Machine Learning Framework for Tool Condition

Monitoring in Milling
by
Omer Faiz
Submitted to the Department of Management and Production
Engineering

Master of Management Engineering

at the
POLITECNICO DI TORINO
04 2021

Author : Omer Faiz

Supervised By : Dr. Giulia Bruno
Assistant Professor,DIGEP
Abstract
In the smart manufacturing era, the dynamics of monitoring and maintenance of the
machines are changed. After the 4th industrial revolution, artificial intelligence and
machine learning techniques are proven to be beneficial for carrying out predictive
maintenance of machines. Internet of Things (IoT) along with the Cyber-Physical
Systems (CPS) has made it possible to conduct a data-driven prognosis of a system.
Predictive maintenance techniques have been developed in order to monitor an in-
service machine for estimating when maintenance should be performed. Big data
analysis and machine learning techniques enable the detection of the current health
state and the remaining useful life of the equipment.
Above mentioned developments have played an efficient role in increasing produc-
tion efficiency by minimizing downtime during the manufacturing processes. This
study describes the application of a tool condition monitoring (TCM) framework to
a real-time milling data-set, with the aim of classifying the tool condition (worn, un-
worn) of the milling tool during a running process. The application of the framework
can help in optimizing the maintenance operations and preventing the breakdown of
the equipment.

2
Acknowlegdement

Finally, I am concluding the journey of my master’s which started two years back, it’s
been an engrossing journey that enabled me to polish my personal and management
skills. Politecnico di Torino provided me with a great opportunity to internationalize,
where I met people from all around the world and learned from them. This is the
time when I would like to thank the people who stood by me during this expedition.
Firstly, I would like to thanks Dr. Giulia Bruno, who guided me professionally
during the whole thesis work.
I am grateful to my parents and siblings for their kind support and prayers which
remained me focused at all times.
I would like to thank my friends Hassaan (Hashmi and Baig) who were always
there to support me with their expertise.
My deep gratitude goes to some friends Abdullah, Faseeh, Izza, and Noor who
provided their exceptional support during my whole master’s degree program.
In the end, I would like to thank everybody who kept me motivated and added
fun to this endeavor.

Regards,
Omer Faiz

3
4
Contents

1 Introduction 13
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2 Thesis Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3 Thesis Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2 Industry 4.0 and Smart Manufacturing 19

2.1 Industrial Revolution . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.1 Industry 4.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1.2 Industrial IoTs . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2 Smart Manufacturing . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Smart Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

3 Artificial Intelligence Based Prognosis for Predictive Maintenance 27

3.1 Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Predictive Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.1 Condition Based Maintenance . . . . . . . . . . . . . . . . . . 28
3.2.2 Cloud-based Predictive Maintenance . . . . . . . . . . . . . . 29
3.3 Tool Condition Monitoring . . . . . . . . . . . . . . . . . . . . . . . . 30

4 Machine Learning Algorithms for Predictive Maintenance 33

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.1.1 Learning Strategies . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5
 4.3 K-Nearest Neighbors (K-NN) . . . . . . . . . . . . . . . . . . . . . . 36
4.4 Support Vector Machine (SVM) . . . . . . . . . . . . . . . . . . . . . 37
4.5 Decision Tree Classification . . . . . . . . . . . . . . . . . . . . . . . 38
4.6 Random Forest Classification . . . . . . . . . . . . . . . . . . . . . . 39
4.7 Artificial Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . 40

5 Manufacturing Processes and control 43

5.1 Use-Case Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Milling Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2.1 Types of Milling . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.2.2 CNC milling machine . . . . . . . . . . . . . . . . . . . . . . . 45
5.3 Tool Wear Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6 Data preparation 53
6.1 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 Data Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.2.1 Data Cleaning and Labeling . . . . . . . . . . . . . . . . . . . 55
6.2.2 Encoding Categorical Data . . . . . . . . . . . . . . . . . . . . 56
6.2.3 Feature Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.2.4 Data Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

7 Methodology Proposed 61
7.1 Data Pre-Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.2 Data Splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.3 Feature Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.4 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.5 Model Application . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
7.6 Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . . 64

8 Results 65
8.1 By using train-test split for 18 experiments . . . . . . . . . . . . . . . 66
8.2 By using cross-validation approach . . . . . . . . . . . . . . . . . . . 66

6
 8.3 Analysis with machining finalized data . . . . . . . . . . . . . . . . . 68
8.4 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
8.5 Feature Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
8.5.1 Principal component analysis . . . . . . . . . . . . . . . . . . 70
8.5.2 Feature Reduction by using SelectFromModel . . . . . . . . . 70

9 Conclusion, Limitations and Future Recomendations 73

7
8
List of Figures

2-1 Smart Manufacturing Framework . . . . . . . . . . . . . . . . . . . . 22

2-2 Evolution of Maintenance Paradigm . . . . . . . . . . . . . . . . . . . 24

3-1 Framework of Condition Based Maintenance . . . . . . . . . . . . . . 29

4-1 K-Nearest Neighbour . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

4-2 Support Vector Machine . . . . . . . . . . . . . . . . . . . . . . . . . 38
4-3 Decision Tree Classification . . . . . . . . . . . . . . . . . . . . . . . 39
4-4 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4-5 Structure of ANNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4-6 Tuning of ANNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

5-1 CNC Milling Machine . . . . . . . . . . . . . . . . . . . . . . . . . . 49

6-1 ’S’ shaped test artifact . . . . . . . . . . . . . . . . . . . . . . . . . . 53

6-2 Python Code: Data Cleaning and Labeling . . . . . . . . . . . . . . . 55
6-3 Python Code: OneHotEncoding . . . . . . . . . . . . . . . . . . . . . 56

7-1 Python Code: Flow Chart . . . . . . . . . . . . . . . . . . . . . . . . 61

7-2 Python Code: Flow Chart . . . . . . . . . . . . . . . . . . . . . . . . 62
7-3 Python Code: Feature Scaling by Standardization . . . . . . . . . . . 63

8-1 Python Code: Principal Component Analysis . . . . . . . . . . . . . 70

9
10
List of Tables

2.1 Smart Manufacturing Framework [24] . . . . . . . . . . . . . . . . . . 23

5.1 Examples for direct and indirect tool wear sensing methods . . . . . . 51

6.1 Machining data-set of a CNC milling . . . . . . . . . . . . . . . . . . 54

6.2 Examples for direct and indirect tool wear sensing methods . . . . . . 59
6.3 Examples for direct and indirect tool wear sensing methods . . . . . . 60

8.1 By using train-test split . . . . . . . . . . . . . . . . . . . . . . . . . 66

8.2 By using cross-validation approach . . . . . . . . . . . . . . . . . . . 67
8.3 Analysis with Machining Finalized experiments . . . . . . . . . . . . 68
8.4 Results after Feature Extraction . . . . . . . . . . . . . . . . . . . . . 69
8.5 Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . . 70
8.6 Feature Reduction by using SelectFromModel . . . . . . . . . . . . . 71

11
12
Chapter 1

Introduction

Manufacturing processes these days are becoming more complex, dynamic, and con-
nected. Industry operations face difficulties with strongly non-linear and stochastic
action due to the numerous complexities and inter-dependencies that occur. The
manufacturing industry is entering a time of incredible progress and transition led by
increased integration of sensors and the Internet of Things ( IoT), improved availabil-
ity of data, and developments in robotics and automation. This leads to extensive
digitization of the factories and challenges industrial firms to analyze, evaluate, and
reassess their existing activities and potential strategic directions in the modern age
[1].

Predictive Maintenance systems are developed to track, diagnose and forecast de-
fects and degradation of system components before criticality. The ultimate aim is
to avoid disruption, recognize root causes for follow-up operation, and allow effective
evidence-based maintenance preparation and optimization [3]. In this regard machine
learning techniques can inductively learn useful data patterns and refer to the pro-
cessing state and outcome reinforcing them. The idea that data is a representation of
the features of the physical machine [4]. The purpose of the prognosis is to estimate
the time variation of system output degradation from its current state to its final
malfunction [5].

13
1.1 Background

The background of manufacturing is closely related to the history of increasing the

efficiency of industrial machinery to remove unnecessary downtime. Research on con-
dition monitoring, root cause analysis and remaining useful life (RUL) prognosis lays
out the fundamental expertise for maintenance and Prognostics Health Management
(PHM) [7].A discipline that measures a health condition and remains a useful life de-
pendent on past and present operational conditions is also referred to as PHM [6]. The
idea is to allow timely detection and isolation of precursors or emerging faults of the
components, forecast their advancement, and promote responsible decision-making
[2].

Classical prognostic approaches fall into two categories: model-based and data-
driven prognostics. Model-based prognosis refers to theories based on statistical mod-
els of system actions resulting from physical laws or the distribution of probabilities.
For example,traditional model-based prognostics include methods based on Wiener
and Gamma processes[8], hidden Markov models[9], Kalman filters[10] and particle
filters[11]. One of the drawbacks of model-based prognosis is that an in-depth analysis
of the underlying physical mechanisms contributing to device failures is needed.

Compared to model-based prognostics, Data-driven prognostics refers to methods

that construct a predictive model using a learning algorithm and a huge volume of
historical data. The specific advantage of data-driven approaches is that an in-depth
interpretation of the physical actions of the device is not necessary. In comparison,
data-driven approaches do not consider any fundamental distributions of probability
[12].

Data-driven prognostics is based on several machine learning algorithms which is

used for the tool condition monitoring. These machine learning algorithms predict
tool wear and the estimate remaining useful life (RUL) of a tool. In this regard
Kothuru [12] did audio-based tool condition monitoring in milling of the work-piece
material With the Hardness Variation Using Support Vector Machines(SVM) and
Convolutional Neural Network CNNs. The proposed SVM and CNN models have

14
shown a significant overall prediction accuracy of 98% and 96.3% in tool wear moni-
toring and 68% of overall prediction accuracy in work-piece hardness variation mon-
itoring [12]. Sohyung Cho [14] did tool breakage detection using a support vector
machine for an end milling operation, and used cutting force and spindle power as
input parameters. The algorithm is computationally efficient and robust in higher
dimensions [14]. Clayton Cooper [15] used convolutional neural network-based tool
condition monitoring in vertical dry milling operations using acoustic signals. They
achieved a reduction in tool wear detection time and demonstrated the untapped
potential of acoustic signal monitoring [15].
A. G, Mebrahitom Yuan [16] did remaining tool life prediction of an end milling
operation based on the force sensor signal, they used Support Vector Regression
(SVR) and neural network for tool wear estimation by using cutting force, acoustic
emissions and vibrations as input parameters. It is found that the upper/lower limits
of predicted tool wear width are closely following the trend of actual wear width of
all the three cutting tools. The maximum error between the median of predicted tool
wear width and the actual wear width is around 5%, which confirms the effectiveness
of proposed method for long-term tool life prediction [16]. X.Li [17] also used a Fuzzy
Neural Network (FNN) model to monitor the tool degradation through tool wear
and it was compared with the conventional Multi Regression Models (MRM) [17].
Dazhong Wu [18] used cloud-based machine learning for tool wear prediction in the
milling process. The prediction of tool wear in milling operations was performed with
the random forest and PRF algorithms with the input parameters of cutting forces,
vibrations, acoustic emissions. The performance of the random forest algorithm was
evaluated using mean squared error, R-square, and training time. The experimental
results have shown that random forests can generate very accurate predictions.[18]

15
1.2 Thesis Objective

The objective of this thesis is to classify the tool condition (worn,unworn) during a
milling process, for predictive maintenance by using a data driven prognosis approach
based on several machine learning algorithms.

Predictive maintenance is a process in which the service life of essential compo-

nents is estimated based on examination or evaluation in order to maintain the quality
of components during the span of their service life. The conventional maintenance
approach is following run-to-failure management based on the philosophy: "if ain’t
broken, then don’t fix it", this approach leads to the time losses in manufacturing
and costly repairs. One of the drawbacks of reactive maintenance is that maintenance
resources (e.g. personnel, equipment and spare parts) are hard to forecast for repairs.
Hence predictive maintenance is an effective methodology which can optimize the
usage of the components by estimating their useful life and predicting the schedule
for repairs.

Tool Condition Monitoring (TCM) is one of the essential steps of predictive main-
tenance. TCM is to monitor pattern values, analyze data on degradation, and using
a monitoring system designed to track condition of a component during the process.
Mostly tool condition monitoring is done by incorporating sensors with the com-
ponents and monitoring the activity continuously. These observations are feed into
machine learning algorithms which in-turn classify the tool condition (worn, unworn),
and predicts the remaining useful life of the components based on historical data.

In today’s manufacturing industry, the end milling process is a widely used process
of machining, monitoring the quality of the tool is very important since it reflects the
quality of the product being produced. This technique can efficiently reduce the cost
of the process by reducing the rejections and wastage occurring due to the use of
the worn tool. Predictive maintenance can be a great breakthrough to achieve an
optimal downtime and finish product quality in a milling process. It can provide the
economies of machines and the dimensional accuracy in the machining process which
is of the foremost importance in the manufacturing industry.

16
1.3 Thesis Structure
∙ Chapter one contains the background as well as the objective of the research.
Moreover it gathers the past study and explained why a data-driven approach
for monitoring of machines could provide better results.

∙ Chapter two contains a detailed discussion about industrial evolution and emerg-
ing smart manufacturing industries. A detailed guide about the maintenance
techniques which are being used in the industry.

∙ Chapter three reviews the literature regarding Machine Learning and Artificial
intelligence. Also it contains a discussion about Prognosis Health Management
(PHM), the detailed description of Tool Condition Monitoring (TCM). More-
over this chapter includes a discussion on cloud-based predictive maintenance.

∙ Chapter five includes a review of Machine Learning algorithms that can be used
for data-driven prognostics for predictive maintenance.

∙ Chapter four includes the topologies of manufacturing processes, a detailed dis-

cussion on milling operation and the possible maintenance operations required
in a milling process.

∙ Chapter six contains the details of data preparation.

∙ Chapter seven contains the methodology which is being proposed to train and
test the data-set. Moreover, it will provide the significance of algorithm selected
to conduct prognosis.

∙ Chapter eight contains detailed discussion of all obtained results of the analysis.

∙ Chapter nine includes the conclusion of the thesis with the the limitations we
faced during the research work, and the recommendations for the future re-
search.

17
18
Chapter 2

Industry 4.0 and Smart

Manufacturing

2.1 Industrial Revolution

2.1.1 Industry 4.0

Manufacturing technologies have rapidly become widespread in the manufacturing

sector. The transition towards intelligent and interconnected manufacturing pro-
cesses, widely recognized as the 4th industrial revolution and referred to as industry
4.0, is the manifestation of three previous revolutions based on mechanization, large-
scale production, and automation, respectively. While the ability to access these
technologies is a significant step towards improving the manufacturing sector, the
transformation from the general capabilities of digital manufacturing to actionable
implementation decisions for manufacturing companies is yet another obstacle.The
automation of manufacturing processes in third industrial revolution acted as a break-
through by offering the infrastructure for gathering useful information and providing
utilization mechanisms [19].
In addition , the ability of hardware and manufacturing devices to communicate
with software systems has led to a new classification of equipment, Cyber-Physical
Systems (CPS) or systems that are designed from and based on the synergy of soft-

19
ware and physical components [20]. Cyber-Physical Systems includes components like
multiple sensory input/output devices, touch screens, cameras, GPS chips, speakers,
microphone, light sensors, proximity sensors, which enhance the man-machine com-
munication. Such systems have become an important part of lean manufacturing
by offering deep insights into the production process. Established closed-loop man-
agement, process optimization and quality control mechanisms are among the most
common applications for CPS [19].

2.1.2 Industrial IoTs

Industrial IoT (IIoT) is the network of intelligent and highly connected industrial
components that are deployed to achieve high production rate with reduced oper-
ational costs through real-time monitoring, efficient management and controlling of
industrial processes,assets and operational time [22].
Internet of Things (IoT) is a growing paradigm which interconnects humans and
objects via internet. Thus, these objects can be identified and communicate over
the internet. With IoT in manufacturing, devices on the factory floor, manufactur-
ing processes, production systems and people are connected through the internet.
IoT capable devices used in the manufacturing process to improve the efficiency and
productivity, which is usually called as smart manufacturing. IoT devices installed
in the smart manufacturing systems keep on gathering information. IoT also has
diverse applications ranging from the integration of control rooms, asset monitor-
ing, problem-proof maintenance and process planning, additive manufacturing and
augmented reality [21].Industrial Internet Consortium (IIC) is an open membership
organization founded by several US-based companies with an idea to catalyze the
concept of industrial internet which has further potential to transform to business in
the industrial sector.
With IIoT, anything from major machines to transformers can be linked to the
Internet, providing status alerts and performance data. In this way, operators will
take preventive action on a possible issue until it costs the company and consumers
billions of dollars.In health care, IIoT and Big Data analytics can drive a variety of

20
positive aspects. Results such as enhancing patient flow, tracking and controlling
the usage of health care equipment, monitoring clinical, financial, and operational
steps, and managing efficiency of the workforce. GE Healthcare reports that these
advancements will reduce the cost of hospital equipment by 15 to 30 percent by an
additional hour of efficiency per shift [23].

2.2 Smart Manufacturing

The Industry 4.0 framework in the manufacturing sector covers a wide range of ap-
plications ranging from product design to logistics. The function of mechatronics, a
basic concept in the design of the manufacturing system, has been modified to make it
compatible with cyber physical systems. Smart product design based on personalized
specifications targeting individualized goods has been proposed. Figure 1 presents,
industry 4.0 framework for smart manufacturing systems. The horizontal axis shows
the typical issues in industry 4.0, including smart design, smart machining, smart
monitoring, smart control smart scheduling, and industrial applications. The verti-
cal axis shows challenges in another dimension of Industry, ranging from sensor and
actuator deployment to data collection, data analysis, and decision-making.
CPS and IoT-based manufacturing systems require the generation of large quan-
tities of data in industry 4.0, and big data processing is essential for the design and
operation of smart manufacturing systems. For example, using the big data ana-
lytics methodology, a comprehensive framework for data-driven risk assessment for
industrial manufacturing systems has been performed based on real-time data. Such
a subject has been widely documented in support of production optimization and
production of cyber-physical systems [24].
Increased manufacturing complexity, dynamic economics, and dramatically dif-
ferent performance goals would require the widespread use of networked, real-time
information-based technologies that turn facilities into knowledge-based facilities, a
reactive, predictive approach to function, threat detection to incident prevention, and
vertical decision-making to distributed intelligence and local decision-making with

21
Figure 2-1: Smart Manufacturing Framework

22
 Smart Machining In Industry 4.0, smart machining can be done with the help of
smart robots and other forms of smart objects that can sense and
communicate with each other in real-time.
Smart Monitoring Monitoring is an important part of the operation, maintenance,
and optimal scheduling for Industry 4.0 manufacturing systems.
The generalized deployment of different sensors has made smart
monitoring possible.
Smart Control High-resolution adaptive production control (i.e. smart control)
can be accomplished in Industry 4.0 by creating cyber-physical pro-
duction control systems. Smart control is primarily performed to
handle various smart machines or resources physically through a
cloud-enabled platform.
Smart Scheduling Smart scheduling primarily uses advanced models and algorithms
to derive information from data captured by sensors. Data-driven
techniques and an advanced decision architecture can be used for
smart scheduling.
Smart Design and Research at the stage of smart design and manufacturing includes
Manufacturing smart architecture, smart prototyping, smart controllers, and smart
sensors. Real-time control and monitoring help the introduction
of smart manufacturing. Supporting technologies include IoT,
STEPNC, 3D printing, industrial robotics, and wireless communi-
cations. Big data analysis plays a key role in smart manufacturing.
Smart Decission- Smart decision-making is at the core of Industry 4.0. The ultimate
Making aim of deploying large-scale sensors is to achieve smart decision
making via a detailed set of data. Realizing smart decision-making
needs exchanging knowledge and communication in real-time.

Table 2.1: Smart Manufacturing Framework [24]

global impact.Smart Manufacturing is an organization that combines the knowledge of

the consumer, its partners, and the public. It responds as an organized,performance-
oriented organization, minimizing energy and material use while optimizing environ-
mental sustainability, health and safety, and economic competitiveness. The company,
operations management, labor, and manufacturing innovations lead to new ways of
thinking about the manufacturing process [25].

23
 Figure 2-2: Evolution of Maintenance Paradigm

2.3 Smart Maintenance

Modern manufacturing systems typically consist of several machines to satisfy the de-
mand for good quality and high functional complexity goods. The probability of ma-
chine failure is aggregated as the number of machines increases in the system. In any
industry, sudden failure can lead to huge economic losses due to machine/production
downtime. For example, a standard car assembly line experiences a $20,000 loss for
every minute of downtime.Therefore it is necessary for large and complex manufac-
turing systems to have successful maintenance operations that improve the status of
machines.During the technological revolutions, maintenance techniques have under-
gone a radical progression and are currently underway Fig[2].
Usually, the maintenance management is categorized into different policies [26]:

∙ Corrective maintenance is an unscheduled repair, where the equipment is al-

lowed to operate until it fails unless a maintenance operation is carried out.

24
 ∙ Preventive maintenance, perhaps the most common maintenance policy, A pe-
riodic collection of operations on the equipment shall be carried out to minimize
the risk of failure. This form of maintenance is carried out while the equipment
is still in operation.

∙ Predictive maintenance includes the use of sensor technology and analytical

tools to predict when equipment failures can occur and to prevent failures from
occurring through maintenance.

Corrective maintenance is the most straightforward technique, among others,

where activities are carried out due to failure to restore the machine to a partic-
ular state. Corrective maintenance can be very expensive, as it is often combined
with unnecessary downtime, which often costs three or four times as much as planned
downtime.Alternatively, preventive maintenance requires steps taken to sustain an
object in a particular condition by providing systematic inspection, assessment, and
prevention of incipient failures. The simple implementation of preventive mainte-
nance is to perform periodic acts based on assumed behavior, such as the mean time
between failures.To prevent unnecessary downtime, a preventive maintenance interval
is needed. It is also conservatively set to be much shorter than the average to allow for
differences between machines. Besides, the deterioration process of machines can be
strongly linked to operations and external influences, making it difficult to calibrate
the actions of machines.
Nowadays, industrial maintenance is primarily reactive and preventive, the pre-
dictive approach is applied only to severe circumstances. Traditionally, these main-
tenance techniques do not take into account the vast amount of data generated on
the shop floor and the evolving Information and Communication Technology (ICT)
available, e.g. Internet of Things (IoT), Big Data, Advanced Data Analytics, Cloud
Computing, and Augmented Reality.However, the maintenance paradigm is shifting
and industrial maintenance is now seen as a strategic factor and a profit-making factor
to ensure efficiency in industrial systems.

25
26
Chapter 3

Artificial Intelligence Based Prognosis

for Predictive Maintenance

3.1 Artificial Intelligence

Nowadays, manufacturing objectives are becoming extremely challenging, with a mul-

titude of demands arising from a growing product and process complexity, increased
uncertainty in consumer demand and expectations, and steadfast competitive pres-
sures from competitors in the market to remain profitable. Seen from a positive
perspective these extreme conditions provides an opportunity for the unique capa-
bilities of Artificial Intelligence (AI) over the daily used conventional tools. AI tools
capable of defining and classifying multivariate, non-linear trends in operating and
output data that are hidden from the plant engineer.AI provides the opportunity to
turn large volumes of complex manufacturing data, which have become commonplace
in today’s factory, into actionable and insightful knowledge. AI tools which estab-
lishing core competencies in maintenance are Prognostic and Health Management
(PHM) and Condition Based Monitoring (CBM). The objective is to allow timely
identification and isolation of precursors or incipient faults of components, predict
their development, and promote reasonable decision-making [1].

27
3.2 Predictive Maintenance
Predictive Maintenance in its broad sense is an application of AI in the maintenance
which apply data analysis techniques to information generated in shop floor processes
for the identification of anomalies in the behavior of properties. This approach ex-
tends PHM and CBM maintenance approaches by considering machine learning and
augmented reality technology to assist maintenance technicians during maintenance
interventions by offering guided intelligent decision-making support articulated by
the use of human-machine interaction technologies [26]. While the main objective of
PHM is to provide the health status and predict the Remaining Useful Life (RUL) of
components or facilities, financial benefits such as operating and repair cost savings
and increased lifespans are also achieved [26].

3.2.1 Condition Based Maintenance

Condition-based maintenance (CBM) makes maintenance decisions based on informa-

tion about the current deterioration of the machine and its evolution. It uses system
degradation information derived from online monitoring systems to optimize CBM
decisions by balancing the probability of failure and achievable profits.It is impor-
tant to understand how the dynamics of production shift under the CBM stop events
and build an optimal CBM control system to minimize the negative effect of CBM
stoppage on output [27].The idea of an AI-based diagnosis is to formulate a task as
a classification and to relate the details contained in the data to the types of fault
and the degree of severity. The framework of CBM is shown in the Fig [3-1] which
describes that CBM is further categorized into three types 1) Physics based 2) Model
based 3) Data-driven/AI.

∙ Physics-based

∙ Model-based

∙ Data-driven/AI

28
 Figure 3-1: Framework of Condition Based Maintenance

3.2.2 Cloud-based Predictive Maintenance

Cloud-based prognosis, which reflects a new form of service-oriented technology to

help several organizations to deploy and operate prognosis services over the internet
can be imagined, powered by the ability of cloud computing and cloud development.
First, the data is collected by monitoring the system state of the machine. Sensors
are connected on the shop floor, remotely and dynamically. Based on in-situ measure-
ments, remote data analysis, and root cause failure diagnosis and prognosis are then
performed.Secondly, collaborative technical teams can have expert knowledge in the
cloud, which is known as a knowledge base that can be accessed across the internet.
The results of the forecasting service and the calculation of time-to-failure would lead
to predictive maintenance planning, which can be used remotely and dynamically on
the factory floor.
Cloud-based prognosis have following characteristics [28]:

∙ Service Oriented ability: All prognostic system functions are provided as

cloud-based services that are accessible through web browsers or the internet.
Local installation and servicing do not require costly software packages.

∙ Accessibility and promotion of robustness: Cloud-based framework as a

29
 flexible and integrated solution, Configurable prognosis services can increase the
robustness of existing manufacturing processes. If required or applicable, the
cloud can choose alternative and pay-as-you-go prognostic services and various
maintenance options.

∙ Resource-aware ability: Cloud-based prognosis allows for tracking and fore-

casting system usage and conditions, locally or remotely, so that maintenance
decisions can become resource-conscious and well-informed.

∙ Collaboration and distribution: Cloud-based prognosis facilitates seamless

and collaborative sharing of process and system data between different applica-
tions at different locations.

3.3 Tool Condition Monitoring

Tool failure is a dynamic phenomenon involving one or more failure modes, including
excessive wear, brittle fracture (chipping), and breakage.In high-speed cutting pro-
cesses, late replacement of damaged tools can lead to machine breakdowns and have
a negative effect on the quality of the product, which will lead to scrapping and high
process costs. Accurate tool condition detection is important to achieve a high level of
competitiveness by increasing process efficiency and standardizing the quality of the
parts made. Tool Condition Monitoring (TCM) systems have therefore been widely
emphasized as an important concept for achieving these industrial requirements. Due
to the difficulties of implementing direct methods online in harsh cutting environ-
ments, more intensive efforts have been made towards indirect methods that define
the tool fault by detecting indicative process-borne features derived from different
signal measurements [29].
TCM techniques include direct measurement and indirect measurement of tool
wear. In direct measurement methods, the tool’s wear parameters are measured by a
microscope surface profiler, etc. These direct methods have the benefit of acquiring
high accuracy Dimension shifts due to tool wear, but they are not technically and

30
economically desirable. The direct methods are inappropriate to field conditions due
to the continues flow of the cutting fluid and disturbance due to offline operations,
which severely limits the application of direct measurement.
Indirect methods are designed on the basis of specific parameters which are directly
correlated with tool wear. These parameters includes machining process data such
as operational parameters and sensor signals (force signal, vibration signal, acoustic
signal, current/power signals, etc) are acquired and the relevant features are ex-
tracted from the data. Moreover, Artificial intelligence techniques are applied on
these extracted features to predict the tool condition. Generally, the indirect TCM
system consists of hardware and software components for signal acquisition, signal
pre-processing, features extraction, features selection and decision making.

31
32
Chapter 4

Machine Learning Algorithms for

Predictive Maintenance

4.1 Introduction
Machine learning is a branch of artificial intelligence that tends to allow machines
to perform their jobs skillfully through the use of intelligent software. Statistical
learning approaches are the foundation of intelligent software used to create machine
intelligence. Since machine learning algorithms need data to be trained, the disci-
pline must be related to the discipline of the database. Similarly, there are familiar
concepts such as Information Discovery from Data (KDD), data mining, and pattern
recognition.
Machine learning algorithms are helpful in bridging this gap of understanding. We
are not aiming to grasp the fundamental mechanisms that help us learn. We write
computer programs that will make machines learn and allow them to perform tasks
such as prediction. The purpose of learning is to create a model that takes feedback
and generates the desired outcome. Often we can grasp the model, but at other times
it can also be like a black box for us the work of which cannot be completed. The
model can be seen as an approximation of the mechanism that we want machines to
imitate. In such a case, we will get errors for some input, but most of the time, the
model is right. The accuracy of the results would therefore be another indicator of

33
the efficiency of the machine learning algorithm.

4.1.1 Learning Strategies

Machine learning is divided into two major learning strategies:

Unsupervised Learning

Unsupervised learning methods are used against data that have no labels.The main
objective of the unsupervised learning method is to explore the data and find some
hidden structure among them.The target of such unsupervised learning problems may
be to discover groups of similar examples within data called clustering, or to deter-
mine the distribution of data within input space, known as density estimation, or to
project data from high-dimensional space to low-dimensional space for the purpose
of dimensional reduction and data visualization. Conventional unsupervised learning
methods in the process industry include main component analysis, independent com-
ponent analysis, k-means clustering, kernel density estimation, self-organizing map,
Gaussian mixture models, manifold learning, vector data description support, and so
on [31].

Supervised Learning

Supervised learning deals with the cases where the data is labeled, discrete and con-
tinuous. The main applications of the supervised machine learning method include
process monitoring, fault classification and identification, online operating mode lo-
calization, soft sensor modeling and online applications, quality prediction and online
estimation, key performance index prediction and diagnostics, etc. Supervised learn-
ing is majorly categories into following:

∙ Regression If the desired output consists of one or more continuous variables,

then the task of supervised learning is called regression. A typical example of
the problem of data regression is the prediction of key performance in a process

34
 in which inputs contain routinely recorded process variables such as temperature
and pressure.

∙ Classification Classification is a type of supervised learning. It specifies the

class to which data elements belong to and is best used when the output has
finite and discrete values. It predicts a class for an input variable as well.e.g.
fault classification, or operating mode classification

In our case we are going to train classification models to predict the state of
milling tool i-e (a) worn or (b) unworn. Moreover in the following sections several
classification models are explained which will be further applied on our data.

4.2 Logistic Regression

Logistic Regression is a Machine Learning technique used for classification problems, a

predictive analysis algorithm centered on the idea of probability. In logistic regression
instead of predicting the exact value for the variable, the probability of happening of
that event is predicted.
The logistic function, also known as the sigmoid function, was developed by statis-
ticians to explain the characteristics of population growth in ecology, gradually grow-
ing and optimizing environmental efficiency. It’s a shaped curve that can take any
real-valued number and map it to a value between 0 and 1, but never precisely within
those limits.

𝑓 (𝑣𝑎𝑙𝑢𝑒) = 1/(1 + 𝑒− 𝑣𝑎𝑙𝑢𝑒) (4.1)

Where e is the origin of the natural logarithms (Euler number or EXP() feature in your
spreadsheet) and the value is the real numerical value that you want to transform.
Logistic regression is a classification algorithm used to allocate observations to a
discreet collection of classes. Examples of labeling issues include e-mail spam or not
spam, internet transactions fraud or not fraud, tumor malignant or benign. Logistic

35
 Figure 4-1: K-Nearest Neighbour

regression transforms the output by using the logistic sigmoid function to return the
likelihood value.

4.3 K-Nearest Neighbors (K-NN)

KNN can be used for both classification and regression predictive problems. However,
it is most commonly used for classification issues in the industry. In order to assess
any methodology, we usually look at three critical aspects:
1. Simple to understand performance
2. Calculation of time
3. Predicting Strength
The KNN algorithm assumes that identical objects occur in close proximity to
each other. In other terms, related objects are close to each other. Note in the
picture above that, much of the time, identical data points are close to each other.
The KNN algorithm believes that this assumption is valid enough for the algorithm
to be useful. . KNN captures the concept of similarity (sometimes called distance,

36
similarity, or proximity).There are several methods to measure the distance, and one
way may be preferable based on the problem we solve. The straight-line distance
(also called the Euclidean distance) is, however, a common and familiar choice.

k is therefore just the number of neighbors "voting" on the proximity of test

example’s class.To choose the K that is appropriate for your results, we run the KNN
algorithm multiple times for various values of K and choose the K that decreases the
amount of errors we find while retaining the algorithm’s ability to correctly make
predictions when it’s given data that it hasn’t seen before.

4.4 Support Vector Machine (SVM)

Support vector machine algorithm finds out a hyperplane in n-dimensional plane that
clearly classifies the data points. There are several potential hyperplanes that could
be chosen to differentiate the two types of data points. Our goal is to locate a plane
with the highest margin, i.e. the maximum difference between the data points in both
groups. Maximizing the margin gap gives some reinforcement such that potential data
points can be classified more confidently.

Hyperplanes are judgment boundaries that help to distinguish data points. Data
points falling on either side of the hyperplane can be assigned to various groups. Also,
the scale of the hyperplane depends on the number of functions. If the number of
input features is 2, the hyperplane is just a line. If the number of input features
is 3, the hyperplane can become a two-dimensional plane. It’s hard to think when
the number of features reaches 3. Support vectors are data points that are closer to
the hyperplane and change the direction and orientation of the hyperplane. Using
these support vectors, we optimize the range of the classifier. Removing the support
vectors would shift the direction of the hyperplane. These are the things that enable
us to develop our SVM.

37
 Figure 4-2: Support Vector Machine

4.5 Decision Tree Classification

The Decision Tree algorithm is a family of supervised learning algorithms. Unlike

other supervised learning algorithms, the decision tree algorithm can also be used
to solve problems of regression and classification. It breaks down the data set into
smaller and smaller subsets as, at the same time, the related decision tree is gradually
created. The end product of this is a tree with decision nodes and leaf nodes. The
decision node (for example, Outlook) has two or three branches (e.g., Sunny, Overcast
and Rainy). Leaf node (e.g. Play) is a classification or a decision. The largest decision
node in the tree that corresponds to the greatest indicator known as the root node.
Decision trees can handle both categorical and numeric results. Fig [5-3] shows an
example of decision tree classification.

Compared to other algorithms, decision trees require less time to prepare data
during pre-processing. The decision tree does not require data normalization. The
decision tree does not even include the scaling of results. Missing values in the
data often do NOT greatly impact the method of creating a decision tree. The
decision tree model is very intuitive and easy to communicate to professional teams
and stakeholders.

38
 Figure 4-3: Decision Tree Classification

4.6 Random Forest Classification

The Random Forest is a supervised learning algorithm. The "forest" that it creates
is a series of decision trees, usually trained by the "bagging" process. The general
principle of the baggage approach is that a mixture of learning models would maximize
the total result.

A benefit of the random forest is that it can be used for both classification and
regression problems, which make up the majority of modern machine learning models.
Let’s look at random forest classification, as classification is often used as the building
block of machine learning. In Fig[5-4] you can see how a random forest with two trees
would look like.

Random forest has almost the same hyperparameters as a decision tree or a bag-
gage classifier. The Random Forest adds additional randomness to the model as the
trees expand. Instead of looking for the most important feature when splitting a
node, it looks for the strongest feature in a random subset of features. This results
in a large range that usually results in a stronger model.

Another high benefit of the random forest algorithm is that it is very straight-
forward to calculate the relative value of each function to the forecast. Sklearn is a
fantastic way to calculate the value of a feature by looking at how often the tree nodes
that use the feature decrease impurity among all trees in the forest. It dynamically

39
 Figure 4-4: Random Forest

measures this score for each training element and averages the scores such that the
sum of all value is equal to one.

4.7 Artificial Neural Network

ANN is a human nervous system-inspired simulation technique that enables learning,
for example from representative data representing a physical event or decision-making
mechanism. The special characteristic of ANN is that it is capable of defining ob-
servational associations between independent and dependent variables and extracting
hidden details and detailed insights from representative data sets.
ANNs consist of a layer of input nodes and a layer of output nodes attached to
one or more layers of hidden nodes. Input layer nodes transfer information to hidden
layer nodes by firing activation functions, and hidden layer nodes fire or stay dormant
based on the proof given. Hidden layers attach weighting functions to the proof, and
when the value of a single node or group of nodes in the hidden layer exceeds a

40
 Figure 4-5: Structure of ANNs

threshold, the value is transferred to one or more nodes in the output layer. ANNs
must be qualified in a significant number of instances (data) [32].
Fig [5-5] shows how a normal neural network feels like. After training the model,
you need to look at two things: bias and variance. The bias-variance tradeoff is a
kind of problem that implies that low bias models would have high variance and vice
versa.High bias leads the model to ignore the relevant correlation between essential
features that can contribute to under-fitting. High variance may cause the model
to be susceptible to random noise in the data set contributing to over-fitting. Our
mission is to eliminate bias and variance and to find an optimum fit line as shown in
the Fig [5-6]

41
Figure 4-6: Tuning of ANNs

42
Chapter 5

Manufacturing Processes and control

5.1 Use-Case Definition

The use-case selected for the thesis is a "Milling Data Set",A series of machining
experiments were run on 2" x 2" x 1.5" wax blocks in a CNC milling machine in
the System-level Manufacturing and Automation Research Testbed (SMART) at the
University of Michigan. Machining data was collected from a CNC machine for varia-
tions of tool condition, feed rate, and clamping pressure. Each experiment produced
a finished wax part with an "S" shape - S for smart manufacturing - carved into the
top face.

5.2 Milling Processes

Among all the manufacturing processes the largely used is milling operation. The
milling process is a type of manufacturing process that allows the production of
complex parts and trimming of several materials such as metal,wood,plastics and
composites. Unlike the other manufacturing processes milling process is performed
on a dedicated milling machined equipped with a multi-sharp tool, mounted on a
spindle, and the process is carried out by the simultaneous movements of motorized
axes. The cutting head of the tool removes the material from the work piece when
comes in contact with the cutter.

43
5.2.1 Types of Milling

There are two major types of milling processes:

∙ Face Milling, in general, is defined as the process of cutting surfaces that are
perpendicular to the cutter axis, or the faces of a part.Face milling generates
a surface normal to the axis of rotation. It is used for wide flat surfaces. The
peripheral portions of the teeth do most of the metal cutting. There are several
forms of face milling:(a) conventional face milling, in which the cutter width
extends beyond the work piece on both sides,so the cutter overhangs the work
on the both sides;(b) partial face milling , where the cutter overhangs the work
on one side only; (c) End milling, In which the cutter diameter is less than the
working distance, the slot is cut into the part;(d) Profile milling, it is a type
of end milling which is used to create a profile if the end of a work piece;(e)
Pocket milling, it is a type of end milling in which tool needs to cut shallow
into the work piece;(f) Surface contouring, which uses a ball-nose cutter along
a curvilinear trajectory to create a three-dimensional surface.

∙ Peripheral milling Peripheral milling generates a surface parallel to the ro-

tation axis. Both flat and formed surfaces can be produced by this method,
with a cross-section of the resulting surface corresponding to the axial contour
of the cutter. The process is sometimes referred to as ’slab’ milling. There are
various forms of peripheral milling: (a) slab milling, the basic form of peripheral
milling in which the diameter of the cutter is greater than the work part width;
(b) slot milling, in which the diameter of the cutter is less than the workpiece
width, creating a slot in the work when the cutter is very thin; (c) side milling,
in which the cutter machines the side of the workpiece; (d) straddle milling,
similar to side milling but the cutting action occurs on both sides of the item;
(e) form milling, in which the teeth of the milling cutter have a shape which
corresponds to the profile of the surface to be produced [25].

44
5.2.2 CNC milling machine

All the milling operations described above are used in conjunction with milling ma-
chines, which provide rotary movement to the cutters, and feed to the workpiece and
arrangement for clamping, automatic feed etc. Milling machines come in three basic
models:

∙ Horizontal Milling Machine

∙ Vertical Milling Machine

∙ Universal Milling Machine

CNC stands for computer numeric controlled. It refers to any machine tool (i.e.
mill, lathe, drill press, etc) that uses a device to electronically regulate the motion
of one or more axes on the machine.CNC Milling Machinery is an incredibly use-
ful milling machine for both commercial and industrial production. The aerospace
industry, the medical industry and the electronics industry can all benefit from the
products of CNC Milling.

Components of CNC milling machine

∙ Frame The frame protects the unit and provides stiffness to withstand cutting
forces. Usually, the base has a detachable column. CNC Milling Machine
Frames are most commonly made from cast iron. Other options include epoxy
granite filling and aluminum welding.

∙ Table The table is where some kind of workholding solution retains the work-
piece for machining. Most of the milling machine tables use T-Slots to connect
the workholder to the bench. By mounting a Fixture Plate on it you can make
the T-Slot table much more convenient and flexible.

∙ Spindle The spindle is the foundation of every milling machine. It consists of

a revolving taper assembly where the holders of the tool can be mounted.

45
 ∙ Axes The axes of the CNC Milling Machine make motion with Cartesian coor-
dinates programmed through g-code and manual jogging from the control panel.
Generally there are three axes corresponding to X, Y, and Z. The optional 4th
Dimension is a CNC Milling Machine accessory. Five axis milling machines are
feasible but not very popular in the DIY CNC world. The spindle rotates a
motor with an optional transmission of some kind.

∙ CNC Controller The CNC Controller is the brain of the system. It includes
the electronics that drive the axle motors to shift the axes. CNC Controllers
are responsible for receiving G-Code and manual inputs from the CNC Control
Panel and translating them to the correct signals to control the Axis Stepper
or Servo Motors.

Accessories of CNC milling machine

∙ Coolant CNC Coolant plays many roles in cutting and comes in various vari-
eties. There are many forms of coolant device typical to DIY CNC including:
– Mist Coolant – Flood Coolant – Air Blast or Cool Air Gun – Dust Collector
Although the Dust Collector does not "cool," it satisfies the primary coolant
that is the chip evacuation.

∙ Lubricating Mechanism CNC Milling Machines require lubrication for their

conductors and in particular, for their pathways. Ways are any device used to
cause axes to slip. Way Oilers can be either manual or automatic. Automatic
electric way oiler contributes dramatically to reliability by maintaining smooth
friction and less road wear.

∙ Powered Draw-bar Powered or Power Drawbars are a great comfort for CNC
Milling Machines and an absolute necessity if the Automatic Tool Changer is to
be used. They effectively allow the button to capture and release tool holders
from the spindle. There are three different styles, including Impact Wrench,
Tormach TTS Style and Pull Stud. Machines without a Power Drawing bar

46
 enable the operator to loosen the drawbar (or spindle-dependent collet) with a
key that is much longer than the push-button type.

∙ Enclosure CNC Milling Machines are fundamentally a messy beast. They’re

going to chuck chips away long ways. If they are fitted with a flood coolant, the
mess is much worse. With the Enclosure, the mess is contained inside, where it
doesn’t get all over the workshop.

Operational Parameters of CNC milling machine

∙ Spindle Speed of Revolution Spindle revolution speed determines the cut-

ting edge velocity relative to the workpiece, i.e. the cutting speed. Since cutting
speed has a significant influence on tool life, the selection of cutting speed is
closely related to the reliability of the tool. Too low or too high cutting speed
can lead to a drastic reduction in tool life. In the meanwhile, the speed of the
spindle revolution in the milling of thin-walled workpieces has a direct influence
on the stability of the cutting process. The spindle speed of the revolution
should therefore be chosen discreetly in the milling process.

∙ Cutting depth and Cutting width The cutting depth and the cutting width
are constrained by the spindle strength, the transmitting power of the machine
tool, the material type, the tool specifications, the coolant, the machining pro-
cess, and the rigidity of the machine tool-tool-workpiece device. And they’re
having a huge influence on tool life. They should then be chosen fairly ac-
cording to the machining standard, the machining efficiency, and the machining
process. Generally, machining performance is the first objective of roughing
machining, such that a greater cutting depth and a larger cutting width are
chosen. The consistency of the workpiece surface is the main objective of the
finishing process so that the depth of cutting and the width of cutting can be
minimized.

∙ Feed Rate The feed rate is the speed at which the cutting tool travels relative
to the workpiece in the milling process. In general, the linear feed rate is used in

47
 practical processing and is defined as feed per minute. The feed rate of milling
can directly impact machining precision, surface quality, workpiece deforma-
tion, and tool life. It is also limited by tool parameters, workpiece material,
tool direction, machine tool stiffness, and feed device efficiency.In the machin-
ing process, the feed rate of milling is selected based on component material,
geometry characteristics, quality specifications, and machine tool capacity.

∙ Tool material In milling operation cutting tools should have following char-
acteristics:

– Hardness is the ability of a material to resist deformation. Cutting tool

material must be harder than the material of workpiece. For this reason
hardness and stength of the cutting tool is maintained at relatively high
temperatures.

– Toughness is the ability of a material to absorb energy and plastically

deform without fracturing. Toughness is necessary in order for the cutting
tool not to chip or fracture, especially during the cutting operation.

– Wear Resistancemeans the attainment of acceptable tool life before tools

need to be replaced. Hardness is the most important property that resist
abrasive wear.

∙ Workpiece Material Mechanical properties of the workpiece material are im-

portant requirements influencing the working conditions. In the case of low
cutting forces, low hardness and tensile strength usually provide better machin-
ability.Common materials that are used in milling involve steel, cast iron, alu-
minium, nickel, zinc, magnesium, titanium and thermoset plastics. Properties of
the work material have a significant influence on the success of milling operation
and the selection of the workpiece material must take into consideration several
factors as cost, strength, resistance to wear and especially its machinability.

48
 Figure 5-1: CNC Milling Machine

5.3 Tool Wear Monitoring

During a milling operation the replacement of a worn tool is a major maintenance

required. Tool wear monitoring in very important and a delicate task to insure the on-
time tool replacement and moreover the quality of the finish product. Sensor-based
approaches, which are considered here, can be divided into mainly two methods in
which we classify tool wear monitoring: (a) Direct method;(b) Indirect Methods.
Direct methods calculate the individual values of such wear parameters (e.g. the
size of the wear area), whereas indirect methods measure relevant process parameters
that are associated with tool wear (e.g. cutting forces or vibrations). It is possible
to conclude the values of these method parameters on the current tool state using an
effective analytical or empirical model. Direct method measuring equipment is very
costly and these systems are vulnerable to errors due to environmental factors in the
machine tool (e.g. chips, coolant, etc.)
Another relevant classification criteria for sensor-based methods depends on the

49
monitoring period: significant parameters are measured in continuous or on-line meth-
ods during the cutting process, while significant parameters are measured in inter-
mittent or off-line methods only throughout the cutting process intervals.Altogether,
a trend towards indirect, continuous methods can be detected. These methods will
be investigated in the table.

50
Wear or process parameters Examples for measurement procedures
and transducers

Shape or position of the cutting edge or Direct methods

the wear area Shape or position of the cutting edge or
the wear area Measurement with opti-
cal methods (e.g. CCD camera or fi-
bre optic sensor) or integration of thin
film sensors into the coating of a cut-
ting tool
Volumetric overall loss of the tool Measurement of size and concentration
of wear particles in the coolant (and
electrochemical analysis) or measure-
ment of radioactivity (for specifically
prepared tools)
Changes of the electrical resistance at Voltage measurement at a specific, con-
the junction of tool and workpiece ductive tool coating
Changes of workpiece dimensions Dimension measurement by means of
micrometers or optical, pneumatic, ul-
trasonic, or electromagnetic transduc-
ers
Change of distance between tool (or Distance measurement by means of mi-
toolholder) and workpiece crometers, pneumatic gauges, displace-
ment transducers (e.g. inductive or ca-
pacitive), or ultrasonic sensors
Cutting forces Indirect Method
Force measurement with strain gauges
or piezoelectric sensors at the tool or
at (or in) the toolholder, piezoelec-
tric force measuring plates or rings at
the turret, force-measuring bearings,
torque measurement at the main spin-
dle
Electrical current, power, or energy Measurement of current or power con-
sumption of spindle or feed motors (e.g.
ampere meter or dynamometer)
Cutting temperature Temperature measurement by means
of thermocouples or pyrometers, re-
flectance of chip surface or chip color
Roughness of the machined surface Measurement with a mechanical stylus
or optical methods (e.g. CCD camera
or fibre optic sensor)
51
Table 5.1: Examples for direct and indirect tool wear sensing methods
52
Chapter 6

Data preparation

6.1 Data Description

Data is the essential component for training a machine learning algorithm. For the
purpose of tool condition monitoring data related to the operational features of CNC
milling machine is required. The data-set used here for the classification of the worn
and unworn tool taken from System-level Manufacturing and Automation Research
Testbed (SMART), University of Michigan. The data-set comprises of a series of
machining experiments were run on 2" x 2" x 1.5" wax blocks in a CNC milling
machine. There are total 18 experiments each experiment produced a finished wax
part with an "S" shape - S for smart manufacturing - carved into the top face, as
shown in Fig [6-1] Eight experiments were run with an unworn tool while ten were
run with a worn tool. Moreover experiments were run with pressures of 2.5, 3.0, and

Figure 6-1: ’S’ shaped test artifact

53
4.0 bar. The data-set obtained from the CNC milling machine is consist of two files.
One data-set contain the general data from 18 different experiments which are given
in the "train.csv" and includes following features Table [6-1]
The other part of data-set is a time series data was collected from 18 experiments
with a sampling rate of 100 ms and are separately reported in files experiment_01.csv
to experiment_18.csv. Each file has measurements from the 4 motors in the CNC (X,
Y, Z axes and spindle). The features available in the machining data-sets are listed
in table [6-2].

Material Feedrate Clamp pressure Tool Condition MachiningFinalize

wax 6 4 unworn yes

wax 20 4 unworn yes
wax 6 3 unworn yes
wax 6 2.5 unworn no
wax 20 3 unworn no

wax 6 4 worn yes

wax 20 4 worn no
wax 20 4 worn yes
wax 15 4 worn yes
wax 12 4 worn yes
wax 3 4 unworn yes
wax 3 3 unworn yes
wax 3 4 worn yes
wax 3 3 worn yes
wax 6 3 worn yes
wax 20 3 worn no
wax 3 2.5 unworn yes
wax 3 2.5 worn yes

Table 6.1: Machining data-set of a CNC milling

54
 Figure 6-2: Python Code: Data Cleaning and Labeling

6.2 Data Pre-processing

Data pre-processing is a data mining strategy that involves transforming raw data into
an usable form. Real-world data is often unreliable, contradictory and/or deficient in
certain behaviors or patterns and is expected to include a variety of mistakes. Pre-
processing data is an established way of tackling these problems. Notice that excessive
data pre-procesing can reduce the quality of data, so a careful data pre-processing
procedure will be adopted.

6.2.1 Data Cleaning and Labeling

Data-cleaning is the method of identifying and fixing (or deleting) corrupt or faulty
information from a record set, table or archive and refers to the detection of missing,
incorrect, corrupted or irrelevant parts of the data and the substitution, alteration or
deletion of contaminated or gross data. Data cleaning can be done interactively with
data wrangling tools, or as batch processing by scripting.

Data labeling is an important part of data pre-processing for ML, particularly for
supervised learning, in which both input and output data are labeled for classification
to provide a learning basis for future data processing.

55
6.2.2 Encoding Categorical Data

Feature encoding is one of the most important preprocessing steps in any machine
learning project. It is the method of converting categorical data into numerical data
in a data-set. We need to perform feature encoding since most models of machine
learning can only interpret numerical data in text form and not data. In our data-set
we have a "Machining Positioning" which is a categorical feature which is needed to be
converted into numerical values. OneHOTEncoder is used to convert the categorical
features into numerical one.

Figure 6-3: Python Code: OneHotEncoding

6.2.3 Feature Scaling

Feature scaling is essential for machine learning algorithm that tune all the features
in the data to a same scale. It refers to putting the values to the same scale so that
no feature is dominated by the other. There are two major scaling techniques which
are used for feature scaling.

Standardization

Standardization of a dataset is a common requirement for many machine learning

estimators: they might behave badly if the individual features do not more or less
look like standard normally distributed data (e.g. Gaussian with 0 mean and unit
variance). By computing the relevant statistics on the samples in the training set,
centering and scaling occur independently on each feature.

𝑧 = (𝑥 − 𝜇)/𝜎 (6.1)

56
For example, several elements used in a learning algorithm’s objective function (such
as the Support Vector Machines RBF kernel or linear model L1 and L2 regularizers)
assume that all features are centered around 0 and differ in the same order. If a
function has a variance that is greater than others in order of magnitude, it could
overpower the objective function and make the estimator unable to correctly learn
from other characteristics as expected.

Normalization

Database normalization is the structuring of the relational database, in order to reduce

the redundancy and improve data integrity. The aim of normalization is to adjust
the numeric column values in the dataset to use a common scale, without distorting
variations in the value ranges or losing information.
Min-max Normalization is one of the most common ways to normalize data.
For every feature, the minimum value of that feature gets transformed into a 0, the
maximum value gets transformed into a 1, and every other value gets transformed
into a decimal between 0 and 1.

𝑥𝑠𝑐𝑎𝑙𝑒𝑑 = (𝑥 − 𝑥𝑚𝑖𝑛 )/(𝑥𝑚𝑎𝑥 − 𝑥𝑚𝑖𝑛 ) (6.2)

There is one drawback of Min-max normalization that it does not handle outliers very
well.

6.2.4 Data Splitting

One of the first decision in developing a model is how to utilize the existing data. A
common technique to split the data is as training and test set. To build models and
feature sets, the training set is used; it is the substrate for estimating parameters,
comparing models, and all the other activities needed to achieve a final model. The
test set is used for estimating a final, unbiased evaluation of the model’s success only
at the end of these activities. It is critical that the test set not be used prior to this
point. Looking at the test sets results would bias the outcomes since the testing data

57
will have become part of the model development process. There is no rule of thumb
that how much data should be in the training and test sets. The proportion of data
can be driven by many factors, including the size of the original pool of samples and
the total number of predictors.

58
Wear or process parameters Examples for measurement procedures
and transducers

X1_ActualPosition actual x position of part (mm)

X1_ActualVelocity actual x velocity of part (mm/s)
X1_ActualAcceleration actual x acceleration of part (mm/s/s)
X1_CommandPosition reference x position of part (mm)
X1_CommandVelocity reference x velocity of part (mm/s)
X1_CommandAcceleration reference x acceleration of part
(mm/s/s)
X1_CurrentFeedback current (A)
X1_DCBusVoltage voltage (V)
X1_OutputCurrent current (A)
X1_OutputVoltage voltage (V)
X1_OutputPower power (kW)
Y1_ActualPosition actual y position of part (mm)
Y1_ActualVelocity actual y velocity of part (mm/s)
Y1_ActualAcceleration actual y acceleration of part (mm/s/s)
Y1_CommandPosition reference y position of part (mm)
Y1_CommandVelocity reference y velocity of part (mm/s)
Y1_CommandAcceleration reference y acceleration of part
(mm/s/s)
Y1_CurrentFeedback current (A)
Y1_DCBusVoltage voltage (V)
Y1_OutputCurrent current (A)
Y1_OutputVoltage voltage (V)
Y1_OutputPower power (kW)
Z1_ActualPosition actual z position of part (mm)
Z1_ActualVelocity actual z velocity of part (mm/s)
Z1_ActualAcceleration actual z acceleration of part (mm/s/s)

Table 6.2: Examples for direct and indirect tool wear sensing methods

59
Wear or process parameters Examples for measurement procedures
and transducers

Z1_CommandPosition: reference z po-

sition of part (mm)
Z1_CommandVelocity reference z velocity of part (mm/s)
Z1_CommandAcceleration reference z acceleration of part
(mm/s/s)
Z1_CurrentFeedback current (A) Z1_DCBusVoltage
voltage (V )
Z1_OutputCurrent current (A )
Z1_OutputVoltage voltage (V )
S1_ActualPosition actual position of spindle (mm )
S1_ActualVelocity actual velocity of spindle (mm/s )
S1_ActualAcceleration actual acceleration of spindle (mm/s/s
)
S1_CommandPosition reference position of spindle (mm )
S1_CommandVelocity reference velocity of spindle (mm/s )
S1_CommandAcceleration reference acceleration of spindle
(mm/s/s )
S1_CurrentFeedback current (A )
S1_DCBusVoltage voltage (V )
S1_OutputCurrent current (A )
S1_OutputVoltage voltage (V )
S1_OutputPower current (A )
S1_SystemInertia torque inertia (kg*m2 )

Table 6.3: Examples for direct and indirect tool wear sensing methods

60
Chapter 7

Methodology Proposed

The proposed approach is to classify the tool as worn and unworn as described in
figure 7-1

Figure 7-1: Python Code: Flow Chart

61
7.1 Data Pre-Processing

As explained in the chapter 6, the raw Data collected from System-level Manufac-
turing and Automation Research Testbed (SMART), University of Michigan is pre-
processed for the analysis.
The whole data-set obtained from the CNC milling machine is exported to python
by using pandas library. Firstly the required data from both the files is gathered and
integrated in a single data-frame. Tool condition is assigned to all 18 experiments
correspondingly. In the data cleaning part only those experiments are picked from
the data-set in which machining is finalized.
In the data-set tool condition is given by the categorical labels, for the ease ap-
plying the machine learning models the data has given binary labels i-e

worn unworn

0 1

7.2 Data Splitting

Firstly, the data splitting is done by using the train_test_split function of the scikit
learn library. It splits the whole data-set, the size of test set is given as 0.3 which
means 30% of the data will be in the test set and remaining 70% data will be in train
set randomly.

Figure 7-2: Python Code: Flow Chart

Secondly, the data is split manually in such a way that test set will take all the
experiments one by one and all remaining will go in the train set. Similarly another
split is done by assigning 2 consecutive experiments to the test set and remaining in

62
the train set. In this way accuracy of the models can be checked for several situations
instead of just one random split which was done firstly.

7.3 Feature Scaling

Feature scaling is done by standardization, using the StandardScaler function of scikit

learn library. Centering and scaling happen independently on each feature by com-
puting the relevant statistics on the samples in the training set. Mean and standard
deviation are then stored to be used on later data using transform.

Figure 7-3: Python Code: Feature Scaling by Standardization

While doing the standardization, the columns which are generated during the
OneHotEncoding are already in the binary form, so these columns are skipped from
the standardization.

7.4 Feature Extraction

Feature extraction is the process of transforming raw data into meaningful features
more suitable for a machine data mining task, which act as inputs for machine learning
algorithms and help in improving the overall predictive model performance. Gener-
ally, feature extraction starts from an initial set of measured data and builds derived
values (features) intended to be explanatory and essential, simplifying the subsequent
learning and modelling phases. A feature is simply an attribute on which the predic-
tion is done.

63
7.5 Model Application
Machine Learning models are trained for the data-set containing containing the pre-
processed data. Machine Learning algorithms which are usually used to classify the
tool condition are decision tree, random forest and neural network. They are applied
and accuracy of the model is tested by comparing the predicted values from the test
set values. Results of the analysis are given in the following chapter.

7.6 Principal Component Analysis

Principal component analysis is the mathematical algorithm that reduces the dimen-
sionality of the data by keeping most of the variations in the data. It reduces the
features by identifying the distances, which are called principal components, along
which the variation of the data is maximal. By using a few components, each sample
can be represented by relatively few numbers instead of by values for thousands of
variables.
Principal component analysis (PCA) is a technique that is useful for the compres-
sion and classification of data. The purpose is to reduce the dimensionality of a data
set (sample) by finding a new set of variables, smaller than the original set of variables,
that nonetheless retains most of the sample’s information. By information we mean
the variation present in the sample, given by the correlations between the original
variables. The new variables, called principal components (PCs), are uncorrelated,
and are ordered by the fraction of the total information each retains.

64
Chapter 8

Results

The criteria selected to evaluate the classification model is accuracy score and cross-
validation score.

∙ Accuracy Score: Accuracy is one metric for evaluating classification models.

Informally, accuracy is the fraction of predictions our model got right. Formally,
accuracy has the following definition:

𝑁 𝑢𝑚𝑏𝑒𝑟𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦𝑆𝑐𝑜𝑟𝑒 =
𝑇 𝑜𝑡𝑎𝑙𝑛𝑢𝑚𝑏𝑒𝑟𝑜𝑓 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠

∙ Cross-validation score: Cross-validation is a resampling procedure used to

evaluate machine learning models on a limited data sample.

The procedure has a single parameter called k that refers to the number of
groups that a given data sample is to be split into. As such, the procedure is
often called k-fold cross-validation. When a specific value for k is chosen, it
may be used in place of k in the reference to the model, such as k=10 becoming
10-fold cross-validation. In our reseach we have used 5-fold cross validation to
validate the machine learning models.

65
8.1 By using train-test split for 18 experiments
In the first analysis we have concatenated all eighteen machining experiments in a
single data-frame. After the concatenation of data data normalization, feature encod-
ing and data splitting is applied which is already explained in chapter 6. Dataframe
in this case includes 25286*49 of matrix. We are going to split this whole data frame,
firstly into dependant and independent variable, secondly we used the function of
train-test split from the model selection module of sklearn library.
After splitting the data-set into train-test sets, machine learning algorithms are
applied whose results are as under:

Machine Learning Algorithm Accuracy

Random Forest 99.193%

Decision Tree 98.956%

Neural Networks 88.755%

KNN 86.776%

kernel SVM 73.014%

Logistic Regression 53.732%

Table 8.1: By using train-test split

8.2 By using cross-validation approach

In the second analysis the data-set is distributed into training and test sets by creating
a loop. During this analysis CNC machine experiments are divided in such a way
that: 1) test set will get 1 experiment and the other 17 experiments will be in the
train set i-e when the loop starts experiment # 1 will go to test set and the remaining
other experiments will go to the training set. In the next loop experiment # 2 will
go to the test set and the other experiments will be in the training set. Moreover
this loop will go all over 18 times, providing us 18 accuracy values for the machine
learning model.

66
 2) test set will get 2 experiments and the other 16 experiments will be in train
set.i-e when the loop starts experiment # 1 and experiment # 2 will go to test set
and the remaining other experiments will go to the training set. In the next loop
experiment # 2 and experiment # 4 will go to the test set and the other experiments
will be in the training set. Moreover this loop will go all over 18 times, providing us
9 accuracy values for the machine learning model. Hence we will be getting 18 and 9
accuracies of one machine learning models at a time respectively.

In this method all possible classification accuracies in a model are extracted and
their mean value will be considered as the model accuracy.

This analysis is done by considering all 18 experiments. In those 18 experiments

8 experiments are with “unworn” tool and 10 experiments are with “worn” tool.

Machine Learning Algorithm Mean Accuracy Mean Cross Val-

idation Score

Random Forest 78.35% 90.50 %

(segmentation 17,1)

Random Forest (segmentation 66.47% 89.46%

16,2)

Decision Tree (segmentation 85.40% 89.49 %

17,1)

Decision Tree (segmentation 76.78% 90.08%

16,2)

Neural Networks (segmentation 44.63% 61.21%

17,1)

Neural Networks (segmentation 51.57% 54.11%

16,2)

Table 8.2: By using cross-validation approach

67
8.3 Analysis with machining finalized data
In the third analysis strategy followed is the same as first and second analysis but
experiments are reduced on the basis of finalization of machining operation. In this
analysis only those experiments are considered for which the machining operation in
finalized. Experiments are excluded on the basis of feature "Machining Finalize".
The analysis is done by considering 14 experiments for which machining is final-
ized. In those 14 experiments 6 experiments have unworn tool and other 8 experiments
have worn tool.

Machine Learning Algorithm Mean Accuracy Mean Cross Val-

idation Score

Random Forest 86.97% 95.74%

(segmentation 13,1)

Random Forest (segmentation 81.80% 95.49%

12,2)

Decision Tree (segmentation 89.89% 92.70%

13,1)

Decision Tree (segmentation 80.47% 93.60%

12,2)

Neural Networks (segmentation 49.07% 55.67%

13,1)

Neural Networks (segmentation 49.46% 56.71%

12,2)

Table 8.3: Analysis with Machining Finalized experiments

8.4 Feature extraction

Firstly, feature extraction is done by considering 14 experiments for which machining
is finalized. In those 14 experiments 6 experiments have unworn tool and other 8
experiments have worn tool. Mean, median, kurtosis and skewness are selected at the

68
same time. Iterations are done with a data-frame of 14×189.

Machine Learning Algorithm Mean Accuracy Mean Cross Val-

idation Score

Random Forest 42.85% 73.33%

(segmentation 13,1)

Random Forest (segmentation 42.85% 46.66%

12,2)

Decision Tree (segmentation 42.5% 46.66%

13,1)

Decision Tree (segmentation 50% 56.66%

12,2)

Neural Networks (segmentation 42.85% 50%

13,1)

Neural Networks (segmentation 28.57% 36.66%

12,2)

Table 8.4: Results after Feature Extraction

8.5 Feature Reduction

Feature reduction, also known as reduction of dimensionality, is the method of reduc-

ing the number of features in a heavy computing resource without losing essential de-
tails. Reducing the number of characteristics requires reducing the number of factors,
making it simpler and quicker for the machine to operate.There are several methods
by which feature reduction is done. Generalized discriminant analysis, auto-encoders,
non-negative matrix factorization, select-from-model, and principal component anal-
ysis are some of the most common. In our analysis we have used principal component
analysis and SelectfromModel techniques for applying feature reduction.

69
8.5.1 Principal component analysis

Principal component analysis is applied to the machine learning model by setting the
component variability as 95%. Results of principal component analysis are shown in
the table below.

Figure 8-1: Python Code: Principal Component Analysis

Machine Learning Algorithm Mean Accuracy Mean Cross Val-

idation Score

Random Forest 42.24% 62.25%

(segmentation 13,1)

Random Forest (segmentation 47% 57.12%

12,2)

Decision Tree (segmentation 50.3% 60.2%

13,1)

Decision Tree (segmentation 49.4% 57.2%

12,2)

Neural Networks (segmentation 46.8% 63.4%

13,1)

Neural Networks (segmentation 44.9% 55.8%

12,2)

Table 8.5: Principal Component Analysis

8.5.2 Feature Reduction by using SelectFromModel

SelectFromModel is an effective tool of scikit learn library, which selects features on

the basis of importance wiegths. Decision Tree and Random Forest models have a

70
built in method which generates the importance of the features according to their
weights in the data-set.
By applying SelectFromModel, features are reduced from 48 to 12. Results for
the analysis are given in the table below.

Machine Learning Algorithm Mean Accuracy Mean Cross Val-

idation Score

Random Forest 86.94% 96.60%

(segmentation 13,1)

Random Forest (segmentation 82.77% 96.56%

12,2)

Decision Tree (segmentation 88.08% 92.47%

13,1)

Decision Tree (segmentation 84.10% 94.86%

12,2)

Table 8.6: Feature Reduction by using SelectFromModel

71
72
Chapter 9

Conclusion, Limitations and Future

Recomendations

In the light of the above-stated results, we can conclude, that we have achieved the
objective of this research which was to classify the tool condition (worn, unworn)
during a milling operation, by using machine learning models. We have mainly used
the following models to classify the tool condition:

∙ Random forest

∙ Decision tree

∙ NeuralNeural networks

∙ K-Nearest Neighbor

∙ Kernel SVM

∙ Logistic Regression

Finally, looking at the evaluation parameters, mean accuracy scores, and mean cross-
validation scores we can conclude that Random forest, Decision tree, and Neural
network approach are providing more accurate predictions as compared to others.
While conducting the analysis with all 18 experiments concatenated to form a data-set
and splitting them by using train-test split method, we have acquired the predicting

73
accuracy of 99.193%. The machine learning model is efficient but assuming it in a
real case scenario it may be overfitting or giving biased results.
we have applied the cross-validation approach for getting the unbiased predicting
accuracies for all the experiments as test set, and giving a single mean accuracy value
for the analysis. This analysis also concluded with a higher mean accuracy value
for Random Forest algorithm, with an accuracy value of 78.35% and cross-validation
score of 90.50%. It shows that the for predicting the tool condition in a milling
operation random forest algorithm can be beneficial.
The research has concluded that the strength of features is more important than
the number of features for training a machine learning model. Principal component
analysis shows that our features are weak to extract any of them as the significant
features. It’s beneficial to have features who have a quite significant variation with
the tool condition at any point in time.
Machine learning framework as discussed above can play an important role in
decreasing the downtime during a milling process and in return can increase the
production efficiency and product quality. It could be possible by the implementa-
tion cyber-physical systems more precisely for attaining data-set with some strong
features.
Now-a-days for any machine learning model the main limitation is the availability
of the data-set with some strong features. Due to the COVID most of the companies
has closed or temporarily stopped their ongoing projects. For the same reason i did
not get the data-set which i was supposed to get from an italian car prototyping
company. The unavailability of the data with required features was the main hurdle
in my research.
Further improvements for the smart manufacturing system for tool condition and
machining process monitoring involves the development of effective data preparation
methodology to produce more general data in the corresponding machining process
and the advanced signal processing methods can be applied for feature generation.
Acoustic and vibration signals can be used for generating strong features. In addition
to the Neural Networks, convolution neural networks (CNN) and reinforced machine

74
learning techniques can be used to get more accurate predictions. Work also remains
regarding tool wear regression-based classification in place of discrete classification;
such an investigation will require close monitoring of tool condition at various points
during tool life. Another important development could be the prediction of the re-
maining useful life of the tool (RUL), which can be beneficial for optimizing the
maintenance operation.

75
76
Appendix

Phython Code

8.1

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd

frames = list()
results = pd.read_csv("train.csv")
for i in range(1,19):
exp = ’0’ + str(i) if i < 10 else str(i)
frame = pd.read_csv("experiment_.csv".format(exp))
row = results[results[’No’] == i]
frame[’TC’] = 0 if row.iloc[0][’tool_condition’] == ’worn’ else 1
frames.append(frame)
df = pd.concat(frames, ignore_index = True)
df.head()
df.replace(’Machining_Process’: ’Starting’:’Prep’,’end’:’End’,
inplace=True)
print(df[’Machining_Process’].value_counts().sort_index())

x = df.iloc[:,:-1].values

77
y = df.iloc[:,48].values

from sklearn.compose import ColumnTransformer

from sklearn.preprocessing import OneHotEncoder
ct = ColumnTransformer([(’encoder’, OneHotEncoder(), [47])], remainder=’passthrough’)
x= np.array(ct.fit_transform(x))

from sklearn.model_selection import train_test_split

x_train, x_test, y_train, y_test = train_test_split(x,y,test_size=0.25,random_state=0)

from sklearn.preprocessing import StandardScaler

sc_x = StandardScaler()
x_train[:,8:] = sc_x.fit_transform(x_train[:,8:])
x_test[:,8:] = sc_x.transform(x_test[:,8:])

from sklearn.tree import DecisionTreeClassifier

classifier = DecisionTreeClassifier(criterion= ’entropy’, random_state=0) classifier.fit(x_train,y_train)
from sklearn.model_selection import cross_val_score
clf = cross_val_score(classifier,x_train,y_train)
y_pred = classifier.predict(x_test)

from sklearn.metrics import confusion_matrix, accuracy_score

cm = confusion_matrix(y_test, y_pred)

print (cm)
ac = accuracy_score(y_test,y_pred)
print (ac)
imp= classifier.feature_importances_

78
8.2

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import random

imported all the libraries

frames = list()
results = pd.read_csv("train.csv")
for i in range(1,19):
exp = ’0’ + str(i) if i < 10 else str(i)
frame = pd.read_csv("experiment_.csv".format(exp))
row = results[results[’No’] == i]
if row.iloc[0][’machining_finalized’] == ’yes’:
frame[’TC’] = 0 if row.iloc[0][’tool_condition’] == ’worn’ else 1
frames.append(frame)
else:
i+=1

imported all experiments in which machining is finalized

giving the class labels worn = 0 and unworn = 1
appended all the experiments in the list called "frames"

accuracy = list()
cross_val_scorelist= list()

Def a variable "accuracy" in the list type to store the accuracy scores.
Def a variable "cross_val_scorelist" in the list type to store the cross val-

79
idation scores.

for c in range(0,14):
generated a "for loop" in the range 0-14 for spliting the train and test sets
in each iteration"
df_testlist = list()
df_trainlist = list()
Def two list type variables "df_testlist" and "df_trainlist" to store train
and test experiments at each instance.
for a in range (0,14):
if a==c:
print(’c’, a)
df_test= frames[a]
if a==0:
df_test.replace(’Machining_Process’: ’Starting’:’Prep’,’end’:’End’, inplace=True)
df_testlist.append(df_test)
appended the experiment which is test set for the instance to the list

else:
print(a,’ ’)
df_train= frames[a]
if a==0:
df_train.replace(’Machining_Process’: ’Starting’:’Prep’,’end’:’End’, inplace=True)
df_trainlist.append(df_train)
appended all other experiments in the train set list.
a+=1
print(’———’)
df = pd.concat(df_trainlist, ignore_index = True)
concatenated all the experiments in the train list for applying the model.

80
 x = df.iloc[:,:-1].values
assigned the columns for the independent variable x
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder
ct = ColumnTransformer([(’encoder’, OneHotEncoder(), [47])], remainder=’passthrough’)
x= np.array(ct.fit_transform(x))
OneHotEncoding is done to convert the categorial features into numerical
one.

y = df.iloc[:,48].values
assigned the column for dependent variable "Tool Condition"

x_test = df_test.iloc[:,:-1].values
x_test= np.array(ct.fit_transform(x_test))
y_test = df_test.iloc[:,48].values

Feature scaling by using Standardization

from sklearn.preprocessing import StandardScaler
sc= StandardScaler()
x= sc.fit_transform(x)
x_test= sc.transform(x_test)

Applying machine learning model to the data set i-e x and y will be
independent and dependent variables respectively.
Classifier is trained for the respective model (Random Forest, Decision
Tree, and Neural Networks) with the given dataset.
from sklearn.ensemble import RandomForestClassifier
classifier = RandomForestClassifier(n_estimators = 50, criterion= ’entropy’, ran-
dom_state= 0)
classifier.fit(x,y)

81
 cross_val_score returns score of test fold where cross_val_predict
returns predicted y values for the test fold
from sklearn.model_selection import cross_val_score
clf = cross_val_score(classifier,x,y)
cross_val_scorelist.append(clf)

Machine learning model(Random Forest) predict the value of y for the

corresponding x_test.
y_pred = classifier.predict(x_test)

Accuracy is one metric for evaluating classification models.

Informally, accuracy is the fraction of predictions our model got right.
from sklearn.metrics import accuracy_score
ac = accuracy_score(y_test,y_pred)
accuracy.append(ac)

c+=1

import statistics
mean accuracy score for all iterations.
mean= statistics.mean(accuracy)
print (mean)
mean cross validation score for all iterations.
cvm = statistics.mean(clf)
print (cvm)

82
8.4

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import random
imported all the libraries required.

means = list()
medians = list()
kurtosiss = list()
skewnesss = list()
stds=list()
frames = list()
results = pd.read_csv("train.csv")

imported all experiments in which machining is finalized giving the

class labels worn = 0 and unworn = 1 appended all the experiments in
the list called "frames"

for i in range(1,19):
exp = ’0’ + str(i) if i < 10 else str(i)
frame = pd.read_csv("experiment_.csv".format(exp))
row = results[results[’No’] == i]
mean = frame.mean(axis=0)
median = frame.median(axis=0)
kurtosis = frame.kurtosis(axis=0)
skewness = frame.skew(axis=0)
std= frame.var(axis=0)

Calculated the features (mean, median,kurtosis and skewness) for the

feature extraction

83
 mean[’TC’] = 0 if row.iloc[0][’tool_condition’] == ’worn’ else 1
means.append(mean)
median[’TC’] = 0 if row.iloc[0][’tool_condition’] == ’worn’ else 1
medians.append(median)
kurtosis[’TC’] = 0 if row.iloc[0][’tool_condition’] == ’worn’ else 1
kurtosiss.append(kurtosis)
skewness[’TC’] = 0 if row.iloc[0][’tool_condition’] == ’worn’ else 1
skewnesss.append(skewness)
frame[’TC’] = 0 if row.iloc[0][’tool_condition’] == ’worn’ else 1
frames.append(frame)
i+=1

making the lists of required features respectively.

df_me = pd.concat(means, axis=1 , join=’inner’)

df_md = pd.concat(medians, axis=1 , join=’inner’)
df_ku = pd.concat(kurtosiss, axis=1 , join=’inner’)
df_skw = pd.concat(skewnesss, axis=1 , join=’inner’)
df_std = pd.concat(stds, axis=1 , join=’inner’)

Concatenated all the features to convert list into a dataframe.

Taking transpose of features for further analysis.
df_me=df_me.transpose()
df_md=df_md.transpose()
df_ku=df_ku.transpose()
df_skw=df_skw.transpose()
df_std=df_std.transpose()
Concatenated all the features in a single dataframe. df = pd.concat([df_me.drop(’TC’,axis=1),d
axis=1 , join=’inner’)

84
 df.insert(0, "No", [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17], True)

Def a variable "accuracy" in the list type to store the accuracy scores.
Def a variable "cross_val_scorelist" in the list type to store the cross
validation scores.
accuracy = list()
cross_val_scorelist= list()

generated a "for loop" in the range 0-18 for spliting the train and test
sets in each iteration"
for c in range(0,18):

Def two list type variables "df_testlist" and "df_trainlist" to store

train and test experiments at each instance.
df_testlist = list()
df_trainlist = list()
for a in range (0,18):
if a==c:
print(’c’, c)
df_test= df[df["No"] == c]
df_test=df_test.drop(labels= ’No’,axis= 1)
df_testlist.append(df_test)

appended the experiment which is test set for the instance to the list
else:
print(a,’ ’)
df_train= df[df["No"] == a]
df_train= df_train.drop(labels=’No’,axis= 1)
df_trainlist.append(df_train)

85
a+=1
print(’———’)
me = pd.concat(df_trainlist, ignore_index = True)

concatenated all the experiments in the train list for applying the model

x = me.iloc[:,:-1].values

assigned the columns for the independent variable x

y = me.iloc[:,188].values
y=y.astype(’int’)
x_test = df_test.iloc[:,:-1].values

y_test = df_test.iloc[:,188].values
y_test=y_test.astype(’int’)

Feature scaling by using Standardization

from sklearn.preprocessing import StandardScaler
sc_x = StandardScaler()
x = sc_x.fit_transform(x)
x_test = sc_x.transform(x_test)

Applying machine learning model to the data set i-e x and y will be
independent and dependent variables respectively. Classifier is trained for
the respective model (Random Forest, Decision Tree, and Neural Net-
works) with the given dataset.

from sklearn.tree import DecisionTreeClassifier

classifier = DecisionTreeClassifier(criterion= ’entropy’, random_state=0)

86
classifier.fit(x,y)

cross_val_score returns score of test fold where cross_val_predict

returns predicted y values for the test fold

from sklearn.model_selection import cross_val_score

clf = cross_val_score(classifier,x,y)
cross_val_scorelist.append(clf)

Machine learning model(DecisionTreeClassfier) predict the value of y

for the corresponding x_test.
y_pred = classifier.predict(x_test)

from sklearn.metrics import accuracy_score

ac = accuracy_score(y_test,y_pred)
accuracy.append(ac)

c+=1
import statistics
Accuracy is one metric for evaluating classification models. Informally,
accuracy is the fraction of predictions our model got right.
mean= statistics.mean(accuracy)
print (mean)
mean cross validation score for all iterations.
cvm = statistics.mean(clf)
print (cvm)

87
8.5

import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import random

imported all the libraries

frames = list()
results = pd.read_csv("train.csv")
for i in range(1,19):
exp = ’0’ + str(i) if i < 10 else str(i)
frame = pd.read_csv("experiment_.csv".format(exp))
row = results[results[’No’] == i]
if row.iloc[0][’machining_finalized’] == ’yes’:
frame[’TC’] = 0 if row.iloc[0][’tool_condition’] == ’worn’ else 1
frames.append(frame)
else:
i+=1

imported all experiments in which machining is finalized

giving the class labels worn = 0 and unworn = 1
appended all the experiments in the list called "frames"

accuracy = list()
cross_val_scorelist= list()

Def a variable "accuracy" in the list type to store the accuracy scores.
Def a variable "cross_val_scorelist" in the list type to store the cross val-

88
idation scores.

for c in range(0,14):
generated a "for loop" in the range 0-14 for spliting the train and test sets
in each iteration"
df_testlist = list()
df_trainlist = list()
Def two list type variables "df_testlist" and "df_trainlist" to store train
and test experiments at each instance.
for a in range (0,14):
if a==c:
print(’c’, a)
df_test= frames[a]
if a==0:
df_test.replace(’Machining_Process’: ’Starting’:’Prep’,’end’:’End’, inplace=True)
df_testlist.append(df_test)
appended the experiment which is test set for the instance to the list

else:
print(a,’ ’)
df_train= frames[a]
if a==0:
df_train.replace(’Machining_Process’: ’Starting’:’Prep’,’end’:’End’, inplace=True)
df_trainlist.append(df_train)
appended all other experiments in the train set list.
a+=1
print(’———’)
df = pd.concat(df_trainlist, ignore_index = True)
concatenated all the experiments in the train list for applying the model.

89
 x = df.iloc[:,:-1].values
assigned the columns for the independent variable x
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder
ct = ColumnTransformer([(’encoder’, OneHotEncoder(), [47])], remainder=’passthrough’)
x= np.array(ct.fit_transform(x))
OneHotEncoding is done to convert the categorial features into numerical
one.

y = df.iloc[:,48].values
assigned the column for dependent variable "Tool Condition"

x_test = df_test.iloc[:,:-1].values
x_test= np.array(ct.fit_transform(x_test))
y_test = df_test.iloc[:,48].values

Feature scaling by using Standardization

from sklearn.preprocessing import StandardScaler
sc= StandardScaler()
x= sc.fit_transform(x)
x_test= sc.transform(x_test)

from sklearn.decomposition import PCA

pca = PCA(0.95)
pca.fit(x)
x= pca.transform(x)
x_test= pca.transform(x_test)
imported PCA from sklearn.decomposition for applying Principal Com-
ponent Analysis.
the algorithm will select the number of components while preserving 95%

90
of the variability in the data.

Applying machine learning model to the data set i-e x and y will be
independent and dependent variables respectively.
Classifier is trained for the respective model (Random Forest, Decision Tree,
and Neural Networks) with the given dataset.
from sklearn.ensemble import RandomForestClassifier
classifier = RandomForestClassifier(n_estimators = 50, criterion= ’entropy’, ran-
dom_state= 0)
classifier.fit(x,y)

cross_val_score returns score of test fold where cross_val_predict re-

turns predicted y values for the test fold
from sklearn.model_selection import cross_val_score
clf = cross_val_score(classifier,x,y)
cross_val_scorelist.append(clf)

Machine learning model(Random Forest) predict the value of y for the correspond-
ing x_test.
y_pred = classifier.predict(x_test)

Accuracy is one metric for evaluating classification models.

Informally, accuracy is the fraction of predictions our model got right.
from sklearn.metrics import accuracy_score
ac = accuracy_score(y_test,y_pred)
accuracy.append(ac)

c+=1

import statistics

91
mean accuracy score for all iterations.
mean= statistics.mean(accuracy)
print (mean)
mean cross validation score for all iterations.
cvm = statistics.mean(clf)
print (cvm)

92
Bibliography

[1] Arinez, J. F., Chang, Q., Gao, R. X., Xu, C., and Zhang, J. (August 13, 2020).
"Artificial Intelligence in Advanced Manufacturing: Current Status and Future Out-
look." ASME. J. Manuf. Sci. Eng. November 2020; 142(11): 110804.
[2] Lee, J., Wu, F., Zhao, W., Ghaffari, M., Liao, L., and Siegel, D., 2014, “Prog-
nostics and Health Management Design for Rotary Machinery Systems —Reviews,
Methodology and Applications,” Mech. Syst. Signal Process., 42(1–2), pp. 314–334.
[3] ge.com/research/project/predictive-maintenance.
[4] Lecun, Y., Bengio, Y., and Hinton, G., 2015, “Deep Learning,” Nature, 521(7553),
pp. 436–444.
[5] Gao, R., Wang, L., Teti, R., Dornfeld, D., Kumara, S., Mori, M., and Helu, M.,
2015, “Cloud-Enabled Prognosis for Manufacturing,” CIRP Ann., 64(2), pp. 749–772.
[6] R. Teti, K. Jemielniak, G. O’Donnell, and D. Dornfeld, 2010, “Advanced monitor-
ing of machining operations,” CIRP Ann-Manuf Techn, 59(2), pp. 717-73.
[7] Arinez, J. F., Chang, Q., Gao, R. X., Xu, C., and Zhang, J. (August 13, 2020).
"Artificial Intelligence in Advanced Manufacturing: Current Status and Future Out-
look." ASME. J. Manuf. Sci. Eng. November 2020; 142(11): 110804.
[8] X.S. Si, W. Wang, C.H. Hu, M.Y. Chen, and D.H. Zhou, 2013, “A wiener-process-
based degradation model with a recursive filter algorithm for remaining useful life
estimation,” Mechanical Systems and Signal Processing, 35(1), pp. 219-237.
[9] M. Dong and D. He, 2007, “Hidden semi-Markov model-based methodology for

93
multi-sensor equipment health diagnosis and prognosis,” European Journal of Oper-
ational Research, 178(3), pp. 858-878.
[10] Saha, K. Goebel, and J. Christophersen, 2009, “Comparison of prognostic algo-
rithms for estimating remaining useful life of batteries,” T I Meas Control.
[11] M. E. Orchard and G. J. Vachtsevanos, 2009, “A particle-filtering approach for
on-line fault diagnosis and failure prognosis,” T I Meas Control.
[12] B. Sick, 2002, “On-line and indirect tool wear monitoring in turning with artificial
neural networks: a review of more than a decade of research,” Mechanical Systems
and Signal Processing, 16(4), pp. 487-546.
[13] Kothuru, A., Nooka, S. P.,and Liu, R. (August 3,2018). "Audio-Based Tool
Condition Monitoring in Milling of the Workpiece Material With the Hardness Vari-
ation Using Support Vector Machines and Convolutional Neural Networks." ASME.
J.Manuf. Sci. Eng.November 2018; 140(11):111006.
[14] Sohyung Cho, Shihab Asfour, Arzu Onar, Nandita Kaundinya, Tool breakage
detection using support vector machine learning in a milling process, International
Journal of Machine Tools and Manufacture, Volume 45, Issue 3,2005,
[15] Clayton Cooper, Peng Wang, Jianjing Zhang, Robert X. Gao, Travis Roney, Ihab
Ragai, Derek Shaffer, Convolutional neural network-based tool condition monitoring
in vertical milling operations using acoustic signals, Procedia Manufacturing, Volume
49, 2020,
[16] Wu, Zhenhua. "Cutting Tool Condition Monitoring and Prediction Based on
Dynamic Data Driven Approaches." Proceedings of the ASME 2015 International
Manufacturing Science and Engineering Conference. Volume 1: Processing. Char-
lotte, North Carolina, USA. June 8–12, 2015
[17] Li, X., et al. Fuzzy neural network modelling for tool wear estimation in dry
milling operation. in Annual conference of the prognostics and health management
society. 2009.

94
[18] D. Wu, C. Jennings, J. Terpenny and S. Kumara, "Cloud-based machine learning
for predictive analytics: Tool wear prediction in milling," 2016 IEEE International
Conference on Big Data (Big Data), Washington, DC, 2016, pp. 2062-2069
[19] Kurfess, T., Saldana, C., Saleeby, K., and Parto-Dezfouli, M. (August 28, 2020).
"INDUSTRY 4.0 AND INTELLIGENT MANUFACTURING PROCESSES: A RE-
VIEW OF MODERN SENSING TECHNOLOGIES." ASME. J. Manuf. Sci. Eng
[20] I. Dumitrache and S. I. Caramihai, "The Enterprise of Future as a Cyber-Physical
System," IFAC Proceedings Volumes, vol. 46, no. 9, pp. 1310-1315, 2013/01/01/2013
[21] Alothman, Hussam Ali, Khasawneh, Mohammad T., and Nagarur, Nagen N. "In-
ternet of Things in Manufacturing: An Overview." Proceedings of the ASME 2018
International Mechanical Engineering Congress and Exposition. Volume 2: Advanced
Manufacturing. Pittsburgh, Pennsylvania, USA. November 9–15, 2018
[22] W.Z. Khan, M.H. Rehman, H.M. Zangoti, M.K. Afzal, N. Armi, K. Salah, Indus-
trial internet of thing: Recent advances, enabling technologies and open challenges,
Computers Electrical Engineering, Volume 81,2020, 106522,ISSN 0045-7906,
[23] Bergonzi, L, Colombo, G, Rossoni, M, Furini, F. "Data and Knowledge in
IIoT-Based Maintenance Application." Proceedings of the ASME 2017 International
Mechanical Engineering Congress and Exposition. Volume 11: Systems, Design, and
Complexity. Tampa, Florida, USA. November 3–9, 2017
[24] Zheng, P., wang, H., Sang, Z. et al. Smart manufacturing systems for Industry
4.0: Conceptual framework, scenarios, and future perspectives. Front. Mech. Eng.
13, 137–150 (2018).
[25] Jim Davis, Thomas Edgar, James Porter, John Bernaden, Michael Sarli, Smart
manufacturing, manufacturing intelligence and demand-dynamic performance, Com-
puters Chemical Engineering, Volume 47, 2012,
[26] A. Cachada et al., "Maintenance 4.0: Intelligent and Predictive Maintenance
System Architecture," 2018 IEEE 23rd International Conference on Emerging Tech-

95
nologies and Factory Automation (ETFA), Turin, 2018,
[27] Li, Y, Jia, D. "A Real-Time Analysis of Condition-Based Maintenance in a
Multistage Production System." Proceedings of the ASME 2018 13th International
Manufacturing Science and Engineering Conference. Volume 1: Additive Manufactur-
ing; Bio and Sustainable Manufacturing. College Station, Texas, USA. June 18–22,
2018
[28] Yang, Y, Gao, RX, Fan, Z, Wang, J, Wang, L. "Cloud-Based Prognosis: Per-
spective and Challenge." Proceedings of the ASME 2014 International Manufactur-
ing Science and Engineering Conference collocated with the JSME 2014 International
Conference on Materials and Processing and the 42nd North American Manufacturing
Research Conference. Volume 1: Materials; Micro and Nano Technologies; Proper-
ties, Applications and Systems; Sustainable Manufacturing. Detroit, Michigan, USA.
June 9–13, 2014.
[29] Hassan, M., Sadek, A., Attia, M. H., and Thomson, V. (December 18, 2017). "A
Novel Generalized Approach for Real-Time Tool Condition Monitoring." ASME. J.
Manuf. Sci. Eng. February 2018
[30] M. P. Groover, Fundamentals of modern manufacturing: Materials, Processes,
and Systems,1996.
[31] Z. Ge, Z. Song, S. X. Ding and B. Huang, "Data Mining and Analytics in the
Process Industry: The Role of Machine Learning," in IEEE Access, vol. 5, pp. 20590-
20616, 2017, doi: 10.1109/ACCESS.2017.2756872.
[32] Rehan Sadiq, Manuel J. Rodriguez, Haroon R. Mian, Empirical Models to Predict
Disinfection By-Products (DBPs) in Drinking Water: An Updated Review, Editor(s):
Jerome Nriagu, Encyclopedia of Environmental Health (Second Edition), Elsevier,
2019, ISBN 9780444639523,
[33]

96