction of motor failure time using an artificial
neural network. Sensors, 19(19):4342, 2019.

[40] N Tandon and A Choudhury. A review of vibration and acoustic measurement methods for
the detection of defects in rolling element bearings. Tribology International, 32(8):469–
480, August 1999. doi:10.1016/S0301-679X(99)00077-8.

[41] Ken Mao, DG Chetwynd, and M Millson. A new method for testing polymer gear wear rate
and performance. Polymer Testing, 82:106323, 2020.

[42] Anurag Choudhary, Deepam Goyal, Sudha Letha Shimi, and Aparna Akula. Condition
monitoring and fault diagnosis of induction motors: A review. Archives of Computational
Methods in Engineering, 26:1221–1238, 2019.

[43] Yaguo Lei, Jing Lin, Ming J Zuo, and Zhengjia He. Condition monitoring and fault diag-
nosis of planetary gearboxes: A review. Measurement, 48:292–305, 2014.

[44] R.B. Randall. Vibration-based Condition Monitoring: Industrial, Aerospace and Auto-
motive Applications. Wiley, 2011. URL: https://books.google.pt/books?id=
2v4RDQEACAAJ.

[45] Saeed Shokrollahi, Farhad Adel, and Hamid Ahmadian. An investigation into the ac-
celerometer mounting effects on signal transmissibility in modal measurements. Scientia
Iranica, 24(5):2436–2444, 2017.

[46] Case western reserve university bearing data center 2021, Aug 2021. Accessed on 23-04-
2023. URL: https://engineering.case.edu/bearingdatacenter.

[47] Dhiraj Neupane and Jongwon Seok. Bearing Fault Detection and Diagnosis Using Case
Western Reserve University Dataset With Deep Learning Approaches: A Review. IEEE
Access, 2020. doi:10.1109/ACCESS.2020.2990528.

[48] Paderborn university bearing dataset, Aug 2021. Accessed on 23-04-2023. URL:
https://mb.uni-paderborn.de/en/kat/main-research/datacenter/
bearing-datacenter/data-sets-and-download.

[49] Patrick Nectoux, Rafael Gouriveau, Kamal Medjaher, Emmanuel Ramasso, Brigitte
Chebel-Morello, Noureddine Zerhouni, and Christophe Varnier. Pronostia: An experimen-
tal platform for bearings accelerated degradation tests. In IEEE International Conference
on Prognostics and Health Management, PHM’12., pages 1–8. IEEE Catalog Number:
CPF12PHM-CDR, 2012.

[50] Oliver Mey, Willi Neudeck, Andre Schneider, and Olaf Enge-Rosenblatt. Machine
Learning-Based Unbalance Detection of a Rotating Shaft Using Vibration Data. In 2020
25th IEEE International Conference on Emerging Technologies and Factory Automation
(ETFA), pages 1610–1617, Vienna, Austria, September 2020. IEEE. doi:10.1109/
ETFA46521.2020.9212000.

[51] Kahiomba Sonia Kiangala and Zenghui Wang. Initiating predictive maintenance for a
conveyor motor in a bottling plant using industry 4.0 concepts. The International Jour-
nal of Advanced Manufacturing Technology, 97(9-12):3251–3271, August 2018. doi:
10.1007/s00170-018-2093-8.
REFERENCES 89

[52] N Ahobal et al. Study of vibration characteristics of unbalanced overhanging rotor. In IOP
Conference Series: Materials Science and Engineering, volume 577, page 012140. IOP
Publishing, 2019.
[53] Hongmei Liu, Lianfeng Li, and Jian Ma. Rolling Bearing Fault Diagnosis Based on STFT-
Deep Learning and Sound Signals. Shock and Vibration, 2016:1–12, 2016. doi:10.
1155/2016/6127479.

[54] Yunus Emre Karabacak, Nurhan Gürsel Özmen, and Levent Gümüşel. Intelligent worm
gearbox fault diagnosis under various working conditions using vibration, sound and
thermal features. Applied Acoustics, 186:108463, January 2022. doi:10.1016/j.
apacoust.2021.108463.

[55] Meng Ma, Chuang Sun, and Xuefeng Chen. Deep Coupling Autoencoder for Fault Di-
agnosis With Multimodal Sensory Data. IEEE Transactions on Industrial Informatics,
14(3):1137–1145, March 2018. doi:10.1109/TII.2018.2793246.
[56] Eduardo C. Nunes. Anomalous sound detection with machine learning: A systematic re-
view, 2021. arXiv:2102.07820.
[57] Yuma Koizumi, Shoichiro Saito, Hisashi Uematsu, Noboru Harada, and Keisuke Imoto.
Toyadmos: A dataset of miniature-machine operating sounds for anomalous sound detec-
tion, 2019. arXiv:1908.03299.
[58] Harsh Purohit, Ryo Tanabe, Kenji Ichige, Takashi Endo, Yuki Nikaido, Kaori Suefusa, and
Yohei Kawaguchi. Mimii dataset: Sound dataset for malfunctioning industrial machine
investigation and inspection, 2019. arXiv:1909.09347.
[59] Noboru Harada, Daisuke Niizumi, Daiki Takeuchi, Yasunori Ohishi, Masahiro Yasuda, and
Shoichiro Saito. Toyadmos2: Another dataset of miniature-machine operating sounds for
anomalous sound detection under domain shift conditions, 2021. arXiv:2106.02369.
[60] Yaoguang Wang, Yaohao Zheng, Yunxiang Zhang, Yongsheng Xie, Sen Xu, Ying Hu, and
Liang He. Unsupervised anomalous sound detection for machine condition monitoring
using classification-based methods. Applied Sciences, 11(23):11128, 2021.
[61] Paul Primus, Verena Haunschmid, Patrick Praher, and Gerhard Widmer. Anomalous sound
detection as a simple binary classification problem with careful selection of proxy outlier
examples. arXiv preprint arXiv:2011.02949, 2020.
[62] Adam Bondyra, Marek Kołodziejczak, Radosław Kulikowski, and Wojciech Giernacki. An
Acoustic Fault Detection and Isolation System for Multirotor UAV. Energies, 15(11), June
2022. doi:10.3390/en15113955.
[63] CK Madhusudana, Hemantha Kumar, and S Narendranath. Face milling tool condition
monitoring using sound signal. International Journal of System Assurance Engineering
and Management, 8:1643–1653, 2017.
[64] Yicheng Yu, Ali Safari, Xudong Niu, Bruce Drinkwater, and Kirill V Horoshenkov. Acous-
tic and ultrasonic techniques for defect detection and condition monitoring in water and
sewerage pipes: A review. Applied Acoustics, 183:108282, 2021.
[65] Suvi Santa-aho, Antti Nurmikolu, and Minnamari Vippola. Automated ultrasound-based
inspection of rails. International Journal of Railway, 10(2):21–29, 2017.
90 REFERENCES

[66] Arash Behnia, Hwa Kian Chai, and Tomoki Shiotani. Advanced structural health monitor-
ing of concrete structures with the aid of acoustic emission. Construction and Building Ma-
terials, 65:282–302, August 2014. doi:10.1016/j.conbuildmat.2014.04.103.

[67] ThomasD. Rossing, editor. Springer Handbook of Acoustics. Springer New York, New
York, NY, 2007. doi:10.1007/978-0-387-30425-0.

[68] C.U. Grosse, M. Ohtsu, D.G. Aggelis, and T. Shiotani. Acoustic Emission Testing: Ba-
sics for Research – Applications in Engineering. Springer Tracts in Civil Engineering.
Springer International Publishing, 2021. URL: https://books.google.pt/books?
id=EYk4EAAAQBAJ.

[69] Tomoki Shiotani. Evaluation of repair effect for deteriorated concrete piers of intake
dam using ae activity. In Acoustic Emission Testing, volume 13 of Advanced Materials
Research, pages 175–180. Trans Tech Publications Ltd, 9 2006. doi:10.4028/www.
scientific.net/AMR.13-14.175.

[70] Christian U. Grosse and Florian Finck. Quantitative evaluation of fracture processes in con-
crete using signal-based acoustic emission techniques. Cement and Concrete Composites,
28(4):330–336, April 2006. doi:10.1016/j.cemconcomp.2006.02.006.

[71] Manfred Joswig. Pattern recognition for earthquake detection. Bulletin of the Seismological
Society of America, 80(1):170–186, 1990.

[72] Wahyu Caesarendra, Buyung Kosasih, Anh Kiet Tieu, Hongtao Zhu, Craig A.S. Moodie,
and Qiang Zhu. Acoustic emission-based condition monitoring methods: Review and appli-
cation for low speed slew bearing. Mechanical Systems and Signal Processing, 72-73:134–
159, May 2016. doi:10.1016/j.ymssp.2015.10.020.

[73] Rende Liu. Condition monitoring of low-speed and heavily loaded rolling element bearing.
Industrial Lubrication and Tribology, 59(6):297–300, 2007.

[74] Davide Crivelli, Simon Hutt, Alastair Clarke, Pietro Borghesani, Zhongxiao Peng, and
Robert Randall. Condition monitoring of rotating machinery with acoustic emission: A
british–australian collaboration. In Asset Intelligence through Integration and Interoper-
ability and Contemporary Vibration Engineering Technologies: Proceedings of the 12th
World Congress on Engineering Asset Management and the 13th International Conference
on Vibration Engineering and Technology of Machinery, pages 119–128. Springer, 2019.

[75] Enrique Caso, Alfonso Fernandez-del Rincon, Pablo Garcia, Miguel Iglesias, and Fernando
Viadero. Monitoring of misalignment in low speed geared shafts with acoustic emission
sensors. Applied Acoustics, 159:107092, 2020.

[76] S. Bagavathiappan, B.B. Lahiri, T. Saravanan, John Philip, and T. Jayakumar. Infrared
thermography for condition monitoring – A review. Infrared Physics & Technology, 60:35–
55, September 2013. doi:10.1016/j.infrared.2013.03.006.

[77] A.W. Kandeal, M.R. Elkadeem, Amrit Kumar Thakur, Gamal B. Abdelaziz, Ravishankar
Sathyamurthy, A.E. Kabeel, Nuo Yang, and Swellam W. Sharshir. Infrared thermography-
based condition monitoring of solar photovoltaic systems: A mini review of recent ad-
vances. Solar Energy, 223:33–43, July 2021. doi:10.1016/j.solener.2021.05.
032.
REFERENCES 91

[78] Olivier Janssens, Raiko Schulz, Viktor Slavkovikj, Kurt Stockman, Mia Loccufier, Rik
Van De Walle, and Sofie Van Hoecke. Thermal image based fault diagnosis for ro-
tating machinery. Infrared Physics & Technology, 73:78–87, November 2015. doi:
10.1016/j.infrared.2015.09.004.

[79] Anam Abid, Muhammad Tahir Khan, and Javaid Iqbal. A review on fault detection and
diagnosis techniques: basics and beyond. Artificial Intelligence Review, 54:3639–3664,
June 2021. doi:10.1007/s10462-020-09934-2.

[80] Reemon Z Haddad and Elias G Strangas. On the accuracy of fault detection and separation
in permanent magnet synchronous machines using mcsa/mvsa and lda. IEEE Transactions
on Energy Conversion, 31(3):924–934, 2016.

[81] Ali Bou Nassif, Manar Abu Talib, Qassim Nasir, and Fatima Mohamad Dakalbab. Ma-
chine Learning for Anomaly Detection: A Systematic Review. IEEE Access, 9:78658–
78700, 2021. URL: https://ieeexplore.ieee.org/document/9439459/, doi:
10.1109/ACCESS.2021.3083060.

[82] Erica Raviola and Franco Fiori. A Low-Cost, Small-Size, and Bluetooth-Connected Module
to Detect Faults in Rolling Bearings. Applied Sciences, 10(16):5645, August 2020. doi:
10.3390/app10165645.

[83] Muhammad Masood Tahir, Abdul Qayyum Khan, Naeem Iqbal, Ayyaz Hussain,
and Saeed Badshah. Enhancing Fault Classification Accuracy of Ball Bearing
Using Central Tendency Based Time Domain Features. IEEE Access, 5:72–83,
2017. URL: http://ieeexplore.ieee.org/document/7572105/, doi:10.
1109/ACCESS.2016.2608505.

[84] K Balachandar and R Jegadeeshwaran. Friction stir welding tool condition monitoring using
vibration signals and random forest algorithm–a machine learning approach. Materials
Today: Proceedings, 46:1174–1180, 2021.

[85] Candice Bentéjac, Anna Csörgő, and Gonzalo Martínez-Muñoz. A comparative analy-
sis of gradient boosting algorithms. Artificial Intelligence Review, 54:1937–1967, March
2021. URL: https://link.springer.com/10.1007/s10462-020-09896-5,
doi:10.1007/s10462-020-09896-5.

[86] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings
of the 22nd acm sigkdd international conference on knowledge discovery and data mining,
pages 785–794, 2016.

[87] Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye,
and Tie-Yan Liu. Lightgbm: A highly efficient gradient boosting decision tree. Advances
in neural information processing systems, 30, 2017.

[88] Anna Veronika Dorogush, Vasily Ershov, and Andrey Gulin. Catboost: gradient boosting
with categorical features support. arXiv preprint arXiv:1810.11363, 2018.

[89] Dahai Zhang, Liyang Qian, Baijin Mao, Can Huang, Bin Huang, and Yulin Si. A data-
driven design for fault detection of wind turbines using random forests and xgboost. Ieee
Access, 6:21020–21031, 2018.
92 REFERENCES

[90] Zeki Murat Çınar, Abubakar Abdussalam Nuhu, Qasim Zeeshan, Orhan Korhan, Mo-
hammed Asmael, and Babak Safaei. Machine Learning in Predictive Maintenance towards
Sustainable Smart Manufacturing in Industry 4.0. Sustainability, 12:8211, October 2020.
doi:10.3390/su12198211.

[91] Lu Shuang and Li Meng. Bearing fault diagnosis based on pca and svm. In 2007 Interna-
tional conference on mechatronics and automation, pages 3503–3507. IEEE, 2007.

[92] Miryam Elizabeth Villa-Pérez, Miguel A Alvarez-Carmona, Octavio Loyola-Gonzalez,

Miguel Angel Medina-Pérez, Juan Carlos Velazco-Rossell, and Kim-Kwang Raymond
Choo. Semi-supervised anomaly detection algorithms: A comparative summary and fu-
ture research directions. Knowledge-Based Systems, 218:106878, 2021.

[93] Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation Forest. In 2008 Eighth IEEE
International Conference on Data Mining, pages 413–422. IEEE, December 2008. doi:
10.1109/ICDM.2008.17.

[94] Hongfei Wang, Wen Jiang, Xinyang Deng, and Jie Geng. A new method for fault detection
of aero-engine based on isolation forest. Measurement, November 2021. doi:10.1016/
j.measurement.2021.110064.

[95] Shen Yin, Xiangping Zhu, and Chen Jing. Fault detection based on a robust one class
support vector machine. Neurocomputing, 145:263–268, 2014.

[96] Xinya Wu, Yan Zhang, Changming Cheng, and Zhike Peng. A hybrid classification autoen-
coder for semi-supervised fault diagnosis in rotating machinery. Mechanical Systems and
Signal Processing, 149:107327, 2021.

[97] Arden Dertat. Applied Deep Learning - Part 3: Autoencoders,

October 2017. URL:
https://towardsdatascience.com/
applied-deep-learning-part-3-autoencoders-1c083af4d798.

[98] Sebastian Thrun, Wolfram Burgard, and Dieter Fox. Probabilistic robotics. MIT Press,
2005.

[99] Moslem Azamfar, Jaskaran Singh, Inaki Bravo-Imaz, and Jay Lee. Multisensor data fusion
for gearbox fault diagnosis using 2-D convolutional neural network and motor current sig-
nature analysis. Mechanical Systems and Signal Processing, 144:106861, October 2020.
doi:10.1016/j.ymssp.2020.106861.

[100] HT Shi and XT Bai. Model-based uneven loading condition monitoring of full ceramic ball
bearings in starved lubrication. Mechanical Systems and Signal Processing, 139:106583,
2020.

[101] Tai Guo and Zhongmin Deng. An improved emd method based on the multi-objective opti-
mization and its application to fault feature extraction of rolling bearing. Applied Acoustics,
127:46–62, 2017.

[102] Dong Liu, Hongtao Zeng, Zhihuai Xiao, Lihong Peng, and OP Malik. Fault diagnosis of
rotor using emd thresholding-based de-noising combined with probabilistic neural network.
Journal of Vibroengineering, 19(8):5920–5931, 2017.
REFERENCES 93

[103] Roberto M Souza, Erick GS Nascimento, Ubatan A Miranda, Wenisten JD Silva, and Her-
man A Lepikson. Deep learning for diagnosis and classification of faults in industrial rotat-
ing machinery. Computers & Industrial Engineering, 153:107060, 2021.

[104] Yaqiong Lv, Wenqin Zhao, Zhiyao Zhao, Weidong Li, and Kam K.H. Ng. Vibration signal-
based early fault prognosis: Status quo and applications. Advanced Engineering Informat-
ics, 52:101609, April 2022. doi:10.1016/j.aei.2022.101609.

[105] Jacek Dybała and Radosław Zimroz. Rolling bearing diagnosing method based on empirical
mode decomposition of machine vibration signal. Applied Acoustics, 77:195–203, 2014.

[106] Saeed Nezamivand Chegini, Ahmad Bagheri, and Farid Najafi. Application of a new ewt-
based denoising technique in bearing fault diagnosis. Measurement, 144:275–297, 2019.

[107] Yaguo Lei and Ming J Zuo. Fault diagnosis of rotating machinery using an improved hht
based on eemd and sensitive imfs. Measurement Science and Technology, 20(12):125701,
2009.

[108] Tao Liu and Dongxin Lu. The application and development of iot. In 2012 International
symposium on information technologies in medicine and education, volume 2, pages 991–
994. IEEE, 2012.

[109] Mustapha Reda Senouci and Abdelhamid Mellouk. Deploying wireless sensor networks:
theory and practice. Elsevier, 2016.

[110] Sakina Elhadi, Abdelaziz Marzak, Nawal Sael, and Soukaina Merzouk. Comparative Study
of IoT Protocols. SSRN Electronic Journal, 2018. doi:10.2139/ssrn.3186315.

[111] Jie Ding, Mahyar Nemati, Chathurika Ranaweera, and Jinho Choi. IoT Connectivity
Technologies and Applications: A Survey. IEEE Access, 8:67646–67673, 2020. doi:
10.1109/ACCESS.2020.2985932.

[112] Nurzaman Ahmed, Hafizur Rahman, and Md I Hussain. A comparison of 802.11 ah and
802.15. 4 for iot. Ict Express, 2(3):100–102, 2016.

[113] Qahhar Muhammad Qadir, Tarik A. Rashid, Nawzad K. Al-Salihi, Birzo Ismael, Alexan-
der A. Kist, and Zhongwei Zhang. Low Power Wide Area Networks: A Survey of Enabling
Technologies, Applications and Interoperability Needs. IEEE Access, 6:77454–77473,
2018. doi:10.1109/ACCESS.2018.2883151.

[114] Jin-Shyan Lee, Yu-Wei Su, and Chung-Chou Shen. A comparative study of wireless pro-
tocols: Bluetooth, uwb, zigbee, and wi-fi. In IECON 2007-33rd Annual Conference of the
IEEE Industrial Electronics Society, pages 46–51. Ieee, 2007.

[115] Accessed on 02-07-2023. URL: https://spectraquest.com/simulators/

details/mfs/.

[116] S. A. McInerny and Y. Dai. Basic vibration signal processing for bearing fault detection.
IEEE Transactions on Education, 46(1):149–156, February 2003. doi:10.1109/TE.
2002.808234.

[117] Alan S. Morris and Reza Langari. Translational motion, vibration, and shock measurement.
In Measurement and Instrumentation, pages 569–602. Elsevier, 2021. doi:10.1016/
B978-0-12-817141-7.00019-0.
94 REFERENCES

[118] Martin Doseděl and Zdeněk Havránek. Design and performance evaluation of smart vi-
bration sensor for industrial applications with built-in mems accelerometers. In 2018 18th
International Conference on Mechatronics-Mechatronika (ME), pages 1–8. IEEE, 2018.

[119] Matej Andrejašic. Mems accelerometers. In University of Ljubljana. Faculty for mathe-
matics and physics, Department of physics, Seminar, volume 49, 2008.

[120] Analog Devices. ADXL354/ADXL355 (Rev. B), 2020.

[121] NXP Semiconductors. MMA8451Q, 3-axis, 14-bit/8-bit digital accelerometer, 2017.

[122] 2019 Adafruit | Mouser. URL: https://pt.mouser.com/ProductDetail/

485-2019.

[123] EVAL-ADXL355Z Analog Devices | Mouser. URL: https://pt.mouser.com/

ProductDetail/584-EVAL-ADXL355Z.

[124] Siti Aisyah Zawawi, Azrul Azlan Hamzah, Burhanuddin Yeop Majlis, and Faisal Mohd-
Yasin. A review of mems capacitive microphones. Micromachines, 11(5):484, 2020.

[125] Homin Song, Jongwoong Park, and John S Popovics. Development of an MEMS ul-
trasonic microphone array system and its application to compressed wavefield imaging
of concrete. Smart Materials and Structures, 29(10):105011, October 2020. doi:
10.1088/1361-665X/ababe5.

[126] GRAS 40PH-10. URL: https://www.ni.com/pt-pt/support/model.

gras-40ph-10.html.

[127] knowles. SPH0645LM4H-1 Rev A Datasheet, 2017.

[128] 3421 Adafruit | Mouser. URL: https://pt.mouser.com/ProductDetail/

485-3421.

[129] Knowles Electronics. Zero-Height SiSonic TM Microphone, 2013.

[130] Ultrasonic Mic Board. URL:

https://www.veldshop.nl/en/
ultrasonic-mic-board.html?source=facebook.

[131] Andrew P. Hill, Peter Prince, Jake L. Snaddon, C. Patrick Doncaster, and Alex Rogers.
AudioMoth: A low-cost acoustic device for monitoring biodiversity and the environment.
HardwareX, 6:e00073, October 2019. doi:10.1016/j.ohx.2019.e00073.

[132] Texas Instruments. OPAx350 High-Speed, Single-Supply, Rail-to-Rail Operational Ampli-

fiers MicroAmplifier Series, 2015.

[133] TDK Corporation. Singing Capacitors (Piezoelectric Effect), 2006.

[134] Melexis. MLX90614 family Single and Dual Zone Infra Red Thermometer in TO-39 Fea-
tures and Benefits, 2009. ISBN: 3901090614.

[135] 101020895 Seeed Studio

Mouser.| URL: https://pt.mouser.com/
ProductDetail/713-101020895.

[136] NXP Semiconductors. i.MX RT1060 Crossover Processors for Consumer Products.
REFERENCES 95

[137] NXP Semiconductors. I2c-bus specification and user manual. NXP Semiconductors, 4,
2021.

[138] BOB-12009 SparkFun | Mouser. URL: https://pt.mouser.com/ProductDetail/

474-BOB-12009.

[139] Matteo Frigo and Steven G. Johnson. The design and implementation of FFTW3. Proceed-
ings of the IEEE, 93(2):216–231, 2005. doi:10.1109/JPROC.2004.840301.

[140] Espressif Systems. ESP8266 Datasheet.

[141] Ai Thinker. ESP12E-Datasheet.pdf.

[142] Espressif Systems. esp32-c6_datasheet.

[143] FTDI Chip. FT232R - Datasheet.

[144] Mohammad Nasar and Mohammad Abu Kausar. Suitability of influxdb database for iot
applications. International Journal of Innovative Technology and Exploring Engineering,
8(10):1850–1857, 2019.

[145] Guangyao Zhang, Yi Wang, Xiaomeng Li, Yi Qin, and Baoping Tang. Health indica-
tor based on signal probability distribution measures for machinery condition monitor-
ing. Mechanical Systems and Signal Processing, 198:110460, September 2023. doi:
10.1016/j.ymssp.2023.110460.