,,Dunarea de Jos University of Galati (UDJG) [602095]

,,Dunarea de Jos” University of Galati (UDJG)
Faculty of Engineering – Department of Manufacturing Engineering

Université M’Hamed Bougara de Boumerdes (UMBB)
Faculty of Engineering Sciences – Department of Mechanical Engineering

Predictive Maintenance for a Grinding Machine
Tool for Bearing Components Manufacturing,
Using ANN Method

Superviso rs:
Prof. Catalina MAIER (UDJG)
Assoc Prof. Florin SUSAC (UDJG)
Assoc Prof. Mohamed A .MELLAL (UMBB)

Master student: [anonimizat]. Abdelkader RIAL

Galati – 2019

2
ALGERIAN NATIONAL LABORATORY FOR MAINTENANCE
EDUCATION (ANL MEd)
Description
The Algerian National Laboratory in Maintenance Education, ANL -MEd, has the mission to create
the next generation educated workforce in industry. There are two major driving forces for the
ANL -MEd project proposal: i) Matching the educational and training programmes at universities
to the needs of industry and generally of the Algerian economy, for creation of new jobs. ii)
Creation of a strong coalition between university – industry – governmental organizations for long –
term collaboration in education, tra ining and research, for revitalization of Algerian economy and
in particular of Algerian industry.
ANL -MEd assembles for 36 months a consortium with 14 partners with unique combination of
skills and expertise. The consortium, coordinated by USTHB, has a hi erarchical structure that
ensures an efficient communication and cooperation. Four European universities with solid
competence in maintenance engineering and management will contribute to development of
teaching material and training students, teacher, tra iners and industry staff. The four Algerian
universities will collaborate with EU partners and the Algerian industrial partners to develop and
implement the specialization and training programmes, and to create the ANL -ORG, the national
laboratory which wi ll coordinates all activities related to maintenance education. The key factor
for implementation of the project objectives is the active collaboration between academic and
industrial partners. Therefore, important resources have been allocated for creatin g a harmonious
working environment – ANL -ORG – with activities for creating synergies between project
partners and stakeholders. This will contribute to strengthening the active cooperation between
university and industry, as well between Algeria and EU.
Objectives
The goal of the ANL -MEd is to provide Algerian industry with a new generation of skilled
personnel, especially at mid – and higher levels, along with ensuring flexible and continuing
education and training of industry personnel at all levels. Fle xibility is necessary to adapt and
update the knowledge according to the day -to-day progress of science and technology. The
partners in the project have been selected above all to suit to the genuine structure of Algerian
academia and industry and to align it to a modern and dynamic European standard. The reference
line is represented by the two main character istics of the Algerian industry:
– Relative low productivity determined especially by the lacking knowledge in modern
manufacturing technology, maint enance and asset integrity.
– Few large companies and many SMEs with very small resources for internal development.

3
ACKNOWLEDGEMENT

This work was supported by the Erasmus+ Student: [anonimizat]ersity of Gal ati, and University M’Hamed Bougara of Boumerdes, within the framework of
KA2 Erasmus+ project ANL Med, Reference no.: 586035 -EPP-1-2017 -1-DZ EPPKACBHE -JP,
academic year: 2018 -2019 .

I, want to t hank Professor Viorel PAUNOIU, head of Department of Manufacturing Engineering
from Faculty of Engineering, “Dunarea de Jos” University of Galati, ROMANIA , for his support
during all this period .

I would like to express my gratitude to my main supervisor Professor Catalina MAIER from the
Department of Manufacturing Engineering, Faculty of Engineering, “Dunarea de Jos” University
of Galati , ROMANIA for her thoughtful advice s and h er orientations during the establishment of
this work , as well as Professor Florin SUSAC, who has been a truly dedicated mentor .

I would like to acknowledge URB Group Rulmenti SA Bârlad, for providing me the database used
in this thesis, Q A -Director of the factory mister Mitica AFTENI is also acknowledged.

I would also like to acknowledge Professor Smail ADJERID and Professor. Djamel BENAZZOUZ
the coordinators of this project , from Department of Mechanical Engineerin g, Faculty of
Engineering Sciences, Unive rsity M'Hamed Bougara Boumerdes, ALGERIA for their infinite
support during the establishment of this work, and a special thanks goes to Professor Mohamed
Arezki MELLAL my supervisor, from Department of Mechanical Engineerin g, Faculty of
Engineering Sciences,University M'Hamed Bougara, Boumerdes, ALGERIA , for his constant trust
in my work.

I, sincerely would like to thank my colleagues from ALGERIA, with whom, I shared this
exceptional experience.

Finally, I must express my deep gratit ude to my parents, my brother Walid and my sister, for their
constant support and continuous encouragement.

4
TABLE OF C ONTENTS

ALGERIAN NATIONAL LA BORATORY FOR MAINTEN ANCE EDUCATION (ANL
MED) ………………………….. ………………………….. ………………………….. ………………………….. ………….. 2
ACKNOWLEDGEMENT ………………………….. ………………………….. ………………………….. ………… 3
TABLE OF CONTENTS ………………………….. ………………………….. ………………………….. …………. 4
LIST OF FIGURES ………………………….. ………………………….. ………………………….. …………………. 7
LIST OF TABLES ………………………….. ………………………….. ………………………….. …………………… 9
LIST OF ABBREVIATION S ………………………….. ………………………….. ………………………….. …. 10
ABSTRACT ………………………….. ………………………….. ………………………….. ………………………….. . 11
KEY WORDS ………………………….. ………………………….. ………………………….. ………………………… 11
RESUME ………………………….. ………………………….. ………………………….. ………………………….. ….. 11
ملخص ………………………………………………………. ………………………………………………………. ………….. 12
GENERAL INTRODUCTION ………………………….. ………………………….. ………………………….. .. 13
CHAPTER I. ………………………….. ………………………….. ………………………….. …………………………. 15
PRESENTATION OF THE WORKING ENVIRONMENT ………………………….. …………….. 15
I.1. GENERAL PRESENTATION OF “DUNAREA DE JOS” UNIVERSITY OF GALATI ……………………. 15
I.2. GENERAL PRESENTATION OF URB GROUP – RULMENTI S.A BÂRLAD ………………………….. . 18
CHAPTER II. ………………………….. ………………………….. ………………………….. ………………………… 23
INTRODUCTION TO MAIN TENANCE ………………………….. ………………………….. ……………. 23
II.1. INTRODUCTION ………………………….. ………………………….. ………………………….. ………………. 23
II.2. DEFINITION OF MAINTENANCE ………………………….. ………………………….. ……………………… 23
II.3. TYPES OF MAINTENANCE ………………………….. ………………………….. ………………………….. …. 25
II.3.1. Corrective or Accidental maintenance ………………………….. ………………………….. …….. 28
II.3.1.1. Palliative corrective maintenance ………………………….. ………………………….. ……… 28
II.3.1.2. Curative corrective maintenance ………………………….. ………………………….. ………. 29
II.3.2. Preventive maintenance ………………………….. ………………………….. ………………………… 29
II.3.2.1. Systematic preventive maintenance ………………………….. ………………………….. ….. 30
II.3.2.2. Conditional preventive maintenance ………………………….. ………………………….. …. 30
II.3.3. Predictive maintenance ………………………….. ………………………….. …………………………. 32
II.3.4. Zero hours maintenance (Overhaul) ………………………….. ………………………….. ……….. 33
II.3.5. Periodic maintenance (Time Based Maintenance TBM) ………………………….. ………… 34
II.4. PREDICTIVE MAINTENANC E ORGANIZATION ………………………….. ………………………….. …….. 34

Table of C ontents

5
II.4.1. Inspection methodology ………………………….. ………………………….. ………………………… 35
II.4.2. Fault Analysis ………………………….. ………………………….. ………………………….. …………. 36
II.4.2.1. the objective of fault analysis ………………………….. ………………………….. …………… 36
II.4.2.2. Data to be collected to study a fault ………………………….. ………………………….. ….. 36
II.4.2.3. Causes of failures ………………………….. ………………………….. ………………………….. . 37
II.4.2.4. Preventive measures to be taken in cases of failures ………………………….. ……….. 39
II.4.3. Prioritization ………………………….. ………………………….. ………………………….. …………… 40
II.5. CONCLUSION ………………………….. ………………………….. ………………………….. …………………. 41
REFERENCES ………………………….. ………………………….. ………………………….. ………………………… 42
CHAPTER III. ………………………….. ………………………….. ………………………….. ………………………. 44
THEORETICAL ASPECTS REGARDING ARTIFICIAL NEURAL NETWORK
MODEL FOR PREDICTIVE MAINTENANCE ………………………….. ………………………….. …. 44
III.1. INTRODUCTION ………………………….. ………………………….. ………………………….. ……………… 44
III.2. BASIC NEURAL NETWORKS ………………………….. ………………………….. …………………………. 45
III.2.1. Properties of the biological neuron ………………………….. ………………………….. ………… 45
III.2.2. the artificial neuron ………………………….. ………………………….. ………………………….. …. 46
III.2.3. the neural network ………………………….. ………………………….. ………………………….. ….. 47
III. 2.4. History ………………………….. ………………………….. ………………………….. …………………. 52
III. 2.5. Fields of application of artificial neural networks ………………………….. ……………….. 55
III. 2.6. Architecture of Neural Networks ………………………….. ………………………….. …………. 56
III.3. CONCLUSION ………………………….. ………………………….. ………………………….. ………………… 59
REFERENCES ………………………….. ………………………….. ………………………….. ………………………… 60
CHAPTER IV. ………………………….. ………………………….. ………………………….. ……………………….. 64
DEVELOPMENT OF PREDI CTIVE MAINTENANCE MO DEL USING THE NEURAL
NETWORK ALGORITHM, F OR A GRINDING MACHIN E-TOOL ………………………….. 64
IV.1. INTRODUCTION ………………………….. ………………………….. ………………………….. ……………… 64
IV.2. WHY PREDICTIVE MAINTE NANCE MODEL ? ………………………….. ………………………….. …….. 64
IV.3. GRINDING PROCESS PRES ENTATION ………………………….. ………………………….. ………………. 65
IV.4. ARTIFICIAL NEURAL NETWORK MODELIZATION ………………………….. …………………………. 68
IV.4.1. Steps in the building of data base to use for ANN modelling ………………………….. … 68
IV.4.2. Identification of input and output variables ………………………….. ………………………… 73
IV.4.3. Comparison between EASY NN and NN MODEL ………………………….. ……………… 73
IV.4.4. Phases 1: Training 30 data sets ………………………….. ………………………….. ……………… 75
IV.4.5. Phases 2: Training 25 data sets and validation with 5 data sets ………………………….. 84
IV.5. CONCLUSION ………………………….. ………………………….. ………………………….. ………………… 93
REFERENCES ………………………….. ………………………….. ………………………….. ………………………… 94
CHAPTER V. ………………………….. ………………………….. ………………………….. ………………………… 95

Table of C ontents

6
GENERAL CONCLUSIONS & ORIGINAL CONTRIBUT IONS ………………………….. ……. 95
V.1. THE MAIN CONTRIBUTI ONS OF THIS THESIS ………………………….. ………………………….. ……… 95
V.1.1. Part one – state of the art ………………………….. ………………………….. ……………………….. 96
V.1.2. Part Two – Scientific Contributions ………………………….. ………………………….. ………… 97
V.1.3. Part Three – Industrial Exploitation ………………………….. ………………………….. ………… 98
V.2. PROSPECT ………………………….. ………………………….. ………………………….. ……………………… 99
V.2.1. Scientific perspectives ………………………….. ………………………….. ………………………….. 99
V.2.2. Industrial exploitation prospects ………………………….. ………………………….. …………… 100
V.2.3. Future work ………………………….. ………………………….. ………………………….. …………… 100
ANNEX 1. ………………………….. ………………………….. ………………………….. ………………………….. .. 101
INITIAL DATA BASE FR OM URB GROUP – RULME NTI S.A. BÂRLAD ……………….. 101
ANNEX 2. ………………………….. ………………………….. ………………………….. ………………………….. .. 116
NEW DATA BASE ………………………….. ………………………….. ………………………….. ………………. 116
ANNEX 3. ………………………….. ………………………….. ………………………….. ………………………….. .. 120
INTRODUCTION TO EASY NN-PLUS ………………………….. ………………………….. ……………. 120
ANNEX 4. ………………………….. ………………………….. ………………………….. ………………………….. .. 123
THE RESULTS OF THE P REDICTED VALUES USIN G NN MODEL …………………….. 123

7
LIST OF FIGURES

Figure I. 1. Map of Romania ………………………….. ………………………….. ………………………….. …………………. 15
Figure I. 2. Entrance of the Faculty of Engineering ………………………….. ………………………….. ………………. 18
Figure I. 3. URB Group ………………………….. ………………………….. ………………………….. ……………………….. 19
Figure I. 4. URB Group – Rulmenti S.A. ………………………….. ………………………….. ………………………….. .. 20
Figure I. 5. Sales regions of URB ………………………….. ………………………….. ………………………….. ………….. 20
Figure I. 6. URB Group Softwares ………………………….. ………………………….. ………………………….. ………… 21
Figure I. 7. Tooling workshop ………………………….. ………………………….. ………………………….. ………………. 22

Figure II. 1 . Forms of maintenance according to standard NF EN 13306 (2010) 24
Figure II. 2 .Concepts, strategies, activities, maintenance operations 25
Figure II. 3. Principle of Conditional Maintenance Follow -up 32
Figure II. 4. The goals of predict ive maintenance 33
Figure II. 5. Steps to do for solve a problem 36

Figure III. 1. Biological structure of neurons ………………………….. ………………………….. ………………………. 45
Figure III. 2. Biological Neuron / Artificial Neuron Mapping ………………………….. ………………………….. .. 46
Figure III. 3. The generic artificial neuron ………………………….. ………………………….. ………………………….. 46
Figure III. 4. Structure of a neural network / Layers of artificial neural network ………………………….. ….. 48
Figure III. 5. Simplified view of the neural network organization (all neurons are fully connected) ……. 49
Figure III. 6. The most important steps in ANN construction and testing ………………………….. ……………. 50
Figure III. 7. Different typ es of unconnected and recurrent networks ………………………….. …………………. 56
Figure III. 8. A feed -forward, back -propagation neural network ………………………….. ………………………… 57
Figure III. 9. Partially recurrent neural network connections ………………………….. ………………………….. …. 58
Figure III. 10. Interconnexion structure or a feedback network model ………………………….. ………………… 58

Figure IV. 1. Comparison between NN MODEL and EASY NN ………………………….. ……………………….. 74
Figure IV. 2. Controls ………………………….. ………………………….. ………………………….. ………………………….. 76
Figure IV. 3. Input/output Data Entry with Easy NN -Plus ………………………….. ………………………….. …….. 77
Figure IV. 4. Graphic representation of a Neural Network in Easy NN -Plus ………………………….. ………… 78
Figure IV. 5. Neural model archite cture ………………………….. ………………………….. ………………………….. …. 78
Figure IV. 6. Learning progress ………………………….. ………………………….. ………………………….. …………….. 79
Figure IV. 7. Prediction ………………………….. ………………………….. ………………………….. ……………………….. 80
Figure I V. 8. Column values ………………………….. ………………………….. ………………………….. …………………. 81
Figure IV. 9. Example errors ………………………….. ………………………….. ………………………….. ………………… 81
Figure IV. 10. Input importance ………………………….. ………………………….. ………………………….. ……………. 83

List of F igures
8
Figure IV. 11. Sensitivity ………………………….. ………………………….. ………………………….. ……………………… 83
Figure IV. 12. Controls parameters for 2 -nd phase ………………………….. ………………………….. ……………….. 84
Figure IV. 13. Table of IN/OT ………………………….. ………………………….. ………………………….. ………………. 85
Figure IV. 14. Process learning progress ………………………….. ………………………….. ………………………….. … 86
Figure IV. 15. Prediction ………………………….. ………………………….. ………………………….. ……………………… 87
Figure IV. 16. Column values ………………………….. ………………………….. ………………………….. ……………….. 88
Figure IV. 17. Examples errors ………………………….. ………………………….. ………………………….. ……………… 89
Figure IV. 18. Comparison between predicted values and experimental values for training examples …. 90
Figure IV. 19. Comparison between predicted values and experimental values for validation examples . 91
Figure IV. 20. Input values ………………………….. ………………………….. ………………………….. …………………… 92
Figure IV. 21. Sensitivity ………………………….. ………………………….. ………………………….. ……………………… 92

9
LIST OF T ABLE S

Table IV. 1. Difference between machining and grinding ………………………….. ………………………….. ……… 67
Table IV. 2. Initial data base ………………………….. ………………………….. ………………………….. …………………. 68
Table IV. 3. New data base ………………………….. ………………………….. ………………………….. …………………… 69
Table IV. 4. Machine Tools codification ………………………….. ………………………….. ………………………….. … 70
Table IV. 5. Defect codes ………………………….. ………………………….. ………………………….. ……………………… 70
Table IV. 6. Maintenance data base ………………………….. ………………………….. ………………………….. ……….. 72
Table IV.7. Evaluation neural model performance ………………………….. ………………………….. ……………….. 82
Table IV.8. Evaluation of neural model performance for training examples ………………………….. …………. 90
Table IV.9. Evaluation of neural model performance for validating ………………………….. ……………………. 91

10
LIST OF ABBREVIATIONS

TBM: time based maintenance.
TPM: total productive maintenance.
ANN: artificial neural network.
MLP: Multilayer perceptron.
PMC: perceptron multicouche.
RFR: radial basic feature network.

11
ABSTRACT

The work consists in developing, testing and validating in a manufacturing company a neural model
for predictive maintenance of a grinding machine -tool type that has, during manufacturing process ,
the highest number of interruptions due to unpredicted maintenance actions. Based on t hree years
maintenance history, a data base will be realized. This data base will be used to generate a neu ral
model of the predictive maintenance. The neural model will be interrogated and , based on the
results, the predictive maintenance plan will be drawn up. Applying this neural model and the
predictive maintenance plan in the industrial environment could lead to cost reduction.
Key words
Maintenance ; Predictive maint enance ; Grinding machine; Bearings ; Artificial neural networks ;
Diagnosis .
RESUME

Le travail consiste à développer, tester et valider dans une entreprise un modèle neuronal pour la
maintenance prédictive d'un type de machine -outil à rectifier, qui a au cours du processus de
fabrication le plus grand nombre d'interruptions dues à des opérations de maintenance imprévues.
Sur la base d'un h istorique de la maintenance sur trois ans, une base de données sera réalisée. Cette
base de données sera utilisée pour générer un modèle neuronal pour la maintenance prédictive. Le
modèle neuronal sera interrogé et, bas é sur l es résultats obtenus , le plan de la maintenance
prédictive sera élaboré. L'application de ce modèle neuronal et du plan de maintenance prédictive
dans l'environnement industriel pourrait entraîner une réduction des coûts.
Mots clé s
Maintenance ; Maintenance prédictive ; Machine -outil de rectification ; Roulements ; Réseau de
neurone artificiel ; Diagnostic.

12
ملخص

يتمثل العمل في تطوير واختبار والتحقق من صحة نموذج عصبي في شركة تصنيع من أجل الصيانة التنبؤية لنوع من األدوات
اآللية التي يتعين تصحيحها، والتي شهدت أثناء عملية التصنيع أكبر عدد من االنقطاعات بسبب عمليات الصيانة غير المتوقعة.
استنادا إلى تاريخ الصيانة لمدة عامين، سيتم بناء قاعدة بيانات. سيتم استخدام قاعدة البيانات هذه إلنشاء نموذج عصبي، وفي
الخطوة األخيرة، سيتم االستعالم عن النموذج العصبي، وبناءً على النتائج، سيتم تطوير خطة الصيانة التنبؤية. قد يؤدي تطبيق
هذا النموذج العصبي في البيئة الصناعية إلى توفير التكاليف.
الكلمات المفتاحية
الصيانة؛ الصيانة التنبؤية؛ أداة آلة لطحن؛ محامل؛ شبكة عصبية اصطناعية؛ التشخيص.

13
General Introduction

The end -of-studies project Industrial Maintenance Master, carried out at the University of Galati,
is part of the student mobility project of the European Union « Algerian National Laboratory for
Maintenance Education Project No. 586035 -EPP-1-2017 -1-DZ-EPPKA2 -CBHE -JP”. The
universities DUNAREA DE JOS UNIVERSITY OF GALATI (Romania) and UNIVERSITY
M'HAMED BOUGARA OF BOUMERDES (Algeria) are partners in this international project.
The end of studies project is entitled: 'Predictive maintenance for a grinding machine tool for
bearing components manufacturing, using ANN method'.
Manufactu ring industries are steadily rising to reach high levels of productivity. The success of
these companies relies to a large extent on maintaining the reliability of the equipment. Increased
equipment efficiency translates into improved productivity and prof itability. In the industrial
sector, it is imperative to maintain a high level of equipment reliability through rigorous
maintenance programs. As a result, there is a growing need for maintenance management programs
to eliminate unforeseen events. Breakdow ns, minimize unplanned downtime and reduce
maintenance costs.
Failure prediction is essential for predictive maintenance because of its ability to prevent failures
and maintenance costs. At the present, artificial neural networks are the main approaches used in
diagnosis.
The goal of our work is to develop, test, and validate in a manufacturing company a data -driven
neural model to predict the future performance of a mechanical system for preventive mainte nance
of a type of machine tool . This type of machine tool h as the largest number of interruptions during
the manu facturing process due to unplanned maintenance interventions , in the considered industrial
company. The artificial neural network technique is based on a complete technical evaluation of
events occurring during the experimental or operational phases.
The g rinding machine tools are an integral and indispensable part of modern industry. Failure of
different machines component can compromise production and increase maintenance costs.

General I ntroduction

14
Predictive maintenance has become synonymous with the indispensable monitoring of grinding
machine tools to prevent catastrophic faults and failures that can alter their operation. It is for this
reason that we have developed our neural model.
To achieve the assigned objective of this study, we designed and r ealized a database fro m the
original database of Bârla d-Romania that allowed us to develop our neural model by applying a
technique of artificial intelligence.
The work in this thesis has been divided into five chapters.
In the first chapter we present the industrial environment and Dunarea de Jo s University of Galati.
The second chapter presents itself as state of the art, it is an introduction in the field of maintenance,
monitoring and diagnosis.
The third chapter is devoted to the basic neural network e lement s, the properties of a biological
neuron, and some types of neural networks, ending with a conclusion for th is chapter.
We divided the fourth chapter into two part theoretical part contains, a general description of a
machine grinding tools and a practical part contains the development of a neural model from a
database to use it at the predictive maintenance, and to predict the average error percentage.
This thesis end with a general conclusion and a synthesis of the contributions made as well as t he
tracks defining possible perspectives for future works.

15
CHAPTER I.
Presentation o f the Working E nvironment

My internship took place in Romania, in a laboratory of the Department of Manufacturing
Engineering, Faculty of Engineering, Dunarea de Jos University of Galati. During my internship I
visited the society URB Group – Rulmenti S.A. from Bârla d. In my project I have used a data base
obtained from this manufacturing c ompany in order to generate a neural model for predictive
maintenance of grinding machine -tools .

I.1. General Presentation of “Dunarea de Jos” University of Galati

Galati city , which is known to be the fifth most important town of the country, is set at the east of
the land, close to the border with Moldova and Ukraine, and on the banks of the end of Danube.
This region owns a deeply rooted industrial base, as the most major Ro manian steelworks (Arcelor
Mittal Galați) or the dockyard can bear witness.

Figure I. 1. Map of Romania

CHAPTER I. Presentation of the Working E nvironment
16
The main asset of Galați, in terms of education and scientific research, lies in the Dunărea de Jos
University. It is made up of fifteen faculties with more than thirty departments. It has diverse
unique fields of education in the country, such as naval engineering or fishery. It also organizes
programs for doctoral and master degrees in various technical fields, such as chemistry, physics,
mathematics, economy, food technology and fishing, automatic control and computation
techniques, artificial intelligence, or even social and humanistic sciences. During the years,
specialists covering a wide range of education fields ha ve been trained in this University, like
engineers, teachers and programmers.

Dunarea de Jos University of Galati is the most important institution of higher education in the
South -East of Romania.
Dunarea de Jos University of Galati functions according to the university Charter, whose
provisions are in agreement with the national legislation and with the principles of the European
Space and Higher Education, being recognized by all members of the university commu nity.

The history of higher education in Galati covers the following stages:
1948: establishment of the Land Improvement Institute;
1951: establishment of the Naval -Mechanical Institute;
1953: merging the Naval -Mechanical Institute with the Agronomic In stitute, and with the Fish
Farming and Fishing Institute (transferred from other university centres), and the establishment of
the Technical Institute in Galati;
1955: merging of the Technical Institute with the Food Industry Institute in Bucharest;
1957: transforming the Technical Institute into the Polytechnic Institute;
1959: establishment of the Pedagogic Institute and relocation of the Land Improvement Institute
to Iași;
1974: establishment of the University of Galati by merging the Polytechnic Instit ute with the
Pedagogic Institute (State Council Decree of 20 March 1974);
1991: the University of Galati becomes Dunarea de Jos University of Galati (Government
Decision of 4 January 1991).

CHAPTER I. Presentation of the Working E nvironment
17
In the structure of the above mentioned institutes, there were a series of study programs that
were unique in the country: Naval Constructions, Harbours and Ship Exploitation, Food
Industry, Fish Farming Technology, Cooling Devices – which meant that an important creation
process on elaborating educational curric ula and syllabi, lectures, laboratory equipment etc.,
presently being used in other university center s around the country, was fully the work of the
academics in Galati higher education.

The internship took place more accurately in a division of the Un iversity, w hich is the Faculty of
Engineering . This entity is an historic one of the university, because it directly results from the
former Technical Institute of Galați.

Initially reputable for its competence in two domains, namely naval construction and technology
of ships and ports, this division has regularly grown through the years, by extending its
specializations such as:

– Refrigeration and T echnology of machine building in 1960
– Thermal machines and W elding technology in 1978
– Metallurgical engineering , Mechatronics and Robotics and Economical Engineering in
1990 .

Become in the late nineties one of the largest faculty of mechanics in the country, a breaking up
took place, which resulted in a transfer of competences to the F aculty of Engineering of Brăila and
the Technical C ollege. Nowadays, the mission assumed by the faculty of mechanics is to form
specialists through training (normal and post university studies) in areas such as mechanical
engineering, industrial engineering, or environmental engineering.
This mission includes the development of scientific research centers in strong fields, promoting
the national and international cooperation of inter -university and economic environment, and
contributing to the universal herita ge of knowledge. Over the years, the faculty of mechanics has
forged strong ties with the field of industry, even with multinational companies such as Dacia,
Fiat, or Mittal Steel. In terms of figures, this entity represented in 2008:

CHAPTER I. Presentation of the Working E nvironment
18
– More than 7 0% of res earch projects accepted for the whole university
– More than 1400,000 € allocated for research, a tenfold growth over the last 3 years
– More than 50 papers published in professional journals
– More than 30 scientific papers published in conferences

Figur e I. 2. Entrance of the Faculty of Engineering

I.2. General Presentation of URB Group – Rulmenti S.A B ârlad
The manufacturin g company, Rulmenti S.A. from Bârla d, Romania, is a part of the large URB
Group.

CHAPTER I. Presentation of the Working E nvironment
19

Figure I. 3. URB Group
The short history of this company is:
1953 -Bârlad Factory started production ;
1973 -The Forging technology from Haterbur and Wagner Germany was applied successfully ;
1975 -1982 – Modernization of turning and grinding workshops ;
1992 -1996 – Continuous modernization of various machinery and equipment ;
2000 – Finalization of privatization process ;
2001 -2005 – A period of big investments in machine technologies ;
2006 – Founding Anadolu Rulman factory (ART) in Turkey ;
2007 – Acquisition of MGM factory in Hungary ;
2011 – Starting the building of India factory ;
2018 – Continued investment .

CHAPTER I. Presentation of the Working E nvironment
20

Figure I. 4. URB Group – Rulmenti S.A.
The Group URB produce different kinds of bearings (cylindrical, spherical, radial, axial), turning
components and other components. The sales regions of their products are presented in the figure
5.

Figure I. 5. Sales regions of URB

CHAPTER I. Presentation of the Working E nvironment
21
The conception activity in URB Group use computer aided design software Pro/ENGINEER, a 3D
CAD/CAM/CAE parametric software, and the computer aided analysis software – ANSIS in order
to assure the good quality of their products.

Figure I. 6. URB Group Softwares
The total area of tooling workshop (figure 7) is 10,900 m2. There are 357 direct machines & 114
auxiliary machines , capable of:

– Turning;
– Milling ;
– Grinding ;
– Metallic structuring (bending, shocking, rolling, welding) ;
– Heat treatment ;
– Titanium covering ;
– Electrical discharging .

CHAPTER I. Presentation of the Working E nvironment
22

Figure I. 7. Tooling workshop
One of the main part of tooling workshop is Grinding Section consisting in following grinding
equipment:

➢ CNC – grinding machines ;
➢ Centerless grinding machines ;
➢ Classic grinding machines ,
And quality c ontrol equipment for :
➢ Dimensional (ID/OD/H, angles, raceways, recess, radius)
➢ Form deviations (ovality, profile deviation, parallelism, conicity, raceway deviations)
➢ Surface checki ng (surface defects, roughness).

My work is focused on this section of the manufacturing company. The objective of my work
consists in develop ing, tes ting and validating in this manufacturing company a neural model for
predictive maintenance of a grinding machine -tools , which has during manufacturing process the
highest number of interruptions due to unpredicted maintenance actions .

23
CHAPTER II .
Introduction to Maintenance

II.1. Introduction

One of the strategic problems facing the manufacturers , from the design of a machine to its
operation, is the maintenance of the operational mode of industria l systems, prediction and
diagnosis of defects presents an important solution and priority for a company anxious to reduce
the costs of maintenance and depreciation of equipment purchase.
The search for performance is of great importance in the organization of companies because a
large company constantly seeks to increase i ts competitiveness to ensure its sustainability; it can
implement a policy of cost reduction or development.
On the other hand, maintenance methods use information based on experience for predicting future
performance. During the experimental phases these methods need a systematic discipline which is
based on a complete technical evaluation of the events that have occurred, t o give product design
information that can predict the future behavior of new hardware.
In this chapter we define some concepts relate d to maintenance and industrial diagnostics,
predictive maintenance techniques.

II.2. Definition of Maintenance

A first normative definition of maintenance was given by AFNOR in 1994, namely: "all actions to
maintain or restore a property in a specified state or able to ensure a specific service " [1].
Since 2001, it has been replaced by a new definition, now European [2]: "Set of all the technical,
administrative and management actions during the life cycle of a good, intended to maintain it or
to restore it to a state in which it can perform the required function."

CHAPTER II. Introduction to Maintenance
24

Figure II. 1. Forms of maintenance according to standard NF EN 13306 (2010)
The European Federation of National Maintenance Societies (EFNMS) propos es a similar
definition : "All actions which have the objective of retaining or restoring an item in which they
can perform their required function. The actions include the combination of all techni cal and
corresponding administrative, manag erial and supervisory actions ".
Maintenance can be said to encompass a set of concepts, strategies, activities and operations
(Figure II.1.2), all aimed at correcting and /or preventing the failure or impairment of the value of
technical systems.
A technical system is defined as any part, component, device, subsystem, functional unit,
equipment or system that can be individually maintained.
Failure is considered the end of an item's ability to perform a n action or provide a service as
needed. [4]
So maintenance is an activity required to keep assets close to their original state to ensure that the
asset can perform the function for which it was developed or even purchased.

CHAPTER II. Introduction to Maintenance
25

Figure II. 2.Concepts, strategies, activities, maintenance operations
II.3. Types of m aintenance
When talking or writing about types of maintenance, it is in fact about classifying maintenance
tasks into different categories.
Types of maintenance in accordance with the specialization of the technician
In the first place, maintenance can be divided into the following general types, in accordance with
the specializ ation of the technician who carries out the tasks:
• Operational maintenance, the one carried out by operation personnel.
• Mechanical maintenance, made by mechanic specialists.
• Electrical maintenance, carried out by electric specialists.
• Maintenance of instrumentation, executed by instrumentation specialists.
• Control maintenance, made by control specialists.
• Technical cleanings, carried out by technicians specialized in cleaning certain part of the
equipment which require specific and complex procedure.
Types of maintenance in accor dance with scope
In accordance with the scope of the tasks, maintenance can be divided into the following types:

CHAPTER II. Introduction to Maintenance
26
• Routine maintenance, generally carried out by production personnel.
• Underway maintenance, carried out by maintenance staff yet not needing plan t, area or a whole
system stoppage. Thus only implying a narrow number of equipment and subsystems which
do not affect production.
• Minor revisions or inspections, when a limited amount of elements are inspected or substituted.
• Major revisions, when tasks t o be accomplished imply the substitution of a great number of
wearing elements, or the inspection of certain internal parts that demand great dismantling.
Types of maintenance in accordance with anticipate faults
Lastly, according to anticipate faults, there are three major groups of maintenance tasks. In several
cases, when talking about `types of maintenance´, one maybe thinking about this last category.
The types of maintenance according to anticipate faults are the following:
• Corrective maintenance , gets done after the failure has been made and has as its essential
goal, its own adjustment.
• Diagnosis , has as its essential goal the knowing of the machines or installation condition
in order to decide if an intervention has to be done on it. Generally, an observation or
measurement is related to the condition of an installation. Diagnosis tasks frequentl y
include an estimation of its criticality, evaluating the potential severity in cases when the
potential and the degradation trend of failure gets to materialise. These type of tasks used
to be known as predictive maintenance , even though nowadays this na me is disused,
favouring the concept of diagnosis , which is a wider concept and which defines better the
wilfulness of the tasks. The diagnosis includes four types of tasks:
• Easy inspection tasks, very often generally carried out by production technician. It involves
simple sensorial inspections, commonly made with the senses, without needing measure
tools or additional technical means. Thus including visual inspection, observation of
strange noise and vibration emissions and the recognition of abnormal odo urs. It also
includes the reading and record of operating parameters, with instruments installed in the
equipment. They require basic training and they can be carried out by any technician. Due
to its simplicity these inspections can be done very often.
• Online diagnosis tasks, which are carried out basing on the readings obtained from the
inline mounted instrumentation and which are received in the control system.

CHAPTER II. Introduction to Maintenance
27
• Offline diagnosis tasks, which are carried out with instruments that are especially
assembled for making the observation or measurements which are used to diagnose. Such
an example are the analysis of vibration, thermography, the analysis of ultrasound, oils,
fumes produced by combustion and so forth . These are made with offline instruments.
• Detail ed inspection of tasks, carried out by specialized maintenance technicians and which
may or may not require the stoppage of the equipment and systems to make these
inspections. They often demand disassembly or at least, profound observation. As well as
this inspections need specific education for the technician that carries them out, who has to
be able to distinguish between a standard situation, a situation which requires observation
in order to verify its evolution, and an appalling situation which demand s immediate
intervention? Among these detailed inspections, the followings stand out :
• Mechanic verification, like clearance, alignment, thickness, bolt tightening,
starting, operation and stoppage measurements.
• Electrical verification, such as grounding verification, verification of the operation
of emergency stop, connexion verification, and so forth.
• Verification of measuring instruments and functional check of control links. They
may require an intervention in order to adjust certain parameters to default values.
• Check of certain ways of operation or the benefits of an item.
Preventive maintenance ,
• Carried out before a failure made and which has as its main propose prevent from
happening. Preventive maintenance can be further divided into four subtypes, always
taking into account that it is actually about subtypes of maintenance inside the category of
‘preventive maintenance’:
• Conductive maintenance the one carried out by operation technicians and that generally
refers to senses check, data samples, change in fuel and/or adjustment of parameters.
• Systematic maintenance, made from time to time or when a certain number of hours has
passed.
• Hard -time maintenance, overhaul, major revision or zero hours, which is the combination
of tasks made after some time of equipment, system or installation operation, and which
has as its goal return the inspected ensemble to its initial state (as when it has zero hours
of functioning.

CHAPTER II. Introduction to Maintenance
28
• Improvement maintenance, w hich is the combination of tasks which are carried out in a
part of the installation, and has as goal the avoidance of a certain failure to be made or
made again. Although some authors doesn’t consider modifications as ‘maintenance tasks’,
it is logic to c onsider them this way, when they have as goal the avoidance of failures. And
they shouldn’t be consider ‘maintenance tasks’ when they only find improvements in
security, in the environmental or production impact, without affecting failure probability.

II.3.1. Corrective or Accidental maintenance
Corrective maintenance is maint enance performed after failure, [3] defines it as a maintenance
performed after detection of a failure and intended to return a property in a state in which it can
perform a required function. [5]
It is also characterized by its random nature and needs competent human resources and material
resources such as: spare parts and tools, available on site. Among these objectives i s the correction
of defects present in the various equipment and which are communicated to the maintenance
service by the users of the same equipment.
This type of maintenance is generally suitable for equipment for which [6]:
• The consequences of the failure are not critical.
• Repair is easy and does not require a lot of time.
• Investment costs are low.

Two forms of corrective maintenance can be distinguished.

II.3.1.1 . Palliative corrective maintenance
This maintenance is with the curative maintenance, one of the two subdivisions of what is called
corrective maintenance; it mainly consists of temporary actions.
The palliative maintenance designates a repair which allows the propert y to work again while
waiting for a curative intervention which makes it possible to temporarily put back the material to
an acceptable level of performance, the intervention is thus temporary, the standard AFNOR [7]
describes it as: "Corrective maintenance action to enable a good to temporarily perform all or part
of a required function, commonly known as troubleshooting."

CHAPTER II. Introduction to Maintenance
29
II.3.1.2 . Curative corrective maintenance
The curative maintenance qualifies a repair by which a good returns to its initial state, the
interventions are of definitive character, the intervention which follows the failure allows the
restoration of the optimal level of p erformance of the material.
Maintenance Terminology defines it as [3]: "a corrective maintenance action to restore a good in
a specified state to enable it to perform a required function. The result of the actions carried out
must be permanent ".

II.3.2 . Preventive maintenance
Preventive maintenance is a maintenance performed according to predetermined criteria, aimed at
reducing the probability of failure or the probability that an article will fulfill its functions only to
a lesser extent (degradation of operation) [8] .To help avoid equipment failures in use.
Its mission is to maintain a level of certain service on equipment, programming the interventions
of their vulnerabilities in the most opportune time.
So it is to intervene on equipment before it is faulty, interventions are triggered before failures
based on one or more parameters determined after monitoring the behavior of the machine.
This type of maintenance aims to reduce the failure rate or to make it null, by maintaining the
required level of p erformance before the appearance of the fault, the definition given by the
AFNOR [7] is as follows: "Maintenance performed at predetermined intervals or according to
prescribed criteria and intended to reduce the probability o f failure or deterioration of the operation
of a property ", its purposes are:
• Increase the life of the equipment.
• Reduce the probability of service failures.
• Decrease downtime in case of overhaul or breakdown.
• Prevent and also plan interventions for costly corrective maintenance.
• Allow deciding corrective maintenance in good conditions.
• Decrease the maintenance budget.
• Avoid abnormal consumption of energy, lubricant, etc.
• Remove the causes of serious accidents.
The main forms of preventive maintenance are :

CHAPTER II. Introduction to Maintenance
30
II.3.2.1 . Systematic p reventive maintenance
This is one of Preventive Maintenance subtypes. It refers to operations performed systematically,
either according to a schedule (fixed time period), or according to a periodicity of use (number of
hours of operation, number of units produced, number of movements performed, etc.). No
intervention takes place before the deadline determined in advance. So when the maintenance
intervention is performed at fixed and pre -defined interval s, it is called systematic preventive
maintenance. And results in the periodic replacement of parts, without prior checking and whatever
the state of degradation of the goods, the definition given by the European standard [7] is:
"Preventive maintenance carried out at pre -established time intervals or according to a defined
number of units of use but without prior checking of the condition of the property ".
This form of maintenance requires knowing the behavior of the equipme nt; wear; the modes of
damage; the average time of good operation between two accidents (MTBF) to determine the
periods of interventions.

II.3.2.2 . Conditional p reventive maintenance
It is maintenance subordinate to a predetermined type of event (self-diagnosis, information given
by a sensor, measurement of a wear …) revealing the state of degradation of a goo d. [1]
Systematic preventive maintenance is likely to lead to an excess of unnecessary interventions, and
thus to unacceptable financial losses for the company. For this, other forms of preventive
maintenance have appeared which are based on the monitoring of the real state of the goods: the
conditional and provisional maintenance. Condition al maintenance is defined as: "Preventive
maintenance based on a monitoring of the operation of the property and / or significant parameters
of this operation integrating the resulting actions ". [7]
In these main objectives is the tracking insurance of the equipment during its operation in order to
prevent the expected failures. It does not imply knowledge of the law of degradation and
dismantling of the material.
When an imminent failure occurs, the intervention w ill take place for only certain measurable
parameters that reach a critical threshold.
It is absolutely necessary to find a relation between the measurable parameter and the state of the
material to work with the conditional preventive maintenance; one can give like example of
parameter:

CHAPTER II. Introduction to Maintenance
31
• Physical parameters (pressure, flow, temperature, voltage, intensity…)
• The oil level,
• The frequency of vibration or the level of vibration and noise,
• Measurements of mechanical play,
• The residue conte nt of lubri cants

The benefits of Conditional Preventive Maintenance are:
• The use of equipment components to the maximum of their potential, which makes it possible to
avoid the waste of spare parts and to reduce their stock,
• The elimination of unexpected failures, hence reliability is superior productivity.
• The reduction of production stops.
• The reduction in the duration and costs of interventions, since these are planned.
• The reduction of unscheduled interventions.
• Planning intervention s.
• More targeted intervention.
• Improved availability, security.

Setting up a conditional maintenance consist in following steps:
• A search of the monitoring points of the equipment and the parameters to be measured,
• Establishing the eligibility thresholds for each parameter,
• The choice and purchase of measurement instrumentation,
• Training of qualified personnel for measurements and their operation.

CHAPTER II. Introduction to Maintenance
32

Figure II. 3.Principle of Conditional Maintenance Follow -up

II.3.3 . Predictive maintenance
Predictive maintenance is a technique to predict the future failure point of a machine component,
so that the component can be replaced, based on a plan, just before it fails. Thus, equipment
downtime is minimized and the component lifetime is maximized.
It pursues constantly know and report the status and operational capacity of the installations by
knowing the values of certain variables, which represent such state and operational ability. To
apply this maintenance, it is necessary to identify physical v ariables (temperature, vibration, power
consumption, etc.). Which variation is indicative of problems that may be appearing on the
equipment? This maintenance it is the most technical, since it requires advanced technical
resources, and at times of strong mathematical, physical and / or technical knowledge.
The need to use rotating machines in the industrial environment to force the intervention of
predictive maintenance (PM) which monitors the vibration of these machines to detect emerging
problems and to avoid a catastrophic failure, we find as application of this maintenance monitoring
the infrared image of electrical equipment, motors and other electrical equipment to detect
development problems.
The assurance of the maximum gap between repairs of unexpe cted failures is achieved by
providing the necessary data by means of regular monitoring of a real mechanical status
monitoring for the operating efficiency of the systems treatment [9].

CHAPTER II. Introduction to Maintenance
33
To prevent a loss of operation in a manufacturing process, it is necessary to predict the failure that
was part of advanced predictive maintenance, which has made preventive maintenance (PM) an
effective means of improving reliability [10].
To develop a predictive maintenance model based on a growing prediction methodology several
studies have been conducted. The latter prefer to focus on the construction of the predictive model
and the measurement of its performance so we can sa y that this maintenance uses the diagnosis,
and the monitoring of the state of the machine that's why it to the advantage of reducing the costs
and the time of the maintenance [11] [12], thus to reach a maximum life of machines while
minimizing the risks of failure.

Figure II. 4. The goals of predictive maintenance
II.3.4 . Zero hours m aintenance (Overhaul)
The set of tasks whose goal is to review the equipment at scheduled intervals before appearing any
failure, either when the reliability of the equipment has decreased considerably so it is risky to
make f orecasts of production capacity . This review is based on leaving the equipment to zero hours
of operation, that is, as if the equipment were new. These reviews will replace or repair all items
subject to wear. The aim is to ensure, with high probability, a good working time fixed in advance.
[13]

CHAPTER II. Introduction to Maintenance
34
II.3.5 . Periodic m aintenance (Time Based Maintenance TBM)
The basic maintenance of equipment made by the users of it. It consists of a series of elementary
tasks (data collections, visual inspections, cleaning, lubrication, retightening screws…) for which
no extensive training is necessary, but perhaps only a bri ef training. This type of maintenance is
the based on TPM (Total Productive Maintenance) [13].

This division of types of maintenance has the disadvantage of that each equipment needs a mix of
each of these maintenance types, so that we can not think of applying one of them to a particular
equipment.
In this situation it is convenient to define the c oncept of Maintenance Models . A Maintenance
Model is a mixture of the previous types of maintenance in certain proportions, and it responds
appropriately to the needs of a particular equipment. We think that every equipment will need a
different mix of dif ferent types of maintenance, a particular mix of tasks, so that maintenance
models will be as many as existing equipment.
In the situation of manufacturing company, partner in this work, it result the necessity to realize a
maintenance model mixing the pr esent preventive maintenance component with a predictive one.
Following we present some aspects of the predictive maintenance.
II.4. Predictive maintenance o rganization
Predictive maintenance is a technique to predict the future failure point of a machine component,
so that the component can be replaced, based on a plan, just before it fails. Thus, equipment
downtime is minimized and the component lifetime is maximized.
This technique involves the measurement of various parameters that show a predic table connection
with the component life cycle. Examples of such parameters are as follows:
• Bearings vibration
• Temperature of the electrical connections
• Insulation resistance of the motor coil
The use of predictive maintenance is to establish, firstly, a historical perspective on the relation
between the selected variable and the component life.

CHAPTER II. Introduction to Maintenance
35
Manufacturers of instrumentation and software for predictive maintenance may recommend ranges
and va lues to replace the components of most equipment; this historical analysis makes it
unnecessary in most applications. [14]
II.4.1. Inspection methodology
After performing the predictive maintenance on a machine, can then proceed to the determination
or control of physical variables indicative of the state of the machine. The purpose of this section
is to examine in detail the techniques commonly used for condition -based monitoring, so that they
serve as a guide for their o verall selection. The purpose of monitoring is to obtain an indication of
the machine's condition (mechanical) or state of health, so that it can be used and maintained safely
and economically.
For monitoring, this was understood at the beginning, such as measuring a physical variable
deemed representative of the state of the machine and its comparison with values indicating
whether the machine is in good condition or damaged. With the current automation of these
techniques, the term "monitoring" has also been extended to the acquisition, processing and storage
of data. Depending on the objectives to be achieved by monitoring the condition of a machine, a
distinction must be made between vigilance, protection, diagnosis and forecasting.
• Monitoring of machi nes. Its purpose is to indicate when a problem exists. You must
distinguish between good and bad condition, and if it is bad you should indicate how bad it
is.
• Protection of machines. Its aim is to prevent catastrophic failures. A machine is protected,
when the values that indicate their status reach values considered dangerous, the machine
automatically stops.
• Failure diagnosis. Its aim is to define the specific problem. Its objective is to estimate how
much longer could operate the machine without the ris k of catastrophic failure. [14]
In recent times there has been a tendency to apply predictive maintenance or symptomatic, that is,
by vibration analysis, used oil analysis, wear control, etc…

CHAPTER II. Introduction to Maintenance
36
II.4.2. Fault Analysis
II.4.2.1. the objective of fault analysis
The analysis of faults aims at determining the causes that led to repetitive breakdowns which
entail high costs, in order to avoid losses of productivity so the resolution of these problems is
necessary, to take prevent ive measures in order to avoid this. It is important to emphasize this
dual function of fault analysis:
• Determine the causes of breakdowns .
• Propose measures to avoid these failures once these causes have been identified .
The improvement of the results of the maintenance necessarily involves the study of the incidents
which occurred in the factory and the search for solutions to avoid them. We can take as an example
a broken piece, it is simply modified by a similar piece, it is as if you acted o n the cause that caused
the damage, but only on the symptom, it is necessary to study t he cause and to act accordingly
[14], [15], [19].
Problem solving methods are shown in this figure :

Figure II. 5. Steps to do for solve a problem
II.4.2.2. Data to be collected to study a fault
When studying a fault, it is important to collect all possible data available. Among them, always
must be collected the following ones:

CHAPTER II. Introduction to Maintenance
37
• Detailed accounts which tell what was done before, during and after the breakdown. It is
important to detail the time it occurred, the shift that was present (even those workers
who operated the equipment) and the actions that took place at all times.
• Details of all environmental conditions and external to the machine: external temperature,
humidity (if it is available), equipment cleaning conditions, cooling water temperature,
compressed air humidity, and in general, the terms of any extern al supply that the
equipment needs to function.
• Latest preventive maintenance tasks performed on the equipment, detailing any
abnormalities found.
• Other failures that had the equipment in a given period. In equipment of high reliability
with an MTBF high.
• Internal conditions in which the equipment works. It will be important to note data such as
temperature and pressure at equipment works, supplied flow, and in general, the value of
any variable that we can measure. It is important to focus on the area that has failed, trying
to determine the conditions at that point, but also throughout the equipment, because some
failures have their origin in remote parts of the piece that failed. [14] [15]
II.4.2. 3. Causes of failures
Common causes of failure are usually one or more of these four:
Material failures
It is considered that there has been a failure of the material when the working conditions have been
respected, because a machine failure can seriously affect the productivity and activities of a
company ; it is necessary to ensure the pro per functioning of the machine and equipment and
achieve the longest possible life.
A material can fail in several ways:
– by the wear (for example, anti -frictio n bearings);
– by fatigue (for example, some parts are subjected to cyclic pressure and/or stretching stresses,
in which the applied resistance is not constant, but evolves over time ).

CHAPTER II. Introduction to Maintenance
38
Human mistake by the operation staff
Human error is a major causal factor of failure. May occur due to an error in production personnel
that may be caused by: [16] [14]
• Interpretation mistake of an indicator during normal operation of equipment, which
makes the installation operator take the wrong decision .
• Wrong action faced with a machine failure.
• Lack of clear systematic instructions, such as procedures, technical instructions, etc..,
Or poor implementation of these .
• Lack of training .

Human mistake by maintenance staff
Maintenance staff also make mistakes that lead to a breakdown, a production stop, a decrease in
equipment efficiency, etc…
Among the most common failures caused or aggravated by maintenance staff, include:
• Wrong observations of the inspected parameters.
• Perform assembly and disassembly without following the best industry practices
• Not respecting or not checking adjustment tolerances
• Not respecting or not checking torque setting
• Materials reuse that must be thrown away. For example, reuse of sealing elements
• By the use of inappropriate spare parts: aftermarket parts, which do not meet the required
specifications or spare part that has not been tested before being mounted
• By using the wrong tool.

Abnormal external conditions
When external conditions are different from conditions in which the equipment or installation has
been designed, failures can strike favored by these abnormal conditions. It is the case of equipment
is operating in conditions of temperature, humidity or du st different from those for which it was
designed.

CHAPTER II. Introduction to Maintenance
39
Sometimes, several simultaneous causes bring together the same fault, which greatly complicates
the study of the problem and the solutions contribution. It is important to take into account this, as
with d etermining a single cause in many cases you do not get around the problem, and until all
causes have been solved, you do not have results.
Sometimes, in a fault converge more than one of these causes, which complicates somewhat the
study of failure, becaus e sometimes, it is difficult to determine which one was the main cause and
which one had a minor influence on the failure development.
II.4.2.4. Preventive measures to be taken in cases of failures
Here are some preventive measures that are used depending on the cause of the failure:
Failure of the material
We offer various solutions in case of material failure such as:
• If the fault is caused by wear, you should study ways to reduce wear on the part, if it is not
possible to reduce wear, will be n ecessary to study the life of the part and change it in
advance to failure.
• If the failure is caused by corrosion, the solution will be to apply protective coatings and
devices that reduce it
• If the failure is caused by fatigue, among the solutions to be t aken will be:
➢ Reducing energy and / or frequency of the cyclic stresses to which the part is
subjected.
➢ Changing the material for other material with fewer defects .
• If the failure is caused by dilation, modify the installation so as to allow free expansion and
contraction of the material by thermal effect or by changing supports or by incorporating
elements that absorb expansion and contraction.
Human mistakes by production staff
To avoid the mistakes of the production staff, the first preventive s olution to adopt is to work only
with motivated personnel because the root causes of the accidents are of origin human or
organizational. [17] [14]
The second solution is to train the staff. When it is found that certain deficiencies are due to a lack
of knowledge of some staff, a rapid training action must be organized to solve this problem.
Thirdly, it is possible to make changes to machines to avoi d errors.

CHAPTER II. Introduction to Maintenance
40
Human mistakes by maintenance staff
To prevent failure of maintenance staff in the first place (as in the previous case) the staff must be
motivated and properly trained.
The most effective way of fighting against the mistakes made by maintenance staff is the use of
procedures. The procedures contain detailed information of each of the necessary tasks to perform
a job. They also contain all necessary measures and settings t hat should be made on equipment.
Finally, these procedures detail what checks should be made to ensure that the work has been well
done.
Abnormal external conditions
If it is determined that failure was caused by abnormal external condi tions, the solution to adopt is
simple: correct the external conditions, so as to suit equipment requirements.
II.4.3. Priori tization
The hierarchy of the various generated orders is used, for the resolution of the problems during the
correct management of corrective work orders, and for a good organization of the maintenance
activities which requires a quality instrument and a good communication of the breakdown. [18]
Levels of priority can be many and varied, but in almost all companies with a prioritization system,
there are at least three levels: [14]
• Urgent breakdowns: those that must be resolved immediately, no waiting, as they cause serious
damage to businesses.
• Ma jor faults, although they disrupt the normal operation of the system, can wait for all urgent
faults to be resolved.
• Breakdowns whose solution can be planned. You may want to wait for a stop of the equipment,
or simply, it causes a small mess. Generally, accumulating additional orders on the same
equipment is more interesting.

CHAPTER II. Introduction to Maintenance
41
II.5. Conclusion
Industrial maintenance plays a very important role in an industrial enterprise. Each company
chooses one or more diagnostic methods according to its maintenance policy, according to the
budget allocated to it, and according to the competence and training of its personnel. Most often
manufacturers use vibration analysis, lubricant especially in the case of rotating machinery.
Nevertheless the methods o f artificial intelligence take their place thanks to the different
characteristics and in several cases for their complementarity thanks to the different characteristics.
The first part of this chapter was dedicated to the presentation of some terminologie s, notions and
keywords used in this thesis. Indeed, we have established that the diagnosis of an SAP is done
through three basic functions: detection, localization and identification. In a second part, we
presented a classification of different diagnostic methods. Two main categories of diagnostic
methods were discussed: model -based methods and model -free methods.
In the first part of the chapter was devoted to the presentation of basic concepts because the
diagnosis of probable defects in a machine is don e through three basic concepts: detection,
location, and identification.
On the other hand, we use the second part to understand the diagnostic methods used, which are
model -based methods such as parametric estimation, and modeless methods that are widely used
thanks to its ease of adaptation to complex systems such as: than expert systems and artificial
neural networks.
In this sense, the next chapter will be devoted to the presentation of the basics of neural networks,
and their importance in an industrial environment.

CHAPTER II. Introduction to Maintenance
42
References

[1] ***, "standard NFX 60 -010," in 1994.
[2] ***, "NF EN 13306 X 60 -319," Terminologie de la maintenance, Norme
européenne/Norme française, 2001.
[3] ***, "BS EN 13306:2010," BSI Standards Publication , 2010.
[4] A. Birolini, "Reliability engineering – theory and practice," seventh edition ed., Berlin
Heidelberg, springer -verlag, 2004.
[5] Y. Debbah, A. Cherfia, A. Saadi, ,, "« Application de la méthode des réseaux de neurones
pour la prédiction des vibrations i nduites par des défauts combinés (désalignement et
balourd) »,," in The Second International Conference of Mechanics (ICM’15).,
Constantine, Novembre 2015.
[6] Wu, S. J., Gebraeel, N., Lawley, M. A., Yih, Y, A neural network integrated decision
support system for condition -based optimal predictive maintenance policy., Man and
Cybernetics, Part A: Systems and Humans, IEEE Transactions on Systems,, 2007, pp. 226 –
236.
[7] R. GOURIVEAU, K. MEDJAHER, E. RAMASSO et N. ZERHOUNI, , « PHM –
Prognostics and heal th management – De la surveillance au pronostic de défaillances de
systèmes complexes »,, É. T. d. l’Ingénieur, Ed., 2013, p. (16 pages).
[8] R. E. 1. European Standard EN 13306 Maintenance terminology, "Maintenance
terminology," 2010.
[9] X.-3. EN, "Terminologie de la maintenance.," 2010.
[10] ,. R. Keith Mobley, "” AN INTRODUCTION TO PREDICTIVE MAINTENANCE”".
[11] Jardine, A.K.S., Lin, D., Banjevic, D., A review on machine diagnostics and prognostics
implenting condition -based mainte nance, Mech.Syst.Signal, 2006,, pp. pp. 1483 -1510..
[12] Bertolini, M., Bevilacqua, M., A combined goal programming -AHP approach to
maintenance selection problem, Reliab.Eng.Syst.Saf.91, 2006,, pp. pp. 839 -848..
[13] S. G. Garrido, "« TYPES OF MAINTENANCE »," Petrochemical Maintenance, Remove
Tecnologia S.L, 2009 -2017.

CHAPTER II. Introduction to Maintenance
43
[14] S. G. Garrido, "« PREDICTIVE MAINTENANCE »," Petrochemical Maintenance,
Remove Tecnologia S.L, 2009 -2017,
http://petrochemicalmaintenance.com/predictivemaintenance.html
[15] Frédéric Bernard, Maurice Pillet. Qualita 2001, ., "ANALYSE DES DEFAUTS POUR
L’AMELIORATION DE LA QUALITE (ADAQ).," pp. 9 p. ffhal -00976938, Mar 2001,
Annecy,.
[16] H. MABROUK, "« Approche d’intégration de l’erreur humaine dans le retour d’expérience
»," Vols. ,Synthèse INRETS n° 43, 2003, .
[17] Guarnieri F., Cambon J., Boissières I., , "De l'erreur humaine à la défaillance
organisationnelle: essai de mise en perspective historique – REE. Revue de l'électricité et
de l'électronique," 2008,.
[18] "Maintenanace Management Solution, Gestion d'ordr e de travail," RIMSES, [Online].
Available: https://www.rimses.be/fr/Gestion -d-ordres -de-travail -dans-GMAO. [Accessed
08 06 2019].
[19] R.Isermann, “Process fault detection based on modeling and estimation methods",
Automatica, vol. 20(4), 1984, pp. 387 -404, https://doi.org/10.1016/0005 -1098(84)90098 -0

44
CHAPTER III.
Theoretical Aspects Regarding Artificial Neural Network M odel for Predictive
Maintenance

III.1 . Introduction
The human brain is considered the most complex computational component known to man. The
multiple functionalities of the brain, such as thinking, saving, and problem solving, have inspired
many researchers who have attempted to create a computational model that matches the
functionality of the brain: they have obtained neural calculus.
So the neural networks are known devices of their powe r and their speed in the execution of the
operations they present many advantages compared to the more conventional computing method.
In theory, neural networks are data processing techniques essentially understood at present; they
should be part of the to olbox of all scientists who want to make the most of the data they have,
including running forecasts, designing predictive models, recognizing models or signals, and so
on.
The number of possible neural models is however very large, with more complex model s more
closely representing the function of human neurons. The artificial neural networks considered here,
and those that are generally used for systems and control, tend to be much simpler and easier to
physically perform. [1]
ANN meet several criteria of modern enterprise, including finding solutions quickly to increasingly
complex problems. In addition, ANN are robust, versatile and adaptable.
This technique is part of the design effort and focuses on the development of learn ing algorithms
to provide a system of autonomy and adaptive capacity. Sometimes these intelligent systems even
come to "discover" new solutions to complex and difficult to access problems for a human brain.
[2]
The purpose of thi s chapter is to explain under what circumstances neural networks are preferable
to other data processing techniques and for what purposes they may be useful after general
understanding.

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
45
III.2 . Basic Neural Networks

III.2.1 . Properties of the biological neuron
Neuron structure
A biological neural network is composed of a group or groups of chemically connected or
functionally associated neurons. A single neuron may be connected to many other neurons and the
total number of ne urons and connections in a network may be extensive.
We will present in this part the characteristics of this basic unit of the nervous system from the
biology to modeling (the artificial neuron). The human brain contains about 100 billion neurons.
They br eak down into three main regions : cell (or nerve cell), cell body (the soma) and nucleus.
The cell body analyzes and integrates the received information it ramifies to form what one names
the dendrites. It is through the dendrites that the information is c onveyed from outside to the soma,
body of the neuron. So the dendrites ensure the reception of the nerve impulse sent by the other
neurons. The information processed by the neuron then travels along the (unique) axon to be
transmitted to the other neurons. Transmission between two neurons is not direct. In fact, there is
an intercellular space of a few tens of angstroms (10 -9 m) between the axon of the afferent neuron
and the dendrites (a dendrite) of the afferent neuron. The junction between two neurons is called
the synapse (Figure 1). Synapses provide the connection between neurons via chemical
mechanisms of neurotransmitter exchange between dendrite branches and axon terminations of
other neurons [3].

Figure III. 1. Biological structure of neurons
dendrite
axon cell bod y

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
46
So it can be said that, a neuron can be schematized thus, it makes the sum of all the information
that it receives and it emits a signal provided that the sum is sufficiently high.
The need for information from outside is very necessary for brain learning [4], [5].

III.2.2 . the artificial neuron
Figure 2 shows the biological neuron versus artificial neuron. Each artificial neuron is an
elementary processor. It receives a variable number of inputs from upstream neurons. Each
elementary processor has a single output, which then branches to feed a va riable number of
downstream neurons, each connection is associated with a weight.

Figure III. 2. Biological Neuron / Artificial Neuron Mapping
What is an artificial neuron?
This is the basic element of a neural network.

Figure III. 3. The generic artificial neuron
Inputs: Directly system inputs that may coming from other neurons.
Core: Integrates all inputs and bias and calculates neuron output according to an activation function
that is often non -linear to give greater learning flexibility.
Output : Directly one of the outputs of the system or can be distributed to other neurons.

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
47
Bias: Input always 1 which allows to add flexibility to the network by allowing to vary the neuron
trigger threshold by adjusting the bias weight during learning.
Weight : Multiplying factors that affect the influence of each input on the output of the neuron.
The numerical value of the weight associated with a connection between two units reflects the
strength of the relationship between these two units. If this value is positive the connection is called
exciter, otherwise it is called inhibitory. [6]

III.2.3 . the neural network
What is a neural network?
A neural network is a network or circuit of neurons, or in a modern sense, an artificial neural
network, composed of artificial neurons or nodes. Thus a neural network is either a biological
neural network, made up of real biological neurons, or an artificial neural network, for solving
artificial intelligence (AI) problems. [7]
The Artificial Neural Network (ANN) is a method of artificial intelligence i nspired by the
biological structure – a brain. Similarly to this structure, the artificial neural network consists of a
body called processing element, inputs and outputs. The meaning of each entry is multiplied by the
weight and with bias, they go to the body of the cell, the treatment element. In the first step, the
multiplied inputs are summed by the summation function and in the second step they are
propagated by the transfer function to an output.
They are used to solve the following tasks: association , classification, grouping, pattern
recognition, image processing, control, optimization and modeling, these artificial networks may
be used also for predictive modeling, adaptive control and applications where they can be trained
via a dataset. Self -learn ing resulting from experience can occur within networks, which can derive
conclusions from a complex and seemingly unrelated set of information . [7]
Neural networks learn by adjusting the connection weights during training. Neurons are all -or-none
devices that “fire,” or transmit a message to the next connected neuron, based on meeting or
exceeding some threshold. [8]
The basic design of the artificial neuron is very simple. It consists of three basic types of layers: an
input layer, a hidden layer, and an output layer. The hidden layer can have one or more layers,
usually up to three depending on the complexity of the problem solved. The weighted signal is sent

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
48
to the hidden layers where it did most of the calculations. Finally, it is followed at the exit layer.
[9]

Figure III. 4. Structure of a neural network / Layers of artificial neural network
Artificial neural networks are a well -understood technique for data processing. These techniques
fit perfectly into control strategies. Indeed, they perform identification, control or fil tering
functions, and extend the traditional techniques of nonlinear automation to lead to more efficient
and robust solutions. They are highly connected networks of elementary processors operating in
parallel.
Each elementary processor calculates a single output based on the information it receives. Any
hierarchical network structure is obviously a network. [1] [10]
We can also say that artificial neural networks are mathematical models that reproduce in a simple
way the structure of biological neural networks, whose elements are called neurons and whose
connections between neurons are called links.
The principle of neural networks consists in using the information provided by health indicator s of
the machine or of the monitoring system, temperature, vibratory spectrum, pressure, flow …, in
order to perform recognition and identification of faulty modes or degraded states of this system.
Neural networks are used in predictive maintenance to e stimate and predict the trend of system
state degradation using system characteristic indicator information. [10]
Artificial neurons are processing units that rely on the principles of biological neuron functions. This means
that the artificial neuron receives information, elaborates on it, and then transmits further to other processing
units.
Connections between neurons, corresponding to biological synapses, are called weights and are
especially important for learning and other adaptation processes that occur in neural networks
(Figure III.5). According to biological principles, a neuron in the ANN receives one or more inputs

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
49
and transmits one output that is copied and forwarded to further units. The process of training an
ANN corresponds to the learning process in the human brain – the network draws conclusions of
future outcomes based on previous experience. Neurons (units) in the artificial network are
organized into layers. A basic ANN consists of an input, hidden, and o utput layer of neurons.
Increasing the number of hidden layers and changing the way that neurons are interconnected can
lead to greater complexity of neural networks. Input and output layers in the network serve to
represent the data that is fed into the n etwork and received as the network result, respectively,
whereas the units in hidden layers apply functions and perform calculations on the presented data.
It has already been accentuated that ANNs are nonlinear computational techniques. Prior to
analysis, data are scaled (most software tools use 0 to 1 scaling of the data), in order for different
input parameters to be comparable according to their influence on the outputs. Signals from input
and hidden layers are transferred to hidden and output layers th at are subsequently positioned.
Neurons in the hidden layer apply certain activation function to the transmitted signal that they
receive, and most commonly it is sigmoid function. In each neuron, input signals are [11] multipl ied
by their corresponding weights and converted to the

Figure III. 5. Simplified view of the neural network organization (all neurons are fully
connected)
Output by an activation (transformation) function in the following way:
𝒀𝒒= ∑ 𝑾𝒑,𝒒∗𝑿𝒑𝒛

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
50
Construction and testing the ANN
Now we have covered all the elements that ANNs contain, we can give an overview of the
construction process of the ANN (Figure III.6). The first step in this process is definition of ANN
architecture (topology); the number of neurons in the input, output, and hidden layers, as well as
the number of hidden layers. The number of neurons in the input and output layers are determined
by the studied proble m and it has already been accentuated that the number of neurons in the input
layer ( the number of input variables) should be kept to its minimum. The most difficult task at this
stage is selection of the number of hidden layers and the number of neurons in the hidden layer.
The trial and error approach is still commonly used to resolve th is issue, as well as various
optimization techniques. Once the network architecture (topology) is defined, it is often
necessary to select the type of transfer (activation) function, learning rate, smooth factor, etc.
Some of these paramete rs are predefined in the software used, and some need optimization.
We should always bear in mind the difference between the ANN model (architecture, topology,
and arrangement) and the ANN algorithm (method used for computation of outputs based on
inputs).

Figure III. 6. The most important steps in ANN construction and testing

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
51
The next step is the actual training of the network. The usual practice is to divide data into three
sets: training, testing, and validation. Training and test data sets are presented to the network first,
during the training process. Training data are used to define and optimize neuron weights, whereas
test data are kept aside by the network and used periodically to check predictive ability of the
network (during the training process). Training data should cover as much as possible of the data
variability. As mentioned previously, once the test error stops decreasing or starts to rise, it
is indicative of the network over fitti ng and the training process should be stopped.
If network over fitting is likely to occur, it is advisable to reduce the number of hidden layers or
neurons in those hidden layers.
Validation data are completely kept aside during the training process and are used only when the
training process is complete. Validation data contain samples that were not presented to the
network during the training process, but it is important to note t hat these samples should be within
the data space defined by the training data. Division of data in these three sets (training, testing,
and validation) is not an easy task and there are no strict rules on this issue, for example that 65%
of data should be used for training, 25% for testing, and 10% for validation . [12]
Developed ANNs are often tested using the cross – validation approach. This means that the whole
data set is divided into equal sized subsets. The network is then trained the number of times that is
equal to the number of subsets. In the process of validation, data set values predicted by the network
are compared to those experimentally obtained, and usually correlation coefficient R is calculated
to check the appr opriateness of the prediction .
Where experimentally obtained x values are compared to predicted (by the network) y values.
ANNs are, in general, used to predict outputs for the data not input into the network during its
training and testing. Therefore, it is an interpolation method. The prediction ability of the ANN is
restricted to space limit of input/output data presented to the model for training . Extrapolation
outside this data space should not be performed. This should be borne in mind when experiment s
are planned.
In order to obtain relevant and reliable results by using ANN models, it is recommended that the
number of experimental runs is 10 times greater in comparison to the number of inputs;
or, if this is not feasible, at least 2 to 3 times the number of inputs. It is, therefore, practical to

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
52
first conduct scre ening experimental design , in order to select the most significant input
variables that influence output properties, since t his leads to a reduction in the number of
the ANN inputs (and examples number of the needed for the network training at the same time).
There are many possibilities to optimize an ANN.
Limitation of the ANN
A considerable amount of data is often needed to build a reliable ANN model. Many
important steps in building and testing of the ANN are prone to errors and difficulties in optimal
determination. In addition, each software tool used is different and has its own distinctive
properties. It is always recommended to first thoroughly investigate the principles upon which
the neural networks models are built and tested in the software used . [13] [14] [8]
III. 2.4 . History
The preliminary theoretical base for contemporary neural networks was independently proposed
by Alexander Bain (1873) [15] and William James (1890) [16]. In their work, both thoughts and
body activity resulted from interactions among neurons within the brain.
For Bain [15], every activity led to the fir ing of a certain set of neurons. When activities were
repeated, the connections between those neurons strengthened. According to his theory, this
repetition was what led to the formation of memory. The general scientific community at the time
was skeptical of Bain’s theory [15] because it required what appeared to be an inordinate number
of neural connections within the brain. It is now apparent that the brain is exceedingly complex
and that the same brain “wiring” can handle multiple problems and inputs.
James’s theory [16] was similar to Bain’s [15], however, he suggested that memories and actions
resulted from electrical currents flowing among the neurons in the brain. His model, by focusing
on the flow of electrical currents, did not require individual n eural connections for each memory
or action.
C. S. Sherrington (1898) [17] conducted experiments to test James’s theory. He ran electrical
currents down the spinal cords of rats. However, instead of demonstrating an increase in electrical
current as projec ted by James, Sherrington found that the electrical current strength decreased as
the testing continued over time. Importantly, this work led to the discovery of the concept
of habituation.

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
53
McCulloch and Pitts (1943) [18] created a computational model for neural networks based on
mathematics and algorithms. They called this model threshold logic. The model paved the way for
neural network research to split into two distinct approaches. One approach focused on biological
processes in the brain and the other focused on the application of neural networks to artificial
intelligence .
In the late 1940s psychologist Donald Hebb [19] created a hypothesis of learning based on the
mechanism of neural plasticity that is now known as Hebbian learning. Hebbian learning is
considered to be a 'typical' unsupervised learning rule and its later variants were early models
for long term potentiation. These ideas started being applied to computational models in 1948
with Turing's B -type machines.
Farley and Clark (1954 ) [20] first used computational machines, then called calculators, to simulate
a Hebbian network at MIT. Other neural network computational machines were created by
Rochester, Holland, Habit, and Duda (1956) [21].
Rosenblatt (1958) [22] created the percept ron, an algorithm for pattern recognition based on a two –
layer learning computer network using simple addition and subtraction. With mathematical
notation, Rosenblatt also described circuitry not in the basic perceptron, such as the exclusive –
or circuit, a circuit whose mathematical computation could not be processed until after the back –
propagation algorithm was created by Werbos (1975) [23].
Neural network research stagnated after the publication of machine learning research by Marvin
Minsky and Seymour P apert (1969) [24]. They discovered two key issues with the computational
machines that processed neural networks. The first issue was that single -layer neural networks were
incapable of processing the exclusive -or circuit. The second significant issue was that computers
were not sophisticated enough to effectively handle the long run time required by large neural
networks. Neural network research slowed until computers achieved greater processing power.
Also key in later advances was the back -propagation algorithm which effectively solved the
exclusive -or problem (Werbos 1975) [23].
The parallel distributed processing of the mid -1980s became popular under the
name connectionism. The text by Rumelhart and McClelland (1986) [25] provided a full exposition
on the use of connectionism in computers to simulate neural processes.

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
54
Neural networks, as used in artificial intelligence, have traditionally been viewed as simplified
models of neural processing in the brain, even though the relation between this model and brain
biological architecture is debated, as it is not clear to what degree artificial neural networks mirror
brain function [26].
While initially research had been concerned mostly with the electrical characteristics of neurons, a
particularly important part of the investigation in recent years has been the exploration of the role
of neuromodulators such as dopamine, acetylcholine, and serotonin on behaviour and learning.
Biophysical models, such as BCM theory, have been important in understanding mechani sms
for synaptic plasticity, and have had applications in both computer science and neuroscience.
Research is ongoing in understanding the computational algorithms used in the brain, with some
recent biological evidence for radial basis networks and neural back -propagation as mechanisms
for processing data.
Computational devices have been created in CMOS for both biophysical simulation
and neuromorphic computing. More recent efforts show promise for creating nanodevices for very
large scale principal compon ents analyses and convolution [28]. If successful, these efforts could
usher in a new era of neural computing that is a step beyond digital computing [29], because it
depends on learning rather than programming and because it is fundamentally analog rather
than digital even though the first instantiations may in fact be with CMOS digital devices.
Between 2009 and 2012, the recurrent neural networks and deep feedforward neural
networks developed in the research group of Jürgen Schmidhuber at the Swiss AI Lab IDSIA have
won eight international competitions in pattern recognition and machine learning [30]. For
example, multi -dimensional long short term memory (LSTM) [31, 32] won three competitions in
connected handwriting recognition at the 2009 International C onference on Document Analysis
and Recognition (ICDAR), without any prior knowledge about the three different languages to be
learned.
Variants of the back -propagation algorithm as well as unsupervised methods by Geoff Hinton and
colleagues at the Universi ty of Toronto can be used to train deep, highly nonlinear neural
architectures [33], similar to the 1980 Neocognitron by Kunihiko Fukushima [34], and the
"standard architecture of vision" [35], inspired by the simple and complex cells identified by David
H. Hubel and Torsten Wiesel in the primary visual cortex.

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
55
Radial basis function and wavelet networks have also been introduced. These can be shown to offer
best approximation properties and have been applied in nonlinear system identification and
classifica tion applications [27].
Deep learning feedforward networks alternate convolutional layers and max -pooling layers,
topped by several pure classification layers. Fast GPU -based implementations of this approach
have won several pattern recognition contests, including the IJCNN 2011 Traffic Sign Recognition
Competition [36] and the ISBI 2012 Segmentation of Neuronal Structures in Electron Microscopy
Stacks challenge [37]. Such neural networks also were the first artificial pattern recognizers to
achieve human -competitive or even superhuman performance [38] on benchmarks such as traffic
sign recognition (IJCNN 2012), or the MNIST handwritten digits problem of Yann LeCun and
colleagues at NYU.

III. 2.5 . Fields of application of artificial neural networks
Today, artificial neural networks have many applications in a wide variety of sectors:
• Image processing: character and signature recognition, image compression, pattern recognition,
encryption, classif ication, etc.
• Signal processing: filtering, classification, source identification, speech processing … etc.
Control: process control, diagnostics, quality control, robot servoing, automatic guidance systems
for automobiles and planes … etc.
• Defense : missile guidance, target tracking, face recognition, radar, sonar, lidar, data compression,
noise suppression … etc.
• Optimization: planning, resource allocation, management and finances, etc.
• Simulation: simulation of the flight, simulation of blac k box, meteorological forecast, copy of
model … etc.

In particular, in the field of optimization, we can distinguish the application of ANN to perform
Predictive Maintenance. With the vast amount of time series data constantly being produced by
machines in factories and plants, such as sensor and control values, there is a lot of information
available to predict breakdowns of the machines.
There are different sensory systems from simple inductive capacitors to proximity sensors and
photoelectric or laser sensors and advanced vibration sensors to industrial systems such as thermo –

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
56
graphic cameras, vision systems, process control systems and measurement sensors capable of
capturing large amount of machine data available for further dev elopment. Artificial neural
networks show promising results as a robust tool for evaluating these data to support predictive
maintenance activities. Mainly Multilayer Perceptron’s (MLPs) are used for diagnosing bearing
defects, induction motors, non -destru ctive evaluation of performance and degradation of check
valves and in real -time robotic systems.

III. 2.6. Architecture of Neural Networks
The architecture of neural networks consists of a network of nonlinear information processing
elements that are nor mally arranged in layers and executed in parallel. This layered arrangement
for the network is referred to as the topology of a neural network. These nonlinear information
processing elements in the network are defined as neurons, and the interconnections between these
neurons in the network are called synapse or weights. A learning algorithm must be used to train a
neural network so that it can process information in a useful and meaningful way. [14]
Depending on the type of connections (architectures), the ANN are grouped into two categories
(see Figure III.7).

Figure III. 7. Different types of unconnected and recurrent networks
Feedforward Network
In the following example, the inputs represent values (yes/no) of four of the features used to
determine if a patient who has suffered an animal bite needs to be treated prophylact ically for

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
57
rabies. Each connection is weighted and these weights evolve during training. This simple model
is similar to the perceptron in Figure III. 7, except that it includes a hidden layer and that is
propagates error back to the input weights during tr aining. The hidden layer provides additional
knowledge representation internal to the system, and can provide improvements in classification
performance [39, 40]

Figure III. 8. A feed -forward, back -propagation neural network
This network uses the feed -forward approach of the perceptron, but includes the ability to provide
(propagate) feedback about the accuracy of its output to the components.

Recurrent Networks (Feedback)
The architecture of neural connections can be described as a combinational feed forward network.
In order to insert context, as well as to provide the possibility for feedback self -correction, some
networks add a form of state feedback called back propagat ion. Once a feedback system is in place,
the possibility of machine learning is present.
In the process of machine learning, system behavior and processing are altered based on the degree
of approximation achieved for any specified goal. In today’s systems , the human programmer sets
the specified goals. Applied system feedback, however, allows the AI system to develop alternative
approaches to attaining the set goals.
The implementation of feedback systems can be on a real -time. An artificially produced mac hine-
style implementation of synapses and neurons (neural network) can be set up to operate in the
present state, next state, or via synaptic weightings implemented as matrices. Computer
programming languages and systems have been designed to facilitate th e development of machine
learning for artificial neural network systems.
Recurrent and r eentrant neural networks implement connections from lower layers of neuron into
upper layers of the network of neurons. Such a configuration can be used to provide a

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
58
temporal/sequential dimension to neural network processing. From a control theory point of view,
these connections can be envisioned as feedback.

Figure III. 9. Partially recurrent neural network connections

Figure III. 10. Interconnexion structure or a feedback network model

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
59
III.3 . Conclusion

Neural networks have been developed as an attempt to simulate the highly connected biological
system found in the brain through the use of computer hardware and/or software [41].
Artificial neural networks show great potential in industrial applications, especially in predictive
maintenance tasks. Due to the need for well-trained staff for ANN calculations, the need for special
software and the availability of sensors capable of systematically capturing and storing collected
data, it should be properly assessed whether the use in production real will be meaningful and
profitable. The advantage of these methods is the potential for effective prevention of equipment
failure.
Neural network models allow us to bridge the gap between the behavioral and neuronal level. By
integrating data from different domains into one congl omerate model, we might start to see the
‘bigger picture’. For this approach to be successful, it must stay close to empirical data and provide
concrete predictions which have to be tested experimentally to possibly refine the model. Neural
network models provide a more disciplined approach that is grounded by mathematics and allows
exploration of more complex dynamics.
Neural networks pose several advantages for data mining. They have shown excellent performance
in many settings, particularly in prediction , where the trained network can be used to identify the
probability of a specific outcome, However, neural networks pose several challenges to those using
them for data mining. Most neural network software requires that the architecture be specified in
advance, so that the determination of how many hidden layers (and how many neurons are in those
layers) must be made in advance of training. [8]
Finally, various applications will be described to illustrate the possible defects wh ere neural
networks can provide efficient and elegant solutions to technical problems.

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
60
References

[1] A. ,. G. Dreyfus, ‘Réseaux de neurones : méthodologie et applications’, 2002.
[2] Andrei Doncescu, Les réseaux de neurones artificiels.
[3] M. Parizeau, « Réseaux de neurones » Notes de cours (GIF -21140 et GIF -64326),
Université Laval, Ed., Automne 2 004.
[4] A. Mestan, « Introduction aux Réseaux de Neurones Artificiels Feed Forward », 1 janvier
2008.
[5] P. Wira and Habilitation à Diriger des Recherches, « Approches neuromimétiques pour
l’identification et la commande », Université de Haute Alsace, Ed., Novembre 2009.
[6] L.Personnaz, I.Rivals, .,, ‘ Réseaux de neurones formels pour la modélisation, la
commande, et la classification’, C. éditions, Ed., collection Sciences et Techniques de
l’Ingénieur, ., 2003.
[7] J. J. Hopfield, "Neural networks and physical systems with emergent collective
computational abilities", Proc. Natl. Acad. Sci. U.S.A. 79 ( 8), (1982), p. 2554 –2558. .
[8] John H. Holmes, “Knowledge Discovery in Biomedical Data: Theory and Methods” in
Methods in Biomedical Informatics, 2014.

[9] ""Neural Net or Neural Network – Gartner IT Glossary".," [Online]. Available:
www.gartner.com..
[10] ,. S. Haykin, « Neural Networks : A comprehensive foundation », S. Edition and P. P. H.
PTR, Eds., Delhi 110 092, India: by Pearson Education,Inc, Indian Branch, 482
F.I.E.patparganj,, 1999, p. page 23.
[11] Neural computing in p harmaceutical products and process development Published by
Woodhead Publishing Limited,, 2013.
[12] C. ,. Looney, ‘ Advances in feedforward neural networks: demystifying knowledge
acquiring black boxes ’,, IEEE T Knowl. Data En., ( 1996 ) .

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
61
[13] Djuris, J., Ibric, S., & Djuric, Z., Neural computing in pharmaceutical products and process
development., Computer -Aided Applications in Pharmaceutical Technology,, (2013), p.
91–175..
[14]

[15]

[16]
[17]

[18]

[19]

[20]

[21]

[22]

[23]

[24]

[25]

[26]
Yi-Ping Phoebe Chen,Paolo Carloni, , “ Modern Methods in Natural products chemistry”,
in Comprehensive Natural Products II, , 2010.
Bain (1873). Mind and Body: The Theories of Their Relation. New York: D. Appleton and
Company.
James (1890). The Prin ciples of Psychology. New York: H. Holt and Company.
Sherrington, C.S. (1898). "Experiments in Examination of the Peripheral Distribution of
the Fibers of the Posterior Roots of Some Spinal Nerves". Proceedings of the Royal
Society of London. 190: 45 –186. doi:10.1098/rstb.1898.0002.
McCulloch, Warren; Walter Pitts (1943). "A Logical Calculus of Ideas Immanent in
Nervous Activity". Bulletin of Mathematical Biophysics. 5 (4): 115 –
133. doi:10.1007/BF02478259.
Hebb, Donald (1949). The Organization of Behavior. New York: Wiley.
Farley, B.; W.A. Clark (1954). "Simulation of Self -Organizing Systems by Digital
Computer". IRE Transactions on Information Theory. 4 (4): 76 –
84. doi:10.1109/TIT.1954.1057468.
Rochester, N.; J.H. Holland, L.H. Habit and W.L. Duda (1956). "Tests on a cell assembly
theory of the action of the brain, using a large digital computer". IRE Transactions on
Information Theory. 2 (3): 80 –93. doi:10.1109/TIT.1956.1056810.
Rosenblatt, F. (1958). "The Perceptron: A Probalistic Model For Inform ation Storage And
Organization In The Brain". Psychological Review. 65 (6): 386 –
408. CiteSeerX 10.1.1.588.3775 . doi:10.1037/h0042519. PMID 13602029.
Werbos, P.J. (1975). Beyond Regression: New Tools for Prediction and Analysis in the
Behavioral Sciences.
Minsky, M.; S. Papert (1969). An Introduction to Computational Geometry. MIT
Press. ISBN 978-0-262-63022 -1.
Rumelhart, D.E.; James McClelland (1986). Parallel Distributed Processing: Explorations
in the Microstructure of Cognition. Cambridge: MIT Press.
Russell, Ingrid. "Neural Networks Module" . Retrieved 2012

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
62
[27]

[28]

[29]

[30]

[31]

[32]

[33]

[34]

[35]

[36]

[37]

[38] Billings, S. A. (2013). Nonlinear System Identification: NARMAX Methods in the Time,
Frequency, and Spatio -Temporal Domains. Wiley. ISBN 978-1-119-94359 -4.
Yang, J. J.; et al. (2008). "Memristive sw itching mechanism for metal/oxide/metal
nanodevices". Nat. Nanotechnol. 3 (7): 429 –433, doi: 10.1038/nnano.2008.160,
PMID 18654568.
Strukov, D. B.; et al. (2008). "The missing memristor found". Nature. 453 (7191): 80 –
83. doi:10.1038/nature06932. PMID 1845 1858.
Kurzweil AI Interview with Jürgen Schmidhuber on the eight competitions won by his
Deep Learning team 2009 –2012", 2012.
Graves, Alex; Schmidhuber, Jürgen (2008). "Offline Handwriting Recognition with
Multidimensional Recurrent Neural Networks". In Be ngio, Yoshua; Schuurmans, Dale;
Lafferty, John; Williams, Chris K. I.; Culotta, Aron (eds.). Advances in Neural
Information Processing Systems 21 (NIPS'21). Neural Information Processing Systems
(NIPS) Foundation. pp. 545–552.
Graves, A.; Liwicki, M.; Fern andez, S.; Bertolami, R.; Bunke, H.; Schmidhuber, J. (2009).
"A Novel Connectionist System for Improved Unconstrained Handwriting
Recognition". IEEE Transactions on Pattern Analysis and Machine Intelligence. 31 (5):
855–868. CiteSeerX 10.1.1.139.4502 . doi:10.1109/TPAMI.2008.137. PMID 19299860.
Hinton, G. E.; Osindero, S.; Teh, Y. (2006). "A fast learning algorithm for deep belief
nets" (PDF) . Neural Computation. 18 (7): 1527 –
1554. CiteSeerX 10.1.1.76.1541 . doi:10.1162/neco.2006.18.7.1527. PMID 16764513.
Fukushima, K. (1980). "Neocognitron: A self -organizing neural network model for a
mechanism of pattern recognition unaffected by shift in position". Biological
Cybernetics. 36(4): 93 –202. doi:10.1007/BF00344251. PMID 7370364.
Riesenhuber, M.; Poggio, T. (1999). "Hierarchical models of object recognition in
cortex". Nature Neuroscience. 2 (11): 1019 –1025, doi:10.1038/14819. PMID 10526343.
D. C. Ciresan, U. Meier, J. Masci, J. Schmidhuber. Multi -Column Deep Neural Network
for Traffic Sign Classification. Neural Networks, 2012.
D. Ciresan, A. Giusti, L. Gambardella, J. Schmidhuber. Deep Neural Networks Segment
Neuronal Membranes in Electron Microscopy Image s. In Advances in Neural Information
Processing Systems (NIPS 2012), Lake Tahoe, 2012.
D. C. Ciresan, U. Meier, J. Schmidhuber. Multi -column Deep Neural Networks for Image
Classification. IEEE Conf. on Computer Vision and Pattern Recognition CVPR 2012.

Ch III. Theoretical Aspects Regarding Artificial Neural Network Model for Predictive M aintenance
63
[39] I. Russell, "Neural Networks Module"., 2012.
[40] J. Y.Ziquel, ‘ Apprentissage par Réseaux de neurones’ Rapport de projet,.
[41] Tracy L. Faber, … Ernest V. Garcia,, in Clinical Nuclear Cardiology , , Quantification, and
Display, (Fourth Edit ion) ed., 2010 SPECT processing.

64
CHAPTER IV.
Development of Predictive M aintenance Model Using the Neural Network
Algorithm, for a Grinding Machine -Tool
IV.1. Introduction
In the field of optimization, we can distinguish the application of ANN to perform Predictive
Maintenance Model . With the vast amount of time series data constantly being produced by
machines in factories and plants, such as sensor and control values, there is a lot of information
available to predict breakdowns of the machines.
Artificial neural networks show promising results as a robust tool for evaluating these data to
support predictive maint enance activities.
The objective of this chapter is to develop a predictive maintenance model using the neural
network algorithm for a grinding machi ne tool f rom Bârlad-Romania bearing manufacturing
sector , such as grinding machine tool with biggest number of interruptions due to unplanned
maintenance actions .
From a database dating back to 2019, we tried to develop a neural model. For this, we worked on
Easy NN and NN MODEL using the defect code as the output variable , comparing the results of
the relative average errors in order to make the best choice to build a neural network model.
Objective:
The objective of my study is to increase the performances of the maintenance model used in the
manufacturing comp any URB Group – Rulmenti S.A. Bâ rlad.

IV.2. Why predictive maintenance model?

Studying the data base obtained from the manufacturing company we noticed a big difference
between the moment of defect detection and the moment when the reparation (maintenance action)
become. This difference is due to the existing maintenance model – the preventive maintenance
model – in the company , determining the intervention of maintenance service only at the
programmed moment . Under these conditions a large number of interruptions occur, due to

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
65
unpredicted maintenance actions determin ing high cost in production. It result the necessity to use
a combined maintenance model including the predictive component.

The work consists in developing, testing and validating in a manufacturing company a predictive
maintenance model for a grinding machine -tools , using the neural network algorithm .

IV.3. Grinding process presentation

Grinding is an abrasive machining process that uses a grinding wheel as the cutting tool. It is
performed in grinding machine [1]. Although there exists quite a few differences between
machining and grinding, grinding is also considered as one machining process that can produce
smoother surface .
The similarities between machining and grinding are [2]:

• Both machining and grinding processes follow the basic principal of subtractive
manufacturing approach. Here layers of excess material are removed from a solid 3 -D
blank to obtain final product. On the contrary, in additive manufacturing approach, layer
by layer material is added one over another to build a solid 3 -D product.
• In both the machining and grinding processes, material removal takes place in the form of
solid chips.
• In both the cases material is removed by shearing (ploughing and rubbing occur in grinding
but they do not remove material).
• Presence of cutter is mandatory in b oth the cases for realizing material removal.
• Only mechanical energy is utilized for material removal in both the cases (unlike NTM
processes where various forms of energy like mechanical, electrical, thermal, chemical,
etc. are used to remove material).
• Differences between machining and grinding [2], [3]:

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
66
Machining Grinding
Machining is one bulk material removal
process (i.e. high MRR). Thus it is economical
and suitable to give proper size and also for
semi -finishing. Grinding has low material removal rate and is
preferred only for finishing.
Accuracy and tolerance achieved by
conventional machining operations are not so
good. Achieving tolerance below 50µm is very
difficult. It can provide better accuracy and tolerance.
In grinding, achievable tolerance can be as
low as 2µm.
In machining, cutting tool is usually made of
metals or alloys, which is substantially harder
than work material. However, ceramic,
diamond and cBN tools (non -metallic) are also
available, usually in the form of inserts. Cutting tool for grinding, i.e. the wheel, is
made of abrasive materials (such as alumina,
silica, etc.) bonded in harder medium (like
resin, metal, etc.).
Cutting tool used in machining has specific
geometry. Values of the geometrical features
may vary from one tool to another, but each
tool has pre -defined geometry as per tool
signature. Grinding wheel may have specific pre -defined
features, but the abrasive grits (which actual ly
participate in material removal) have random
geometry and orientation.
Rake angle of a cutting tool may vary from
positive to negative. However, the negative
rake usually does not go above ( –15ș) as it
may severely degrade machinability. Abrasive grits have haphazard rake angle. A
much wider variation of rake angle from
(+75ș) to ( –75ș) is noticed among abrasive
grits.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
67
Machining Grinding
During machining operation, each main
cutting edge of the tool actively and mostly
equally participates in material remov al action. Only few (about 1%) abrasive grits actually
engage in material removal action (shearing).
Some grits just engage in rubbing, scratching
and ploughing. A major percentage of grits do
not even touch the work material in every
rotation.
Generated cutting temperature is
comparatively low, and only a tiny portion (5 –
8%) of it diffuses into the workpiece. Thus
thermal damage of the machined surface is
usually insignificant. Severe heat is generated during grinding, and
also a substantial amo unt of heat diffuses into
the workpiece. This causes thermal damage of
the machined (ground) surface such as
changing hardness.
The maximum cutting speed (rpm) used in
conventional machining is typically limited to
2000 rpm (limited by the capability of machine
tool, especially gears and bearings). Thus
cutting velocity (m/min) is also lower. Rotational speed of grinding wheel is much
higher (2000 – 4000 rpm). Ultra -high speed
grinding (speed around 20,000 rpm) is also
carried out for some s pecific applications.
Specific energy consumption (power per unit
volume of material removed in kW/mm3) is
much lower. Specific energy consumption is 5 –50 times
higher than that of the machining.
There exist some surface hardened and
inherently hard materials that cannot be
machined by conventional methods. Such materials can be finished by grinding
without appreciable problem.
Table IV. 1. Difference between machining and grinding

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
68
IV.4. Artificial Neural Network Modelization

The initial data base represents the maintenance data base from Rulmenti Bârlad for the period
2016 -2019 ; this contains full information about the maintenance: grinding machine name (30
machines), type of defects, moment of defect detection, moment of maintenance intervention &
all.
IV.4.1. Steps in the building of data base to use for ANN mode lling
The data base obtained from the manufacturing company has the form presented in the table 2.

Table IV. 2. Initial data base
Codification and time calculation
We select only the grinding machine -tools – SIW3, SIW4 and SIW5 – from this data base. In
addition to the initial data base we introduce supplementary columns (in red) corresponding to the
machine -tool code , defect code and time, measured against a zero mo ment (1.03.2019, 12:30),
when the defect occurred . It result the table 3.
Note:
-The Grinding machines tools work 5 days a week.
-16 hours (2 cycle of 8 hours) per day.
The use of this information is necessary to calculate the time (column "I").
The industrialist's suggestion was to consider only the year 2019 of the database.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
69

Table IV. 3. New data base
Grinding machine -tools codification i s presented in the table below:

SIW3 SIW4 SIW5
Nr. Inv. MT MT
CODE Nr. Inv. MT MT
CODE Nr. Inv. MT MT
CODE
200588 SIW3 31 200658 SIW4 41 200635 SIW5 50
200640 SIW3 32 200704 SIW4 42 200636 SIW5 51
200615 SIW3 33 200703 SIW4 43 200735 SIW5 52
200616 SIW3 34 200702 SIW4 44 200738 SIW5 53
200760 SIW3 35 200657 SIW4 46 200739 SIW5 54
200580 SIW3 36 200729 SIW4 47 200744 SIW5 55
200730 SIW4 48 200749 SIW5 56
200731 SIW4 49 200757 SIW5 57
200733 SIW4 410 200758 SIW5 58
200745 SIW4 411 200760 SIW5 59
203154 SIW5 510

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
70
46844 SIW5 511
Table IV. 4. Machine Tools codification
The defect codification is performed considering the description of the defect in the data base.
Defect description Defect code
Lack of oil 1
Loss of oil, water, air (cause corrections:
changed hose, adjustable pump pressure …) 2
transmitter 3
general mechanical defect 4
electrical defect disp., contactor, leap 5
The electrical 6
electrical / electronic fault – advance 7
electrical safety 8
general electrical / electronic fault 9
electrical fault – droser 10
mechanical failure – device strap 11
mechanical failure – mass inclination screw 12
diamantare 13
gaskets 14
Table IV. 5. Defect code s

Selection of the grinding machine tool and maintenance frequency calculation

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
71
From this new database, only SIW 5 machines were selected to build a neural network model and
algorithm able to predict the interruptions of these machines. After this selection we could see that
30 example of the database appear, it is the data that will use it for training, testing and validation
of the neuronal network model.
Note
The maintenance data of grinding machine tool, code S6 , only contains mechanical, electrical and
hydraulic defects.
In addition to all information we perform the calculation of the time between two successive
apparitions of the same defect (maintenance frequency) . The maintenance frequency , F represents
the difference between two successive maintenance actions determined by the same defect
By example if we take the date 8/1/2019 at 10:30 where the moment of the first repair and the
14/01/2019 at 12:00 when the time of the second repair starts, in this case we have find 4 days and
1.5 hours for the defect with code 1 (lack of oil).
Finally, the data base to use for neural network model building is presented in the table below :

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
72

Table IV. 6. Maintenance data base

No of
experimDefect
codeReaction
time
[hours]Maintenance
duration
[hours]Maintenance
freqvency
[hours]
1 1 1.35 0.50 65.00
2 1 1.15 0.50 65.50
3 1 0.17 0.50 66.16
4 2 1.00 0.50 256.83
5 2 1.10 1.00 151.50
6 2 0.85 1.00 33.00
7 1 1.50 0.50 149.66
8 2 2.00 1.00 16.00
9 1 1.50 0.50 16.00
10 2 0.90 1.00 182.50
11 1 1.00 0.33 100.83
12 12 0.10 1.00 48.48
13 1 1.60 0.50 132.50
14 1 0.80 1.00 308.50
15 1 1.00 0.50 0.00
16 14 0.25 3.50 16.75
17 1 1.40 0.50 81.00
18 2 0.17 2.00 102.00
19 2 0.25 0.67 210.00
20 11 0.25 2.00 166.00
21 1 1.20 0.50 69.00
22 2 0.50 1.50 439.50
23 1 1.20 0.50 98.50
24 12 0.30 0.50 5.50
25 1 0.80 0.50 46.00
26 2 1.25 2.00 0.00
27 1 1.20 0.50 0.00
28 13 0.50 1.00 464.00
29 2 0.25 0.50 0.00
30 2 0.15 1.00 0.00
AVERAGE 0.86 0.92 109.69
STDEVP 0.52 0.67 120.74

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
73
IV.4.2. Identification of input and output variables
Considering the final resulted data base, we work on 4 variables:
✓ The first variable is the code of defect ( determining the interruption) .
✓ The second variable is the reaction time [hours] (see the database).
T2-T1: Where T2 is the m oment when the repair begins. [d ate/hours]
T1 is the time of appearance of defect [date / hours]
✓ The third variable is the maintenance duration ( repair tim e, T3-T2).
Where : T3 represents the time when the repair will be finished.
✓ For the fourth variable we calculate the frequency of maintenance for each data example
(30 datasets) . This information was used for the entire data set to finally construct a
maintenance table containing these variables (Table IV.6) .
From this table it was possible to select neuronal model input and output variables for predictive
maintenance. In our case the input and output variable was reaction time, maintenance duration
and maintenance frequency , respectively the defect code.
EASY NN is used to introduce the variables of the maintenance table from a file of type Text (Tab
delimited).
We chos e the parameters recommended by the software developer with modification of some
parameters like the number of learning cycles and the target error.
From the maintenance data base , we divided our work into two phases. Firstly, all the data was
used for training, learning, and network interrogation.
For the second phase the data set was divided into two:
✓ 25 data for network training and interrogation
✓ 5 data for validation and network query ; the data sets predicted by the networ k will be
compared with thos e experimentally obtained; a correlation coefficient, R will be
calculated to check the appropriateness of the prediction.
IV.4.3. Comparison between EASY NN and NN MODEL

In the beginning I use NN MODEL and EASY NN in the same time and I compare the results,
and the best result are given by EASY NN, so I decided to continue using only EASY NN.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
74

Figure IV. 1. Comparison between NN MODEL and EASY NN

-Easy NN give us better results than NN MODEL , so I decided to use only Easy NN
The x axis: represent the number of experimental data.
The y axis: represent respectively the experimental values in blue of DC and predicted values in
red.
(See annex 4 to know for the results of the predicted values usi ng NN MODEL )

123456789101112131415161718192021222324252627282930
EXP DC 11122212121121111412211121121211322
ANN DC 111222121111211114121121211212113220246810121416EXP/ANN
Nr of EXPComparaison prredictive values and experimental values using Easy
NN
EXP DC ANN DC
123456789101112131415161718192021222324252627282930
EXP DC 11122212121121111412211121121211322
ANN DC 9999999999999999999999999999990246810121416EXP/ANN
Nr of EXPComparaison predictive values and experimental values using NN
MODEL
EXP DC ANN DC

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
75
IV.4.4. Phases 1: Training 30 data sets
Training Process
The training data is used to define, test and optimize the neuron weights, kept aside by the network
and to check the predictive ability of the network (during the training process). The learning data
should cover as much as possible the variability of the data. As mentioned earlier, once the training
error stops decreasing or starts to increase, it indicates that the network has been too narrow and
the training process needs to be stopped.
Easy NN -Plus i s another standalone commercial program for trainin g and evaluating neural
network, Network inputs and outputs are specified in a grid layout that is somewhat akin to data
entry in a spreadsheet. The number of columns of input and output implicitly declare the number
of input and output neurons, respectively.
• Once the data are entered in to the grid, the specifics of any hidden network layers can be
selected via a dialog box. Finally, training parameters are set in yet another dialog box and
the training begins.
• Once training is finished, results can be viewed in several different ways wit hin Easy NN-
Plus.
• A graph of the network’s error during training and a plot of output values vs. target values
can be displayed within the main interface. Also, a diagram of the resulting network can
be generated, including inter -node connections drawn in proportion to their weights Easy
NN-Plus is fairly simple and intuitive (despite its dearth of instructions) to use but like
several other ANN training programs mentioned here, its standalone nature and lack of
portability make it an unlikely choice for re searchers who hope to collaborate. Further
restricting its flexibility, Easy NN-Plus currently only works on systems with the Microsoft
Windows operating system installed.
• Easy NN can be used to create, control, train, validate and query neural networks.
The learning process will start and it will stop automatically, when the target error is
reached.
Working with Easy NN consists in the following steps:
• The ANN is tested for authenticity, by
Action check Grid; a message will give permission to create a network
Action New Network.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
76
• The t raining involves importing the ‗train dataset‗
• The Neural Network is then generated and various parameters are defined,
• The training begins once the controls are set and is terminated once the criterion is met.
Controls

Figure IV. 2. Controls

I modified some control parameters like cycle before first validating and the target error.
I use the values recommended by the software developer.
Neural network building
• To construct this neural network a data base with 30 rows and 5 columns was used.
• The first column indicates the number of experiments or examples.
• For the other columns, respectively, the inputs and the network output have been specified.
• The number of the Input is of three variables that are respectively:
– Reaction time [hours ].
-Maintenance duration [hours].
-Maintenance frequency [hours].
• The Defect Code was considered as an output variable. (as represents the Figure IV.3 )

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
77

Figure IV. 3. Input/output Data Entry with Easy NN -Plus

• These examples are used for neural network training, learning and interrogating.
• The goal is to achieve an average relative error of less than 5%.
• The aim is to create a neural network that can be trained and validated using the
manufacturing company data.
• Once the neural network has been formed and validated, it can be used to predict the defects
that are likely to be encountered for grinding machines tools.
• Once training is finished, results can be viewed in several different ways within Easy NN-
Plus.
• A diagram of the resulting network can be generated, including inter -node connections
drawn in proportion to their weights (See Figure) .

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
78

Figure IV. 4. Graphic representation of a Neural Network in Easy NN -Plus

Figure IV. 5. Neural model architecture

The concept is very simple based in some input parameters as we know that between the input and
output some dependence exist I used ANN to find however variation of input influences the output .

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
79
The proposed neural model consist in 3 layers .
The first layer also called input layer has 3 neurons corresponding to the input variables, the second
layer called hidden layer with 4 neurons and the layer with 1 neurons also called output layer.
Process learn ing p rogress
• A graph of the network’s error during training and a plot of output values vs. target values
can be displayed within the main interface.
• The learning process will start and it will stop automatically, when the target error is
reached.
• Learning rate selected : 0,6
• Stop when average error is: 0,0001
• Stop learning : 253234 Cycles

Figure IV. 6. Learning progress

The Learning Progress view shows how learning is progressing. This is sufficient for over 253234
learning cycles. The graph is produced by sampling these points. The horizontal axis is nonlinear
to allow the whole learning progress to be displayed. As more cycles are executed the graph is
squashed to the left. The scaled errors for all example rows are used. The red line is the maximum
example error, the blue line is the minimum example error and the green line is the average
example error . The orange line is the average validating error. The errors are also shown on the
Example errors view.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
80
So we can see that the maximum average relative error has been decreased after 250000 learning
cycles.
Comment:
Learning progress graph shows the maxim um, average and minimum training error. The graph is
squashed to the left as learning progresses so that the whole learning curve is visible. Up to 253234
learning cycles can be displayed on the graph.
Predictions

Figure IV. 7. Prediction
The Predictions view has two boxes that show how close the predicted value for each example
output is to the true value. The training examples are shown in separate boxes. Predicted outputs
for training examples g et closer to the true values as training progresses. If the predicted values
are very close to the true values then the dots will be on the diagonal line.
This dispersion of predicted values is a good dispersion.
Column Values
The Column Values view. All Input/output column values are shown graphically. The left hand
vertical scale shows the scaled value from 0 to 1. The right hand scale shows the real value
calculated using the highest and lowest values in the column. The blue line is the true value, the
red line is the running average and the black dotted line is the overall trend. The lavender line is
the Risk number. This is the number of times that any forecasted value had to be adjusted before
the network produced an error close to the error produ ced before the forecast.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
81

Figure IV. 8. Column values
Example errors

Figure IV. 9. Example errors

The Example Errors view shows the absolute error and the relative error produced by up to 30
examples in descending order from the greatest error. The order can be switched to ascending from
the smallest error by clicking on the sort order toolbar icon.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
82
Evaluation of neural network performance

Table IV. 7. Evaluation neural model performance
Starting from the table IV.6 of the maintenance data we add two columns :
– The first column to add (New ANN) contains the predicted values for each example.
– The second column add is the average relative error.
– After calculating the average relative training error.
REA T MIN D MIN F EXP DC ANN DC REL ERR
EXP 01 1.35 0.50 65.00 1 1 0
EXP 02 1.15 0.50 65.50 1 1 0
EXP 03 0.17 0.50 66.16 1 1 0
EXP 04 1.00 0.50 256.83 2 2 0
EXP 05 1.10 1.00 151.50 2 2 0
EXP 06 0.85 1.00 33.00 2 2 0
EXP 07 1.50 0.50 149.66 1 1 0
EXP 08 2.00 1.00 16.00 2 2 0
EXP 09 1.50 0.50 16.00 1 1 0
EXP 10 0.90 1.00 182.50 2 1 50
EXP 11 1.00 0.33 100.83 1 1 0
EXP 12 0.10 1.00 48.48 12 12 0
EXP 13 1.60 0.50 132.50 1 1 0
EXP 14 0.80 1.00 308.50 1 1 0
EXP 15 1.00 0.50 0.00 1 1 0
EXP 16 0.25 3.50 16.75 14 14 0
EXP 17 1.40 0.50 81.00 1 1 0
EXP 18 0.17 2.00 102.00 2 2 0
EXP 19 0.25 0.67 210.00 2 1 50
EXP 20 0.25 2.00 166.00 11 129.090909
EXP 21 1.20 0.50 69.00 1 1 0
EXP 22 0.50 1.50 439.50 2 2 0
EXP 23 1.20 0.50 98.50 1 1 0
EXP 24 0.30 0.50 5.50 12 12 0
EXP 25 0.80 0.50 46.00 1 1 0
EXP 26 1.25 2.00 0.00 2 2 0
EXP 27 1.20 0.50 0.00 1 1 0
EXP 28 0.50 1.00 464.00 13 13 0
EXP 29 0.25 0.50 0.00 2 2 0
EXP 30 0.15 1.00 0.00 2 2 0
AVE.ERR 3.636364
STDEV.P 12.49793

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
83
– The standard deviation has also been calculated.
We notice that the average error is Ꜫ= 3.63% less than 5% so we can use artificial neural
networks for the prediction of defects.
Input i mportance

Figure IV. 10. Input importance

The Input Importance view shows the importance and the relative importance of each Input
column. The Importance is the sum of the absolute weights of the connections from the input node
to all the nodes in the first hidden layer. The inputs are shown in the descending order of
importance from the most important input.
Sensitivity

Figure IV. 11. Sensitivity

The Sensitivity view shows how much an output changes when the inputs are changed. The inputs
are all set to the median values and then each in turn is increased from the lowest value to the
highest value. The change in the output is measured as each input is increased from lowest to
highest to establish the sensitivity to change. Insensitive inputs are not shown. The inputs are
shown in the descending order of sensitivity from the most sensitive input.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
84
IV.4.5. Phases 2: Training 25 data sets and validation with 5 data sets

Training and validating process
The training data is used to define, test and optimize the neuron weights, kept aside by the network
and to check the predictive ability of the network (during the training and validating process). The
learning data should cover as much as possible the variability of the data. As mentioned before,
once the training error thus validating stops decreasing or starting to increase, it indicates that the
network has been too close and that the training process needs to be stopped. We used 25 for
training and 5 for validating .
Controls

Figure IV. 12. Controls parameters for 2 -nd phase

We made only one change from the previous phase we can see that in the control parameters.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
85
Table IN/OT
This time I use 25 examples for training and 5 examples for validating .

Figure IV. 13. Table of IN/OT
Process learning progress
This graphs shows the error as a function of learning cycles we can see the maximum training error
decrease. And we can see also the average validating error equal to 0 and it is a good results.
Learning cycles: 201701
Learning rate: 0.6

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
86

Figure IV. 14. Process learning progress
Predictions
The images billow show the dispersion of predicted values for training and validating examples
and is also a good results. Where:
• X axis for both images show true values for training
• Y axis for both images show predicted values after s caling.
Learning cycles: 201701
Target error: 0.0001
Average training error: 0.000395

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
87

(A)
(B)

Figure IV. 15. Prediction

This is the dispersion of predicted values for training and validating examples and is also a good
results.
Column v alues
The Column Values view. All Input/output column values are shown graphically. The left hand
vertical scale shows the scaled value from 0 to 1. The right hand scale shows the real value
calculated using the highest and lowest values in the column. The blue line is the true value, the
red line is the running average and the black dotted line is the overall trend. The lavender line is
the Risk number. This is the number of times that any forecasted value had to be adjusted before
the network produced an error close to the error produced before the forecast.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
88

Figure IV. 16. Column values

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
89
Example errors

Figure IV. 17. Examples errors

The Error Examples view displays the absolute error and the relative error generated by up to 30
examples 25 for training and 5 for validating, in descending order from the largest error. The order
can be switched in ascending order from the smallest error by clicki ng on the toolbar icon of the
sort order.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
90
Evaluation of neural model performance

Table IV. 8. Evaluation of neural model performance for training examples

Figure IV. 18. Comparison between predicted values and experimental values for training
examples
12345678910111213141516171819202122232425
EXP DC 1222212112 1114 12211 12112 1213 22
ANN DC 1221211112 1114 12211 12112 1213 220246810121416EXP/ANN
Nr. of.expCmparaison between predicted values and experimental values
for training examples
EXP DC ANN DC

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
91

Table IV. 9. Evaluation of neural model performance for validating

Figure IV. 19. Comparison between predicted values and experimental values for validation
examples

– I evaluated the performance and I found Ꜫ= 4% for training examples and Ꜫ=0% for
validation examples, we can say that training average error is acceptable because it is less
than 5 . What concerns the average error of validation has obtained a perfect result.

– What concerns the average error of validation has obtained a perfect result.
Input i mportance
The Input Importance view shows the importance and the relative importance of each Input
column. The I mportance is the sum of the absolute weights of the connections from the input node
to all the nodes in the first hidden layer. The inputs are shown in the descending order of
importance from the most important input.

1 2 3 4 5
EXP DC 1 1 1 1 1
ANN DC 1 1 1 1 100.20.40.60.811.2EXP/ANN
Nr of experementsComparaison between predicted values and
expiremental values for validation examples
EXP DC ANN DC

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
92

Figure IV. 20. Input values

Sensitivity

Figure IV. 21. Sensitivity

The Sensitivity view shows how much an output changes when the inputs are changed. The inputs
are all set to the median values and then each in turn is increased from the lowest value to the
highest value. The change in the output is measured as each input is increased from lowest to
highest to establish the sensitivity to change. Insensitive inputs are not shown. The inputs are
shown in the descending order of sensitivity from the most sensitive input.

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
93
IV.5. Conclusion

In this chapter we have introduced an ANN network operating solution for maintenance and
industrial monitoring applications.
The principle of this solution is to have a neural network loaded to process a database of grinding
machine tools. The second part of this solution, which is just as important as the treatment part,
concerns the learning and updating phase of the ANN network.
We also performed performance tests against several cycles of the neural program, we consider
that this solution opens very interesting perspectives in the treatment of database monitoring
variables.
In this sense, we managed to obtain good results for both phases so we can confirm that this model
is the best model for use in SA Rulmenti Bârlad.
This topical trend is starting to take hold in companies that want to reduce maintenance costs.
Indeed, this neural network utilization solution for predictive maintenance has certai n advantages
that are essentially:
• Better know the budget allocated to maintenance and therefore have the opportunity to
optimize costs,
• Refocus on its real production business and entrust the maintenance function to
professionals.
• Industrial applications of the techniques explored and proposed, in order to converge
towards a more efficient maintenance of the machines.
The solution presented in this chapter is very suitable for such an application. The neural
monitoring program carry out the le arning, especially as the database needed for this learning
(history of machine tool rectification).

Ch IV. Development of Predictive Maintenance Model Using the Neural Network Algorithm, for
Grinding Machine – Tool
94
References

[1] M. Sam, The basics of abrasive cutting, Grinders, 17 December 2016.
[2] S. K. a. S. Schmid, Manufacturing Process for Engineering Materials, Pearson Education
South Asia Pte LTD, ISBN 978 -981-06-9406 -7, 2013.
[3] J. T. Black and R. A. Kohser, DeGarmo’s , John Wiley & Sons Inc, , Materials and
Processes in Manufacturing, ISB N 978 -0-470-92467 -9., 2012.

95
CHAPTER V.
General C onclusions & Original Contributions
The research presented in this thesis deals with the study of artificial neural networks for the
predictive maintenance of grinding machine tools.
In the framework of our work, our study concerns the processing of defects in machine tools
grinding who presents many interruptions resulting from unplanned maintenance operations.
In this sense, we have proposed a new neural network model that takes advantage of the
performance and lear ning simplicity of ANN networks.
The new ANN will be used in the predictive maintenance model to implement at Rulmenti S.A.
Bârla d.
We have improved the prediction performance of the ANN network through the evolution of
learning.
The use of artificial neural networks in manufacturing indu stries in failure prediction is essential
for predictive maintenance due to its ability to prevent failures, reduce maintenance costs, and
achieve high levels of productivity.
As a result, the ability to detect and isolate defects, as well as to predict a future state of a system,
have become priority areas of research.
Grinding machine tools represent a dominant class in production systems and can occupy strategic
positions in them. The monitoring of their operating states is therefore of undenia ble interest, in
order to achieve the desired objectives.

V.1. the main contributions of this thesis

The main contributions of this thesis are grouped in three parts. A first part brings together the
state of the art around three points of interest in close correlation:

CHAPTER V. General Conclusions & Original C ontributions
96
• Presentation of the wo rking environment: Dunarea de Jos University of Galati and URB Group
– Rulmenti S.A. Bârla d (Chapter I).
• Introduction of Maintenance, Monitoring, and Industrial Diagnosis (Chapter II).
• Artificial Neural Network Method (Chapter III).
The second part of our work, articulated around the following two chapters, summarizes the
essence of our scientific contribution. We have proposed (Chapter IV) an architecture of a new
neural network for industrial surveillance (ANN). A set of comparative tests have been applied to
the database (Chapter IV), with the use of learning processes to improve system performance to
achieve an error threshold that does not exceed 5 %, that is, aiming to minimize the average error.
The testing step allowed us to highlight the need to develop a neural model that has better
performance.
• I proposed a neural network (ANN) with a study and simulation of pe rformances that
provide stability.
• I evaluate a network performance (ANN) proposed on problems of defects in the grinding
machine tool.
• I create an intelligent monitoring solution for the predictive maintenance of the grinding
machine tool.
V.1.1 . Part one – state of the art
The first step undertaken in this study is the development of definitions related to industrial
maintenance and its types, the purpose of predictive maintenance in the industrial sector, and the
main causes of failure.
We have given a state of the art also on the networks of artificial neurons, the main fields of
applications, the probable architecture with a precise type of problem studied, the two neuronal
architectures most used in surveillance are the Perceptron Multi Layer s (PMC) with its global
representation of its data space and Radial Basic Feature Networks (RFR).

CHAPTER V. General Conclusions & Original C ontributions
97
V.1.2 . Part Two – Scientific Contributions
Different architectures are used for different problems, the architec ture that suits our database is
the feedforward architecture with a one -way propagation. However, their learning is extremely
heavy.
We then asked ourselves the following question: why did we use this architecture of a neural
network?
We therefore established a study of the experimental data with the data predicted by the neuron
network in several cases in the interrogation phase and it was found that the network is not suitable
when in using the frequency of the mai ntenance and the code of defect at the same time as output,
41.99% and 29.96% , respectively, when the learning is 35017201 cycles, this result exceed s our
average error limit of 5%.
We have, on the one hand, restructured this study so that it is in adequacy with the study of all
data, we used the frequency of the maintenance as an output this time with three entries which are
respectively the defect code , maintenance time and reaction time the result was 49.47% when the
learning cycle is 34088259, we note that the learning progress is stopped automatically.
The results provided are in agreement with the experimental data. On the other hand, we have
deepened this study with other performance tests by computer simulation, with the use of fault
code as output variable obtained better results 3.63% with a learnin g cycle of 253243 for the first
phases.

In the second phase on a successful get a good relative average error for training of 4 %.
Throughout theoretical study with simulation tests, we found that this result validates our 3 -4-1
neural network architecture that has better performance.
Through this database, we showed the simplicity with which the ANN network is able to learn
several data in a simple way. Its ability to generalize locally allows it to recognize sequences close
to those learned an d to detect unknown sequences. This type of application is very useful for the
predictive maintenance of the grinding machine tool. The network learns to recognize sequences

CHAPTER V. General Conclusions & Original C ontributions
98
of good operation and to detect known malfunction sequences from the prediction of fault types
that can be found in grinding machine tools.
So each dysfunction sequence has its own cause. The expert will be able to diagnose the cause of
each malfunction sequence after the fault detection of the SIW 5 machine and teach the ANN
network. Th e limitation of the ANN network in this type of application lies in its inability to learn
complex sequences, that is to say sequences where an event occurs more than once.
The learning process can be used for the early detection of defects that causes deg radation and to
plan interventions and re actions after defect prediction, this type of prediction before the failure
occurs can avoid serious industrial accidents of grinding machine tools and a total stop of
production system , knowing that the main goal i s to ensure the continuity of production.
Being as example the defect with the code 1 (lack of oil) it result that the neural network was able
to predict this defect in all experiments to avoid th em.
We have improved the prediction performance of the ANN n etwork through the evolution of
learning.
This way and the network capacity to learn allows to guarantee a stability of the result with a
prediction error close to the global minimum.
V.1.3. Part Three – Industrial Exploitation
One of the objectives of our work is to develop a neural model for the predictive maintenance of
the SIW5 grinding machine tool, easily used for indus trial maintenance applications.
The Easy NN program was used for AN N control, construction , and network in terrogation to
achieve better prediction based on learning progression in an undetermined time in order to
minimize the average error .
Indeed, many companies opt for this kind of solution that allows them both to better control their
maintenance budget but especially to refocus on their true production business.
The proposed neural model will be used by Rulmenti SA B ârlad to drown up a predictive
maintenance plan for year 2020 that will validate the neural model.

CHAPTER V. General Conclusions & Original C ontributions
99
V.2. Prospect

V.2.1. Scientific p erspectives
The basic remark we can make is that, despite the highly surprising and promising results obtained
by artificial neural networks, they are still far enough to match the reasoning skills of a human
expert.
We have seen that neural networks are very effective in detecting a failure, detection of a
degradation, prediction of defects (lack and loss of oil , by example ), modeling and prediction of a
temporal e volution of a non -linear signal ; on the other hand, the diagnostic function is, in our
opinion, a very complex task and can only be partially solved by the pattern recognition technique.
The main reason is that the human expert in his mission to try to diagnose the cause of a failure of
a whole machine or a subset of this machine, of ten uses other information than the quantitative
values (the sensor data). For example, he uses: his hearing to recognize the abnormal sounds of a
machine, his sense of smell to detect a burning smell and its origin, his vision to control the quality
of the parts produced by the machine and identify the various defects of the parts, his touch to
check the tension of a belt and also to see if there is no oil leak, its memory to remember his
knowledge previously acquired with other machines.
The question th at can arise and that can open two completely antagonistic perspectives is whether
we want to replace at any cost the human expert in order to predict the defects of the machine tool
rectification by this diagnostic task to 100%?
If so, research will focus in this case on neuroscience and the development of artificial neural
networks in order to develop neural architectures that tend to be closer to biological neural
networks.
On the other hand, the second option that seems most interesting to us is how does the human
expert gather all the information needed to make his decision?
We believe that the neural network could offer an interesting way to extract some fuzzy knowledge
from the human expert (neural techniques). Once the architecture is identified (3-4-1), artificial
neural networks can very well be exploited to learn this form. On the other hand, the limit that can
be encountered using an artificial neural network is that generally, for a given application, the size

CHAPTER V. General Conclusions & Original C ontributions
100
of the input vector of a neural n etwork is a priori fixed. This is where the difficulty lies in suddenly
using a network of neurons for the task of prediction, diagnosis and maintenance in general
because, in practice, the information used for diagnosis and maintenance are often different We
can imagine then a solution, that is to say an artificial neural network (ANN) for each type of
defect. Each neural network obviously has its own input vector that best characterizes the duration
of maintenance, the response of maintenance, and the fre quency of maintenance.
This architecture (3-4-1) can offer an interesting solution for global decision -making regarding
fault prediction associated with different causes.
V.2.2. Industrial exploitation prospects
Several perspectives remain to be undertaken concerning industrial exploitation. The first is to test
the capabilities and performance of the charged ANN neural network for the prediction of defects
in the grinding machine tool. Indeed, the ANN network capabilities to detect the defe cts of the
grinding machine SIW 5 (Lack and losses of oil ) as well as its prediction capabilities that were
validated by the simulation on Easy NN. The problem is to any test of a system of detection and
diagnosis of failures lies precisely in the existence of th ese failures.
The second step is obviously to adapt the human machine interface in accordance with the type of
industrial equipment to be monitored. Indeed, the current interface has a generic character that was
used to perform the test and the interrogati on of the neuronal structure.
V.2.3. Futur e work
To generate a neural model for predictive maintenance for number of working hours depending on
defect code and frequency.
To realize a neural model for predictive maintenance for number of working hour s depending on
frequency.

101
ANNEX 1.
Initial Data B ase from URB Group – Rulmenti S.A. Bârla d

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
102

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
103

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
104

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
105

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
106

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
107

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
108

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
109

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
110

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
111

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
112

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
113

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
114

ANNEX 1. Initial Data B ase from URB Group – Rulmenti S.A. Bârla d
115

116

ANNEX 2.
New Data B ase

ANNEX 2. New Data B ase

117

ANNEX 2. New Data B ase

118

CODIFICARI
CODIFICAREA DEFECTELOR
MOTIV-EXPLICATIE COD DEFECT (modelare)
lipsa ulei 1
pierderi ulei, apa, aer (determina remedieri: schimbat furtun, reglat presiune pompa, …) 2
traductor 3
defect mecanic general 4
defect electric disp., contactor, salt 5
electrobrosa 6
defect electric/electronic – avans 7
siguranta electrica 8
defect electric/electronic general 9
defect electric – droser 10
defect mecanic – curea dispozitiv 11
defect mecanic – surub inclinare masa 12
diamantare 13
garnituri 14

ANNEX 2. New Data B ase

119

CODIFICAREA MASINILOR UNELTE – COD MU
Nr. Inv. MU COD MU
200588 SIW3 31
200640 SIW3 32
200615 SIW3 33
200616 SIW3 34
200760 SIW3 35
200580 SIW3 36
Nr. Inv. MU COD MU
200658 SIW4 41
200704 SIW4 42
200703 SIW4 43
200702 SIW4 44
200657 SIW4 46
200729 SIW4 47
200730 SIW4 48
200731 SIW4 49
200733 SIW4 410
200745 SIW4 411
Nr. Inv. MU COD MU
200635 SIW5 50
200636 SIW5 51
200735 SIW5 52
200738 SIW5 53
200739 SIW5 54
200744 SIW5 55
200749 SIW5 56
200757 SIW5 57
200758 SIW5 58
200760 SIW5 59
203154 SIW5 510
46844 SIW5 511

120
ANNEX 3.
Introduction to Easy NN-Plus

Introduction to Easy NN-Plus

Easy NN can be used to create, control, train, validate and query neural networks.

Getting Started

In order to create a neural network, press the New toolbar button or use the File>New menu
command to produce a new neural network,

An empty Grid with a vertical line, a horizontal line and an underline marker will appear. The
marker shows the position where a grid column and row will be produced. Press the enter key and
you will be asked "Create new Example row?" – answer yes.

You will then be asked "Create new Input/output column?" – answer yes.

You have now created a training example with one input. The example has no name and no value.
Press the enter key again an d you will open the Edit dialog. This dialog is used to enter or edit all
of the information in the Grid.

ANNEX 3. Introduction to Easy NN -Plus

121

Edit Dialog
Toolbars

Most menu commands also have toolbar buttons. The toolbars can be positioned anywhere.

ANNEX 3. Introduction to Easy NN -Plus

122

123
ANNEX 4.
The Results of the Predicted Values U sing NN MODEL

REA T MIN D MIN F EXP DC ANN DC REL Err
EXP 01 1.35 0.50 65.00 1 8.99581 799.581
EXP 02 1.15 0.50 65.50 1 9.03184 803.184
EXP 03 0.17 0.50 66.16 1 9.20982 820.982
EXP 04 1.00 0.50 256.83 2 9.09977 354.9885
EXP 05 1.10 1.00 151.50 2 9.06158 353.079
EXP 06 0.85 1.00 33.00 2 9.08167 354.0835
EXP 07 1.50 0.50 149.66 1 8.9866 798.66
EXP 08 2.00 1.00 16.00 2 8.87379 343.6895
EXP 09 1.50 0.50 16.00 1 8.95888 795.888
EXP 10 0.90 1.00 182.50 2 9.10429 355.2145
EXP 11 1.00 0.33 100.83 1 9.06551 806.551
EXP 12 0.10 1.00 48.48 12 9.22079 23.16008
EXP 13 1.60 0.50 132.50 1 8.9651 796.51
EXP 14 0.80 1.00 308.50 1 9.14957 814.957
EXP 15 1.00 0.50 0.00 1 9.04509 804.509
EXP 16 0.25 3.50 16.75 14 9.19867 34.29521
EXP 17 1.40 0.50 81.00 1 8.99018 799.018
EXP 18 0.17 2.00 102.00 2 9.22381 361.1905
EXP 19 0.25 0.67 210.00 2 9.22725 361.3625
EXP 20 0.25 2.00 166.00 11 9.22305 16.15409
EXP 21 1.20 0.50 69.00 1 9.02357 802.357
EXP 22 0.50 1.50 439.50 2 9.23492 361.746
EXP 23 1.20 0.50 98.50 1 9.02977 802.977
EXP 24 0.30 0.50 5.50 12 9.17298 23.5585
EXP 25 0.80 0.50 46.00 1 9.09089 809.089
EXP 26 1.25 2.00 0.00 2 9.00947 350.4735
EXP 27 1.20 0.50 0.00 1 9.00918 800.918
EXP 28 0.50 1.00 464.00 13 9.23856 28.93415
EXP 29 0.25 0.50 0.00 2 9.1809 359.045
EXP 30 0.15 1.00 0.00 2 9.20124 360.062
509.8739