University of Louisville

ThinkIR: The University of Louisville's Institutional Repository

Electronic Theses and Dissertations

5-2018

Maintainability analysis of mining trucks with data

analytics.
Abdulgani Kahraman
University of Louisville

Follow this and additional works at: https://ir.library.louisville.edu/etd

Part of the Other Computer Engineering Commons

Recommended Citation
Kahraman, Abdulgani, "Maintainability analysis of mining trucks with data analytics." (2018). Electronic Theses and Dissertations. Paper
2932.
https://doi.org/10.18297/etd/2932

This Master's Thesis is brought to you for free and open access by ThinkIR: The University of Louisville's Institutional Repository. It has been accepted
for inclusion in Electronic Theses and Dissertations by an authorized administrator of ThinkIR: The University of Louisville's Institutional Repository.
This title appears here courtesy of the author, who has retained all other copyrights. For more information, please contact [email protected].
MAINTAINABILITY ANALYSIS OF MINING TRUCKS WITH DATA
ANALYTICS

By

Abdulgani Kahraman
B.S., Sakarya University, 2011

A Thesis
Submitted to the Faculty of the
J.B. Speed School of Engineering
University of Louisville
In Partial Fulfillment of the Requirements
for the Degree of

Master of Science in Computer Science

Department of Computer Engineering and Computer Science

University of Louisville
Louisville, Kentucky

May 2018
Copyright © 2018 by Abdulgani Kahraman

All rights reserved

MAINTAINABILITY ANALYSIS OF MINING TRUCKS WITH DATA
ANALYTICS

By

Abdulgani Kahraman
B.S., Sakarya University, 2011

A Thesis Approved on

April 24, 2018

by the following Thesis Committee:

__________________________________
Dr. Mehmed Kantardzic, Thesis Director

__________________________________
Dr. Adel Elmaghraby

___________________________________
Dr. James E. Lewis

ii
 ACKNOWLEDGMENTS

I would like to thank my advisor Dr. Mehmed Kantardzic for his great contributions

and support in this study.

I would also like to thank my brother Dr. Mustafa Kahraman for his priceless

support to this research study.

I take this opportunity to thank Dr. Adel Elmaghraby and Dr. James Lewis for

serving on my committee, reading this thesis and providing comments during

defense.

Finally, I am grateful to my family for their love and support, especially my mother.

iii
 ABSTRACT

MAINTAINABILITY ANALYSIS OF MINING TRUCKS WITH DATA

ANALYTICS

Abdulgani Kahraman

April 24, 2018

The mining industry is one of the biggest industries in need of a large budget, and

current changes in global economic challenges force the industry to reduce its

production expenses. One of the biggest expenditures is maintenance. Thanks to

the data mining techniques, available historical records of machines’ alarms and

signals might be used to predict machine failures. This is crucial because repairing

machines after failures is not as efficient as utilizing predictive maintenance.

In this case study, the reasons for failures seem to be related to the order of signals

or alarms, called events, which come from trucks. The trucks ran twenty-four hours

a day, seven days a week, and drivers worked twelve-hour shifts during a nine-

month period. Sequential pattern mining was implemented as a data mining

methodology to discover which failures might be connected to groups of events,

and SQL was used for analyzing the data.

According to results, there are several sequential patterns in alarms and signals

before machine breakdowns occur. Furthermore, the results are shown differently

depending on shifts’ sizes. Before breakdowns occur in the last five shifts a hundred

percent detection rates are observed. However, in the last three shifts it is observed

less than a hundred-percentage detection rate.

iv
 TABLE OF CONTENTS

1 Introduction ............................................................................................................................ 1
1.1 Recent Technological Developments in Mining Industry .............................................. 1
1.1.1 Vehicle Health Monitoring System in Mining Industry ........................................ 3
1.2 Big Data in Mining Industry .......................................................................................... 6
1.3 Maintenance ................................................................................................................... 7
1.3.1 Corrective Maintenance ......................................................................................... 7
1.3.2 Preventive Maintenance ......................................................................................... 7
1.3.3 Predictive Maintenance .......................................................................................... 8
1.3.4 The Cost of Maintenance in Mining Industry ........................................................ 9
1.3.5 Predictive Maintenance and Cost ........................................................................... 10
2 Big Data Analytics and Predictive Maintenance in the Mining Industry: A Literature
Review ........................................................................................................................................... 13
3 Sequence Mining Methodology for Predictive Maintenance of Mining Trucks ................... 17
3.1 Sequential Pattern Mining.............................................................................................. 18
3.2 Categories of Patterns .................................................................................................... 21
3.2.1 Periodic Patterns .................................................................................................... 21
3.2.2 Approximate Patterns............................................................................................. 22
3.2.3 Statistically Significant Patterns ............................................................................ 23
3.3 Dataset Details ............................................................................................................... 24
3.3.1 Primary Features in Dataset and Summarization of General Processes................. 24
3.3.2 Preprocessing of Dataset and Features Selection ................................................... 27
3.4 Implementation of the Sequential Pattern Mining ......................................................... 29
4 Experimental Results: Frequent Sequences .......................................................................... 33
4.1 Results for Failure code 1104 ........................................................................................ 33
4.1.1 Pattern Results for Last 3 Shifts ............................................................................. 33
4.1.2 Pattern Results for Last 5 Shifts ............................................................................. 37
4.2 Results for Failure code 1140 ........................................................................................ 41
4.2.1 Pattern Results for Last 3 Shifts ............................................................................. 42
4.2.2 Pattern Results for Last 5 Shifts ............................................................................. 47
4.3 Results for Failure code 1143 ........................................................................................ 50
4.3.1 Pattern Results for Last 3 Shifts ............................................................................. 50
4.3.2 Pattern Results for Last 5 Shifts ............................................................................. 55
5 Conclusion............................................................................................................................. 60

v
REFERENCES .............................................................................................................................. 62
CURRICULUM VITAE………………………………………………………………….65

vi
 LIST OF FIGURES

Figure 1 - Some recent mining techniques (www.uky.edu, n.d.) .................................................................................... 13

Figure 2- Antennas on the mining trucks [www.highservice.com, n.d.] ......................................................................... 15
Figure 3- Type of maintenance ....................................................................................................................................... 19
Figure 4- Breakdown of direct mining costs in large open pit mines in the US (Fekete, 2015) ...................................... 20
Figure 5- The Data Mining Process (Kantardzic, & Zurada, 2005) ................................................................................ 28
Figure 6- Definitions of Features .................................................................................................................................... 36
Figure 7- An Example of Breakdowns on the Status Table ............................................................................................ 38
Figure 8- An example of Events Column on the Machine Health Status Table .............................................................. 38
Figure 9- Small part of event data columns after preprocessing for reason 1140 ............................................................ 40
Figure 10- An example of patterns for breakdown reason 1140 ..................................................................................... 41
Figure 11- Patterns for failure code 1104, comparing last three shifts and confidence percentages ............................... 44
Figure 12- Sequence group distributions according to breakdown orders for failure code 1104 in last three shifts ....... 45
Figure 13- Sequence group distributions according to breakdown orders for failure code 1104 in last three shifts above
70% ................................................................................................................................................................................. 46
Figure 14- Graphical illustration of patterns and confidence for failure code 1104 in last three shifts ........................... 46
Figure 15- Patterns for failure code 1104, comparing last five shifts and confidence percentages ................................. 49
Figure 16- Sequence group distributions according to breakdown orders for failure code 1104 in last five shifts ......... 50
Figure 17- Graphical representation of sequence group distributions according to breakdown orders for failure code 1104
in last five shifts .............................................................................................................................................................. 50
Figure 18- Graphical illustration of patterns and confidence for failure code 1104 in last five shifts ............................. 51
Figure 19- Patterns for failure code 1140, comparing last three shifts and confidence percentages ............................... 53
Figure 20- Sequence group distributions according to breakdown orders for failure code 1140 in last three shifts ....... 54
Figure 21- Sequence group distributions according to breakdown orders for failure code 1140 in last three shifts above
70% ................................................................................................................................................................................. 56
Figure 22- Graphical illustration of patterns and confidence for failure code 1140 in last three shifts ........................... 56
Figure 23- Patterns for failure code 1140, comparing last five shifts and confidence percentages ................................. 58
Figure 24- Sequence group distributions according to breakdown orders for failure code 1140 in last five shifts ......... 59
Figure 25- Graphical representation of sequence group distributions according to breakdown orders for failure code 1140
in last five shifts .............................................................................................................................................................. 59
Figure 26- Graphical illustration of patterns and confidence for failure code 1140 in last five shifts ............................. 60
Figure 27- Patterns for failure code 1143 with comparing last three shifts and confidence percentages ........................ 61
Figure 28- Sequence group distributions according to breakdown orders for failure code 1143 in last three shifts above
60% confidence............................................................................................................................................................... 63
Figure 29- Graphical illustration of patterns and confidence for failure code 1143 in last three shifts ........................... 64
Figure 30- Patterns for failure code 1143, comparing last five shifts and confidence percentages ................................. 66
Figure 31- Sequence group distributions according to breakdown orders for failure code 1143 in last five shifts ......... 67
Figure 32- Graphical representation of sequence group distributions according to breakdown orders for failure code 1143
in last five shifts .............................................................................................................................................................. 68

vii
Figure 33- Graphical illustration of patterns and confidence for failure code 1143 in last five shifts ............................. 69

viii
 1 INTRODUCTION

In this study, first, the basic mining techniques’ information related to the mining industry

is mentioned. Second, recent technological developments and big data developments in the

mining industry are presented. Third, the maintenance costs for big vehicles and how these

maintenance expenses might be reduced, thanks to data mining, is considered. Lastly, as a

case study, sequential pattern mining is used as a methodology on the dataset for predictive

maintenance and results are shared.

1.1 Recent Technological Developments in Mining Industry

Since early civilization, individuals have utilized mining strategies to extract minerals from

the soil. The citizens of ancient civilizations were all interested in mining. In the past,

mining was slow-going and unsafe. As time has progressed, society has created more

secure and exact strategies for finding and revealing substances found in the soil.

In the beginning, diggers employed primitive devices for burrowing. Mining shafts were

burrowed out by hand, and the entire process was exceptionally long. Inevitably,

individuals started using fire to clear burrows and reach more prominent profundities at a

quicker rate. Amid the 1600s, diggers began utilizing explosives to break up expansive

rocks. Motorized mining apparatuses, such as drills, would not be invented for a few more

decades, and it was not until the Industrial Revolution began in the 1700s that mineworkers

started improving the explosives they operated and created more progressive mining gear,

such as drills, lifts and steam-powered pumps (www.generalkinematics.com, 2015).

1
In today’s technologically-advanced society, mining strategies are continuously

progressing. For instance, improving surface mining procedures, diggers are presently able

to extricate over 85 percent of minerals and 98 percent of metallic minerals without digging

a shaft or imperiling the lives of workers (www.generalkinematics.com, 2015). Newly-

developed machines utilized for grinding and crushing can extricate minerals from the soil

with less energy than ever before.

Miners still use several techniques, such as explosives, trucks, drills and bulldozers,

particularly if they must dig deep into the soil. In any case, innovations have permitted

mineworkers in uncovering minerals with more exactness and less harm to the

encompassing environment. More proficient apparatuses can be utilized to decrease energy

consumption and increase the sum of minerals or metals gathered from the shaft.

Mining has made the world more modern compared to the past, but the threats of mining

have resulted in the deaths of numerous laborers. As innovation progresses, mining

procedures have indeed become more precise and productive. In the future, with

technological developments, it is possible to mine for materials with fully automatic

machines thanks to the Industry 4.0 system.

Figure 1 shows several mining techniques which are separated mainly into two categories:

underground mining and surface mining. While underground mining contains drift, slope,

and shaft mining, surface-mining methods involve area, contour, mountaintop removal,

and auger which are in Figure 1 (www.uky.edu, n.d.). Moreover, there are different kinds

of machines in the mining fields, and these machines are expensive vehicles and keeping

these vehicles in working conditions is vital for companies. Nowadays, thanks to the

monitoring systems, companies can follow and record these machines’ statuses in every

2
moment, and it provides opportunities to take precautions before machine breakdowns

occur.

Figure 1 - Some recent mining techniques (www.uky.edu, n.d.)

1.1.1 Vehicle Health Monitoring System in Mining Industry

With technological developments companies try to record and follow their vehicles. Big

vehicle production companies have their own monitoring systems for watching and taking

control of their big vehicles, and these systems present a lot of advantages for mining

companies (Viger, 2017). The Monitoring System is a recent technology for big and

expensive vehicles, which are consistently worked for numerous hours. When the vehicles

break down, it takes serious costs to repair them. Furthermore, since repairing these big

machines take a significant amount of time, this completely influences the machine’s

availability. Hence, mining machines, especially large ones, are required to reduce the

number of failures and enable operations without intrusion (Murakami, Saigo, Ohkura,

Okawa & Taninaga, 2002). In order for machines to continuously work, it is essential to

3
identify any problems and changes in status by physical examination early on that will

signal the maintenance staff to take reasonable measures without waiting for total failures

of valuable equipment.

Vehicle Health Monitoring System (VHMS) is one example of monitoring systems

available to mining companies, and it is suitable to integrate with existing systems. With

Internet of Things (IoT) every company has an opportunity to collect data, and they can

analyze this data to set more efficient work schedules. Particularly, it is extremely crucial

for companies which have big vehicles, because these kinds of vehicles need expensive

maintenance. As a result of this, analyzing their past records and taking several precautions

are vital not only for companies but also for employees because these precautions can

provide safer workplaces. Currently, many companies have their own databases which

were created by several monitoring systems and IoT.

4
 Figure 2- Antennas on the mining trucks [www.highservice.com, n.d.]

This figure shows the location of antennas, radars and cameras. The last few decades of

technological developments gave a priceless opportunity for companies. Currently,

companies can benefit from technological devices with more affordable prices and it is

possible to follow every step in the workplace. Now, expensive vehicles can be equipped

with high-tech devices such as antennas, GPS, etc., and companies can remotely watch

these vehicles and collect every signal and alarm from them and take crucial actions before

harmful breakdowns occur. For these big trucks, left and right-side mirrors are not enough

for seeing around the trucks, so several cameras are attached to the trucks. Radar and

antennas provide location information and prevent accidents in the mining fields.

Furthermore, there are several chips and sensors on various parts of the trucks which send

5
data to the main monitoring data center, so that companies have enough data about their

trucks’ conditions.

1.2 Big Data in Mining Industry

Big data is advanced data made by the action of computers, portable phones, implanted

frameworks and other organized gadgets. Such information became more predominant as

innovations such as radio frequency identification (RFID) and telematics progressed

(Rouse, 2014). Moreover, machine information has increased utilization of the Internet of

Things(IoT), and other big data management technologies that have been developed.

Big data, where different sensors and equipment create more structured data about their

operations, performances or conditions, and can be used to complete various analyses, such

as process optimization, improved maintenance or machine-to machine communication

(Fekete, 2015).

This thesis proposes plan maintenance and shows the importance of big data and the

analysis of data. Big data is an inevitable reality for every industry, including the mining

industry because prices of ores are very changeable and machines, which are used for

mining, are expensive to maintain. For this reason, collecting records of production and

machine status are vital for mining companies to make more efficient production plans.

Therefore, it is crucial for companies’ futures to record and analyze this data and arrange

their maintenance and production techniques. Furthermore, almost every big mining

company is using machine data, and they save all useful information for their future

operations.

6
Today, thanks to big data and IoT, companies could collect every transaction and details

such as drivers’ names, vehicles’ location, status, weight, speed and so on for each

production step. Moreover, this data can be analyzed, and used to make better and more

productive work schedules and create less risky work places.

1.3 Maintenance

There is an increasing pressure on companies, urged by worldwide competition, to

streamline operations involving item and item-related manufacturing system design, item

manufacturing and system maintenance. Maintenance activities are ordinarily performed

first by integration of maintenance and process engineering functions, then by application

of machines and hardware, and finally, through proactive actions on those machines and

equipment including preventive and predictive maintenance (Bastos, Lopes, & Pires,

2012). In literature, it is possible to find three nonspecific sorts of maintenance: Corrective

Maintenance, Preventive Maintenance, and Predictive Maintenance.

1.3.1 Corrective Maintenance

Corrective Maintenance(CM) actions are not schedulable, and this makes them harder to

plan for and costlier to perform. It is usually not the preferred maintenance because it

occurs suddenly and costs valuable money and time (Adolfsson & Dahlström, 2011).

Corrective maintenance is used when a system or machine fails. It includes repair and

replacement of failed parts to make machines active again.

1.3.2 Preventive Maintenance

Preventive Maintenance (PM) aims at maintaining equipment in satisfactory operating

conditions, and is fulfilled by providing for systematic control, detection, and correction of

7
incipient failures, before they cause great defects and usually a PM’s planning is created

according to equipment manufacturers’ advices (www.revolvy.com, n.d.).

Preventive maintenance is routine and tends to follow planned schedules to prevent

equipment and machinery breakdowns. The work is preemptively carried out on equipment

in order to avoid its breakdown. Despite the benefits of this maintenance, it is not efficient

for companies because it costs valuable money and time and does not allow for the use of

parts of machines that are capable of working. Even when machines do not work, they may

still contain parts that can last their full lifetimes, but preventive maintenance does not

calculate for these situations.

1.3.3 Predictive Maintenance

Predictive Maintenance (PDM) has a significant difference from the other maintenance

types. During regular operation to reduce failures, PDM directly monitors the status and

performance of equipment and provides an opportunity to take precautions before machine

failures (www.emaint.com, 2017). Although PDM is more complex compared to the

others, it provides several advantages thanks to monitoring. PDM reduces maintenance

cost, unnecessary preventative maintenance, unplanned maintenance and provides more

efficient work.

This next figure shows three main maintenance types and their definitions. Predictive

maintenance is the most efficient maintenance. However, corrective maintenance is the

costliest maintenance type.

8
 Figure 3- Type of maintenance

1.3.4 The Cost of Maintenance in Mining Industry

Currently, in every industry, mechanization is used for more effective production, and

some of the most expensive equipment belongs to the mining industry. For example,

according to the United States Census the largest absolute increase in the mining industry’s

capital investments occurred from 2006 to 2015 (up $75.4 billion or 75.9 percent) (US

Census Bureau., 2017). The Electric Power Research Institute (EPRI) has calculated

comparative maintenance costs for different maintenance techniques in US dollars per

horsepower (HP) per year. Researchers found that a preventive (scheduled) maintenance

strategy is the most expensive to run at $24.00 per HP. A corrective (reactive) maintenance

strategy is the second most costly at $17.00 per HP but has the additional cost of

compromising safety. Maintaining a 750 HP motor with a scheduled maintenance strategy

would cost approximately $18,000 per year, while a reactive maintenance strategy would

cost $12,750 a year, according to the EPRI study (www.ni.com, 2015). For example,

according to Caterpillar (CAT), mining trucks have between 2000 and 5000 HP

(www.cat.com, n.d.). When companies calculate this maintenance cost for a mining truck

9
according to the HP calculation, it becomes apparent how costly expenses can become for

the companies. The next figure provides an example that shows the percentage of

maintenance costs in pit mines for the US.

Figure 4- Breakdown of direct mining costs in large open pit mines in the US (Fekete, 2015)

According to Fekete, explanation of this figure: “A mining site’s cost structure consists of

three main parts: maintenance, labor and consumables. Figure 4 shows that maintenance

related costs account for about one-third of total operational costs (Campbell & Reyes-

Picknell, 2006), which makes maintenance the largest controllable cost. It includes items

such as replacement parts, human resources, supplies and other items (Lewis & Steinberg,

2001). Though it demonstrates the expenditures of North American open pit mines, other

locations and mine types show similar proportions. Mining companies can focus on

improving maintenance processes with advanced technologies (such as big data and

connected machines), as this area is the largest contributor for their operational

expenditures. “(Fekete, 2015).

1.3.5 Predictive Maintenance and Cost

After industrial revolution, companies focus on more profit, less expenses and safer

workplaces. For this reason, maintenance, which is one of the biggest expenses for

10
companies, has become more significant. Furthermore, researchers have begun to discover

more efficient maintenance techniques, and currently predictive maintenance has become

more crucial for companies. Many of these crucial reasons for improving mining

equipment reliability and maintainability are summarized as follows: (Peng & Vayenas,

2014).

 to maximize profit

 to reduce the cost of poor reliability/maintainability

 to reduce the use of mining equipment services in an unplanned manner because of

short notice

 to provide more accurate short-term forecasts for equipment operating hours

 to overcome challenges imposed by global competition

 to take advantage of lessons learned from other industrial sectors such as aerospace,

defense, and nuclear power generation

 to improve workplace safety

Bastos claims that 99 percent of machine failures are identified by a few pointers and

fulfilling organizations’ requirements leads to heavy expenses in maintenance systems, and

maintenance, considered non-value adding, which is continuously evaluated for cost

reduction, keeping the machines in excellent working condition (Bastos, Lopes & Pires,

2014). The main goal of predictive maintenance is to find the optimal time for needed

maintenance before harmful events may occur and reduce maintenance cost.

Most of the failures do not happen instantaneously, and more often than not there are a few

sorts of degradation processes or indications of transitions from typical states to failures.

Subsequently, the genuine conditions and their trends ought to be surveyed and anticipated

11
amid the degradation handle, and fitting maintenance actions ought to be taken some time

before a breakdown occurs. This is the fundamental target of predictive maintenance.

12
 2 BIG DATA ANALYTICS AND PREDICTIVE MAINTENANCE IN THE
MINING INDUSTRY: A LITERATURE REVIEW

Machinery diagnostics and maintenance are tremendous and diverse, primarily due to a

wide variety of production systems. Dozens of papers on this subject, which include

theories and applicable methods, show up each year in academic journals, conferences and

technical reports. In this section, several papers are summarized.

Like other companies, information technology is crucial for mining companies and

recently, mining companies have collected huge data by using new high-tech devices such

as GPS, monitoring systems, and fleet management systems, etc. As a result of this,

currently, companies can analyze these datasets (Yildirim & Dessureault, 2007).

“Big Data in the Mining Industry” was a paper was written by Fekete (2015). In the article,

Fekete discusses the challenges of the mining industry, such as, the fast dropping prices of

commodities and the technological developments that have forced companies to update

their structures. Thanks to the Internet of Things (IoT), several mining companies now

have the opportunity to collect big datasets to improve operations and efficiency.

Furthermore, huge maintenance expenses force mining companies to make more

reasonable maintenance schedules. According to Fekete, after interviewing with some

mining companies from Australia, it was concluded that predictive maintenance, Big Data,

IoT and Data Analytics create safe workplaces, provide efficient maintenance and decrease

maintenance expenditures (Fekete, 2015).

Cartella et al. (2014) defines a different approach for predictive maintenance, called

Hidden Semi-Markov Models (HSMMs), and they discuss the theoretical formalization of

13
the model, as well as a few experiments performed using both simulated and factual

information with the aim of technique approval. In this paper, all performed tests are able

to accurately appraise the current state of the machinery and viably foresee a predefined

event with generally low normal absolute error. According to these research results, the

model’s appropriateness to real-world settings can be advantageous, particularly where, in

real-time, the Remaining Useful Lifetime (RUL) of a machine can be calculated, and

results show that HSMM would be beneficial for condition monitoring and gauging useful

lifetime applications (Cartella, Lemeire, Dimiccoli & Sahli, 2014).

Sihong Peng et al. (2014) presents a study about the implementation of genetic algorithms

on Underground Mining Equipment as a case study for predictive maintenance. They

assumed that failures of mining equipment caused by an array of factors followed the

biological evolution theory. A software was created for predictive maintenance according

to their dataset and according to their opinion, these failures follow the natural

advancement theory. They used several case studies to focus on practical investigations of

a Load Haul Dump (LHD) vehicle with two different terms: three and six months.

According to their prediction case studies, a factual test is carried out to look at the

similitude between the anticipated data set with the real-life data set in the same period.

This research aims at comparing real data to the prediction of this software and analyzes

how successful genetic algorithms would be successful for prediction of maintenance. As

a result, these two different time interval studies are investigated, and they did not show

major impacts of chronological sequence in their prediction results (Peng et al. 2014).

A recent case study named “Earthmoving trucks condition level prediction using neural

networks” from Greece by Marinelli et al. (2014) presents an artificial neural network

14
(ANN) model that predicts earthmoving trucks’ condition levels using basic predictors.

The results are compared to the respective predictive accuracy of the statistical method of

discriminant analysis (DA). In this research, it is created an ANN-based predictive model,

and used the capacity, age, kilometers travelled and maintenance level of trucks, which

were collected from 126 earthmoving trucks. As a result, data processing identifies

specifically a connection between kilometers travelled and maintenance level with the

earthmoving trucks’ condition grade. Furthermore, they found that the predictive

performance of the proposed ANN model is very high for the validation process, and

similar findings from the application of DA to the same data set using the same predictors.

These models reached above 92 percent accuracy for prediction of trucks’ condition level.

As a result of that, the prediction decreases downtime, and its reverse influences

earthmoving duration and cost, meanwhile increasing the maintenance and replacement

policies’ impressiveness. This research shows that a sound condition level prediction for

earthmoving trucks is achievable through the utilization of easy to collect data and provides

a comparative evaluation of the results of two widely applied predictive methods

(Marinelli, Lambropoulos, & Petroutsatou, 2014).

A theoretical study by Chen et al. (2016), predicted faults from the data acquisition and

fusion strategies, and it used the fault prediction method based on full-vector spectrum

which belongs to Dr. Bently and Dr. Muszynska. According to this method, the uncertainty

of the spectrum structure can be extracted by the designed data acquisition and fusion

method. This method also shows that the reliability of the diagnosis on fault character was

improved, and it gives the technical foundation for the prediction and diagnosis research

of the fault characters (Chen, Han, Lei, Cui, & Guan, 2016).

15
Ullah et al. (2017) used one type of Machine Learning Methods for Predictive

Maintenance. According to this paper, there were several reasons for increasing the internal

temperature of electrical instruments, specifically contact issues, irregular loads, cracks in

insulation, defective relays, terminal junctions and other similar issues. As a result of these

reasons, they caused intrusive failures and potential damage to power equipment. In this

paper, the authors explained the initial prevention mechanism for power substations using

a computer-vision approach by taking advantage of infrared thermal images. This work

included a total of 150 thermal pictures of different electrical equipment in 10 different

substations in operating conditions, using 300 different hotspots. They used multilayer

perceptron (MLP) to classify the thermal conditions of components of power substations

into defect and non-defect classes. The performance of MLP reached 84 percent of

accuracy with graph cut and this result showed the benefit of the proposed defect analysis

approach (Ullah, Yang, Khan, Liu, Yang, Gao, & Sun, 2017).

16
 3 SEQUENCE MINING METHODOLOGY FOR PREDICTIVE MAINTENANCE

OF MINING TRUCKS

With recent technological developments, collecting and analyzing big data is even vital for

predictive maintenance in companies. There are several techniques for analyzing big data.

In this thesis, sequential pattern mining techniques is used to make predictions about

maintenance timing for mining trucks. The experimental results confirm that sequential

pattern mining is suitable approach for discovering information relevant to machine

failures.

“Data mining is a process of discovering various models, summaries, and derived values

from a given collection of data.” (Kantardzic, 2005). Data mining, which is one of the basic

processes of Knowledge Discovery in Database (KDD), is the procedure of extracting

hidden knowledge or patterns (non-trivial, implicit, previously unknown and potentially

useful) from large information warehouses (Zhao & Bhowmick -2003). The next figure

shows data mining’s main steps.

17
 Figure 5- The Data Mining Process (Kantardzic, & Zurada, 2005)

According to the steps in Figure 5, first state the problem which considers the reasons for

machine failures. Second, data is collected for nine months from eleven mining trucks

which belong to one mining company in North America. Afterwards, for more accuracy,

the dataset must be cleaned, and must be eliminated from useless, duplicate data as a third

step. Forth, for estimating model, which would be more useful for the dataset. Lastly,

implement this technique and evaluate the results of the data mining process.

3.1 Sequential Pattern Mining

Sequential Pattern Mining is one of the most important mining techniques for analyzing

big data. Sequential Pattern Mining (SPM) focuses on finding patterns that occur

consecutively in a database or patterns which would be related to time or other values; the

main aim is to discover related sequence patterns (Chueh, 2010). Implementation of SPM

is very broad and can be used for efficient maintenance of vehicles, natural disasters, sales

record analysis, marketing strategies, shopping sequences, medical treatments and DNA

sequences, etc.; the subsequences and frequent relevant patterns from the given data can

18
also be found by SPM (Ubaidulla, Sushmitha, & Vanitha, 2017). The aim of pattern mining

is to discover useful, recent and unforeseen patterns in databases. A sequence database

includes several sequences.

For instance, consider the following database:

Name Sequences
Seq1 a,(b,c,d),(f,g)
Seq2 (a,d),c,b,(a,b,f)
Seq3 c,(a,d,e,f)
Seq4 d,g,a,e,b,b
Seq5 (c,e,g),(a,b)

This database includes four sequences which are seq1, seq2, seq3, seq4 and seq5. For this

example, take into consideration that the symbols “a”, “b”, “c”, d”, “e”, “f”, “g” and “h”

symbolize different items sold in a supermarket, and “a” could be an “almond”, “b” could

be a “box of cereal”, etc.

Now, a sequence is an ordered list of sets of items. For this example, suppose that each

sequence shows what a customer bought in a supermarket. Consider the second sequence

“seq2”. This sequence shows that the second customer bought items “a” and “d” together,

then bought item “c”, then bought “b”, and then bought “a”, “b”, and “f” together.

For the dataset, like this example, there are several failure codes and many alarms and

signals which would be indicators named detection groups for these specific breakdowns’

codes. To try to discover several patterns called rules, and afterwards the results will be

compared these patterns according to the counts for the last three and the five shifts before

machine failures occur. The strength of a detection rate is measured by its support and

confidence and calculation of these: (Lui Zhang-2000)

19
For example, an example of support, when consider two items A and B, and it can calculate

the frequency of the item in the dataset. If consider a basket containing 7 items (3-oranges,

4-lemons) then support of any precise value can be calculated by the rate of number of

occurrences to the total number of items in the basket (i.e., support(oranges) = 3/7).

An example of confidence, this explains how likely B is purchased when A is purchased.

This defines association between two items. For example, when a person buys tea is more

likely to buy sugar as well or vice versa. This is measured by the proportion of transactions

with item X, in which item Y also appears. As a formulization:

“The support of a rule, X → Y, is the percentage of transactions in T that contains X ∪ Y

and can be seen as an estimate of the probability, Pr(X∪Y). The rule support thus

determines how frequent the rule is applicable in the transaction set T. Let n be the number

of transactions in T. Let n be the number of transactions in T.

The support of the rule X → Y is calculated as follows:

𝐶𝑜𝑢𝑛𝑡(𝑋∪𝑌)
Support =
𝑛

Support is a useful measure because if it is too low, the rule may just occur due to chance.

Furthermore, in a business environment, a rule covering too few cases (or transactions)

may not be useful because it does not make business sense to act on such a rule (not

profitable).

20
The confidence of a rule, X → Y, is the percentage of transactions in T that contain X also

contain Y, and can be seen as an estimate of the conditional probability, Pr(Y | X). It is

computed as follows:

𝐶𝑜𝑢𝑛𝑡(𝑋∪𝑌)
Confidence =
𝐶𝑜𝑢𝑛𝑡 𝑋

Confidence thus determines the predictability of the rule. If the confidence of a rule is too

low, one cannot reliably infer or predict Y from X. A rule with low predictability is of

limited use.” (Lui- Zhang 2000).

3.2 Categories of Patterns

Sequential patterns can be divided into three main categories: periodic patterns, statistically

significant patterns, and approximate patterns. However, there are more varieties of models

for sequential patterns in the literature (Esmaeili, & Fazekas, 2010).

3.2.1 Periodic Patterns

According to Slimani & Lazzez, the main purpose of this model is to discover occurrences

of repeated patterns in data and to try to predict future characteristics of real situations;

however, this model has several disadvantages, for example, misalignment might cause us

to miss some interesting and crucial patterns. Experts have shown, as a solution for this

restriction, a pattern might be filled partly to make the model more flexible. As an example,

in the series ({a}{b}{c}{a}{b}{c}{a}{b}{c}), the pattern {a}{b}{c} is a periodic pattern

because it is repeated with a period equal to three. Each status in the pattern shows the

periodicity, and this previous pattern is called a full periodic pattern. As an example, in a

sequence like this ({a}{b}{c}{b}{a}{c}{a}{b}{a}{a}{c}{b}), a pattern {a}*{b} where *

21
is a wide range of items, there is no full periodic pattern with length 3, and this is called a

partial periodic pattern (Slimani & Lazzez, 2014).

3.2.2 Approximate Patterns

In the real world, there are noisy data in almost every big data, and it is necessary to reduce

the effects of these kinds of data on results. Another different and more flexible method is

approximate patterns, which has a specific calculation with a compatibility matrix:

“For solving the problem of finding approximate patterns, the concept of compatibility

matrix is introduced [4]. This matrix provides a probabilistic connection from observed

values to the true values. Based on the compatibility matrix, real support of a pattern can

be computed. Table 2 gives an example of the compatibility matrix.

Table 2 Compatibility Matrix

For example, an observed I4 corresponds to a true occurrence of I1, I2, I3, and I4 with

probability C(I1,I4)=0.05 , C(I2,I4)=0.10 , C(I3,I4)=0.10 , and C(I4,I4)=0.75 ,respectively.

Compatibility matrix usually is given by some domain expert but there are some ways to

obtain and justify the value of each entry in the matrix so that even with a certain degree

of error contained in matrix, sequential pattern mining algorithm can still produce results

of reasonable quality.

22
A new metric, namely match is defined to quantify the significance of a pattern. The

combined effect of support and match may need to scan the entire sequence database many

times. Similar to other data mining methods, to tackle this problem sampling-based

algorithms can be used.

Consequently, the number of scans through the entire database is minimized.” (Esmaeili,

& Fazekas, 2010).

3.2.3 Statistically Significant Patterns

The calculation of support and confidence are crucial for sequential pattern mining, but

just using these supporting values as a standardization measure might cause one to skip

several important patterns; various data mining applications have tried to find a valid

solution for this problem (Slimani & Lazzez, 2014). On the other side, according to

Esmaeili & Gabor (2010), the number of occurrences (support) might be misleading, and

there is no direct ratio between a repetitive number of patterns and a significance of

patterns. Because of this, in several situations, many occurrences of an expected frequent

pattern may not be as important as a few occurrences of an expected uncommon pattern,

which is called surprising pattern instead of frequent pattern. In addition, the support

threshold must be set very low to discover a small number of patterns with high information

gain, and the information gain metric might be helpful to evaluate the degree of surprise of

the pattern (Esmaeili, & Fazekas, 2010).

The next step is deciding k most significant patterns, and this can be easily achieved by

using a threshold value and the best k patterns that have an information gain greater than

the specified threshold should be returned; however, the problem of the information gain

value is difficult to define the location of the occurrences of the patterns (Slimani & Lazzez,

23
2014). In a statistically significant method, the calculation of info gains and dataset and

other values are crucial to get better results. For example:

Two input patterns’ sequences such as:

S1=({a}{b}{c}{b}{a}{b}{d}{c}{a}{b}{b}{d}),and

S2=({b}{c}{d}{b}{a}{b}{a}{b}{a}{b}{d}{c}), then the pattern {a}{b} has the same

(three times occurred) information gain in the two sequences, it is dispersed in S1 but

repeats consecutively in S2 (Slimani & Lazzez, 2014).

3.3 Dataset Details

In this thesis, the data was collected for nine months by eleven mining trucks which belong

to a mine in North America. The dataset was provided by a mining engineer, Mustafa

Kahraman, who works with the mine. In this dataset, the trucks worked twenty-four hours

a day, seven days a week and every shift represented a twelve-hour period. The dataset has

more than three million rows and more than one hundred columns. This dataset included

three main tables: status, production, and machine health status.

First, the status table is where equipment statuses are recorded: fail status, standby,

production etc. Second, the production table is where truck cycles are saved with all

associated details: shift date, driver name, material type, speed, location and so on. Lastly,

the machine health status table is created by the machine health information for selected

trucks generated by chips and sensors which are connected to different parts of the trucks.

3.3.1 Primary Features in Dataset and Summarization of General Processes

These next tables show the short definitions of primary features for the dataset. Experts

created all these definitions and designed the database. After showing these definitions,

24
these are next steps, making feature selection, select necessary parts of data, eliminate

duplicate and missing values, and implement sequential pattern mining techniques. In all

these processes, SQL Server Management Studio will be used as a tool. The next figures

show some initial columns and short definitions of features for the dataset. In this thesis

the main focused feature is that events column which is created by alarms and signals.

25
Figure 6- Definitions of Features

26
3.3.2 Preprocessing of Dataset and Features Selection

Before analyzing the data, as preprocessing steps, first removed most of the features that

were not going to be used in the analysis. Furthermore, there are Non-Applicable (N/A)

values and missing values. These would not be useful for the analysis and removed this

data. In addition, duplicate data was deleted before implementing data mining techniques.

In making the decision to select table features, it was consulted an expert who worked in

this mining company and is knowledgeable about these trucks and the database. According

to the expert, there would be a lot of distinctive features which may have an impact on

breakdowns of trucks, such as driver mistakes, overloading, locations, and so on.

Furthermore, according to this expert, apart from these reasons there are several chips and

sensors which record the trucks’ status, signals, changes and alarms. After receiving this

information, it was decided to analyze more mechanical failures which may be indicative

of alarms and signals, which are called events in the database. Other identifying features

need more in-depth research, truck expertise, and workplace-specific knowledge.

First, it will be used the Status Table which includes the shift number, time, date, reason

number, category and so on. The category column shows the machine status and when it

shows a breakdown status (when category equals 4), as an example to select one reason

code, for example 1140, and a related time, date, and shift number. Afterwards, it will be

combined this data with the events column from the Machine Health Status. Figure 6 is as

an example of the Status Table:

27
 Figure 7- An Example of Breakdowns on the Status Table

As a next step, it will be selected related events’ column data from the Machine Health

Status Table, which was collected from signals and chips before and during the same

failures’ status. In this table, the events column is the most important column for the

analyses because after selecting related events, it will be preprocessed this data in an

attempt to discover some patterns. Figure 7 is as an example of the Machine Health Status

table:

Figure 8- An example of Events Column on the Machine Health Status Table

28
There are 35 different defined machine breakdown reasons in the dataset, and it will be

implemented the same processes for three different failure codes which are related to more

mechanical problems according to the expert. Next, it will be discovered several patterns

between the same breakdown codes consecutively, and after that it will be discovered

several patterns in the last three and five shifts’ events column data occurring before the

shift in which machine failures occur for the same failure codes. After that, it will be

analyzed how many times these patterns occurred and compare pattern numbers with the

last three and five shifts, which would be related failures according to these patterns’

confidence calculation values.

3.4 Implementation of the Sequential Pattern Mining

Discovering unexpected and useful patterns in databases is the fundamental data mining

task. In recent years, a trend in data mining has been to design algorithms for discovering

patterns in sequential data. One of the most popular data mining techniques for finding

patterns is sequential

pattern mining. It consists of discovering interesting subsequence patterns in a set of

sequences, where the remarkable subsequences can be measured in terms of various criteria

such as their occurrence frequency, length, and so on (Viger, 2017).

After deciding which sequential pattern mining technique would be a more appropriate

technique for analyzing the alarms/signals, SQL Management Studio will be used as a tool,

which has several functions to find patterns with T-SQL. It will be tried to discover some

patterns within each breakdown reason after ordering them based on their date. For

example, figure X shows the short part of event data between two 1140 failure codes, and

it will be checked this data to find 2,3,4, or 5 groups of sequential patterns. Next, it will be

29
compared it with the last three and five shifts’ patterns. If there are any specific patterns in

these last three shifts and five shifts, then it would be a reasonable indicator of this specific

failure code. Additionally, the same processes will be implemented for four different

breakdown codes.

Event Def
0701FFE0 OEM interface timeout
0201FFE0 OEM interface timeout
0B010009 PSC Event Number Changes
0B010000 Active Event Number Changes
0B011519 Tach Right Front - Zero while truck mov
2010006 Running without load
0B01000C Propel Restricted
0B01000A Drive Status Normal
2010005 Stopped without load
0B010014 Service BRK > 8mph
0B01FFE0 OEM Interface Timeout
0B01FFE1 OEM Interface Normal
0201FFE0 OEM interface timeout
0B010009 PSC Event Number Changes
0B010000 Active Event Number Changes
0B011519 Tach Right Front - Zero while truck mov
0B010014 Service BRK > 8mph
2010006 Running without load
2010005 Stopped without load
2010000 Dipper
0B01000C Propel Restricted
0B01000A Drive Status Normal

Figure 9- Small part of event data columns after preprocessing for reason 1140

Figure 8 shows a small part of the data which comes from the preprocessed data. The first

column is the main column, named the event column, which shows codes for events. The

second column, named def, illustrates definitions of events. In the event column, the rows

30
 which are bold and red in color, illustrate a simple example of one pattern for this data.

Normally there are more than seven thousand rows of data after preprocessing between the

same reason codes (code 1140). However, this example is a small part of this data and

contains one pattern example. An example of finding a pattern for this small data result is

illustrated in the next table. These patterns are a combination of common elements with

varied sizes. These patterns occur in the same place and in the same shifts before the same

failure code.

Pattern Groups Pattern All Last 3 Confidence Sequences Pattern

Size Counts Shifts Percentage No Groups

0201FFE0 ,0B010009 2 2 2 2 100% s1

0201FFE0 ,0B010009 ,0B010000 3 2 2 2 100% s1
0201FFE0 ,0B010009 ,0B010000
,0B011519 4 2 2 2 100% s1

Figure 10- An example of patterns for breakdown reason 1140

In this table, the first column shows elements of patterns, the second column is a pattern

size, meaning how many events are included in this pattern, and the third column belongs

to the last three shifts before machine failure occurs. For the fourth column, it shows how

many times this pattern occurred in the last five shifts before the machine failure. The fifth

column shows how

many times, this pattern is found between the same failure reasons. The sixth column is the

confidence value which is calculated by dividing the last three or five shifts’ number of

patterns by the “all count” pattern size. Lastly, the seventh column shows that all these

patterns include common events which means they occurred in the same place in the

dataset, and they have varied sizes and elements. It will be made sequential groups for

them, as illustrated by the last column. After this process, a new table is to show which
31
sequence groups are related to which number of failures. The failure code 1140 occurred

five times, and after finding patterns, it will be separated every group of sequence related

to their common elements. It will be done the same process for four different breakdowns,

after which will be illustrated the results according to their confidence values.

32
 4 EXPERIMENTAL RESULTS: FREQUENT SEQUENCES

In this section, it will be shared the pattern results with several graphical illustrations. It

will be done the same processes for three different failure codes which are more related to

mechanical breakdowns apart from other causes. Mechanical breakdowns have more

expensive maintenance costs and are more related to alarms and signals which referred to

as events in the dataset. The codes are 1104,1140 and 1143. First of all, it will be shown

the groups’ patterns data which are above a minimum of 70 percent confidence. Next, it

will be illustrated these results as graphs according to percentages.

4.1 Results for Failure code 1104

It will be separated each failure code result into two parts according to the last shift

numbers, which are the last three and last five shifts.

4.1.1 Pattern Results for Last 3 Shifts

The next figure belongs to the last three shift patterns and is between two of the same

failure codes called “all counts”. It was calculated these pattern groups’ confidence

percentages and ordered them from largest to lowest values for 1104 breakdown codes.

33
Pattern Groups Pattern All Last Confidence Sequences
Size Counts 3Shifts Percentage No
0E010004 ,3E010002 ,0E01FFE1 ,0E010009 4 2 2 100.00% s1
0E010004,3E010002,0E01FFE1,0E010009,0E010008 5 2 2 100.00% s1
0E010005,0E010004,3E010002,0E01FFE1,0E010009 5 2 2 100.00% s1
3E010000,0E01FFE1,0E01000B,0E01000A,0E010000 5 2 2 100.00% s2
3E010002 ,0E01000B ,0E01000A ,0E010009 4 2 2 100.00% s3
3E010002,0E01000B,0E01000A,0E010009 ,0E010008 5 2 2 100.00% s3
0E010006 ,0E01000B ,0E01000A ,0E010000 4 5 4 80.00% s4
0E010006,0E01000B,0E01000A,0E010000 ,0E010001 5 5 4 80.00% s4
0E010007,0E010006,0E01000B,0E01000A ,0E010000 5 5 4 80.00% s4
0E010002 ,0E010003,0E010009 ,0E010008 ,0E010007 5 4 3 75.00% s5
0E010003 ,0E010009 ,0E010008 ,0E010007 4 4 3 75.00% s5
0E010003,0E010009 ,0E010008 ,0E010007 ,0E010006 5 4 3 75.00% s5
0E010008 ,0E010007 ,0E010006 ,0E010005 4 4 3 75.00% s5
0E010008,0E010007 ,0E010006 ,0E010005 ,0E010004 5 4 3 75.00% s5

Figure 11- Patterns for failure code 1104, comparing last three shifts and confidence percentages

After making groups for these patterns, it will be created a table for sequence groups which

have a hundred percent confidence and related failure times. For example, groups of s1

occurred before the third and seventh breakdown for 1104. It will be shown this whole

groups of sequences and their related order of failures in the next two figures. The first one

will be illustrated it as a table, and second one will be as a graphical according to failure

timing and sequence numbers.

34
 Fault
Faults No
No 100%
f7
f1
f6
f2 s3
f5
f3 s1,s2
f4
f4 f3
f5 f2
f6 f1
f7 s1,s3 Sequence
Table s1 s2 s3 No

Graph

Figure 12- Sequence group distributions according to breakdown orders for failure code 1104 in last three shifts

In this table, it can be seen there are seven faults for reason 1104. Before failures occur,

there are three different sequential pattern groups. S3 occurred before the second failure

and s1, s2 before the third breakdown and s1, s3 before the seventh breakdown in last three

shifts. It can be seen these information as a table and graph in Figure 12.

When detection rate is calculated which equals by dividing counts of filled rows in the

table by fault occurring times. For this example, detection rate equals the count of filled

rows in the table which is 3 divided by total faults’ count (7), so it comes to a 43% detection

rate. This percentage means that the system can detect 4 out of 10 breakdowns situations.

For more accuracy, decreasing the confidence value to 70%, which might increase the

detection rate. The next figure shows sequence groups between 70% and 100% and a

distribution of these groups as a table and graph.

35
 Faults Fault
No >70% No
f1 f7
f2 s4,s5 f6
f3 s4,s5 f5
f4 f4
f5 f3
f6 f2
f7 f1
Sequence
s1 s2 s3 s4 s5 No
Table Graph

Figure 13- Sequence group distributions according to breakdown orders for failure code 1104 in last three shifts
above 70%

From this table it can be seen s4 and s5 occurred before the second and third failures, and

this did not change the detection rate. As it can be seen from Table in Figure 12; s1, s2 and

s3 occurred before faults 1 and 2.

The below graph of these patterns was created according to percentages of confidence.

Comparing Confidence and Counts of Pattern Groups

120%
Confidence Percentages

100%
80%
60%
40%
20%
0%
3 4 5 5
Counts of Sequence Groups

Figure 14- Graphical illustration of patterns and confidence for failure code 1104 in last three shifts

For confidence values from 100% to 70%, the count of patterns slightly increased because

when the confidence percentage threshold is decreased, it increased the number of patterns.

36
4.1.2 Pattern Results for Last 5 Shifts

The next table belongs to the last five shift patterns and is between two of the same failure

codes called “all counts”, and it is calculated these pattern groups’ confidence percentages

and ordered them from largest to lowest values for 1104 breakdown codes. This time,

obviously, there are more pattern groups because of the last shift size. When increasing the

shift size, it increased the pattern counts.

This table shows the pattern groups’ counts and confidence values for the last five shifts.

37
Pattern Groups Pattern All Last Confidence Sequences
Size Counts 5Shifts Percentage No
0E010004,3E010002,0E01FFE1,0E010009 4 2 2 100% s1
0E010004,3E010002,0E01FFE1,0E010009 ,0E010008 5 2 2 100% s1
0E010005,0E010004,0E010000,0E010001 ,0E010002 5 2 2 100% s1
0E010005,0E010004,0E010007,0E010006 ,3E010002 5 2 2 100% s1
0E010005,0E010004,3E010002,0E01FFE1 ,0E010009 5 2 2 100% s1
3E010000,0E01FFE1,0E01000B,0E01000A,0E010000 5 2 2 100% s2
3E010002,0E01000B,0E01000A,0E010009 4 2 2 100% s3
3E010002,0E01000B,0E01000A,0E010009,0E010008 5 2 2 100% s3
0E010000,0E010001,3E010001 3 2 2 100% s4
0E010001,3E010001 2 2 2 100% s4
0E010002,0E010003,0E010009,0E010008 ,0E010007 5 4 4 100% s5
0E010003,0E010009,0E010008,0E010007 4 4 4 100% s5
0E010003 ,0E010009 ,0E010008,0E010007 ,0E010006 5 4 4 100% s5
0E010008,0E01000B,0E01000A ,0E01FFE0 4 2 2 100% s6
0E010009,0E010008,0E01000B,0E01000A,0E01FFE0 5 2 2 100% s6
0E01000A,0E010009 ,0E010008 ,0E010007 4 2 2 100% s7
0E01000A,0E010009,0E010008,0E01000B ,0E010000 5 2 2 100% s7
0E01000B,0E01000A,0E010009,0E010008 ,0E010007 5 2 2 100% s7
0E01FFE0 ,0E01000A 2 2 2 100% s8
0E01FFE0,0E01FFE1,0E010009,0E010008,3E010002 5 2 2 100% s9
0E01FFE0 ,3E010002 ,0E01FFE1 3 2 2 100% s10
0E01FFE1 ,0E010008 ,0E010009 ,0E010000 4 2 2 100% s11
0E01FFE1 ,0E01000A 2 2 2 100% s12
3E010002 ,0E010002 2 2 2 100% s13
3E010002 ,0E010002 ,0E010003 3 2 2 100% s13
0E010009 ,0E010008 ,0E010007 3 8 7 88% s14
0E010009 ,0E010008 ,0E010007 ,0E010006 4 7 6 86% s14
0E010006 ,0E01000B ,0E01000A ,0E010000 4 5 4 80% s15
0E010006,0E01000B,0E01000A,0E010000,0E010001 5 5 4 80% s15
0E010007,0E010006,0E01000B,0E01000A,0E010000 5 5 4 80% s15
0E010008 ,0E010007 2 9 7 78% s16
0E010004 ,0E010000 ,0E010001 3 4 3 75% s17
0E010006,0E010002,0E010003 ,0E010009 ,0E010008 5 4 3 75% s18
0E010006 ,0E010009 ,0E010008 ,0E010005 4 4 3 75% s18
0E010006,0E010009,0E010008 ,0E010005 ,0E010004 5 4 3 75% s18
0E010007 ,0E010005 2 4 3 75% s19
0E010008 ,0E010007 ,0E010006 3 8 6 75% s19
0E010008 ,0E010007 ,0E010006 ,0E010005 4 4 3 75% s19
0E010008,0E010007,0E010006 ,0E010005 ,0E010004 5 4 3 75% s19

38
0E010009,0E010008,0E010007 ,0E010006 ,0E010005 5 4 3 75% s19
0E01000A ,0E010009 ,0E010008 ,0E01000B 4 4 3 75% s19

Figure 15- Patterns for failure code 1104, comparing last five shifts and confidence percentages

39
After making groups for these patterns, it is created another table (Figure 16) for these

sequence groups which have a hundred percent confidence and related failure times. For

example, groups of s1 occurred before the third and seventh breakdown for 1104.

Faults No 100%
s8,s13
f1
s3,s4,s5,s10,s12,s13
f2
s1,s2,s5,s7,s12
f3
s8
f4
s4,s9,s11
f5
s6,s10,s11
f6
s1,s3,s6,s9
f7

Figure 16- Sequence group distributions according to breakdown orders for failure code 1104 in last five shifts

In this table, it can be seen there are seven faults for reason code 1104. However, there are

13 different sequential pattern groups because of shift sizes. Comparing the last three shift

patterns, the number of patterns increased because this time it was used the last five shifts

in the dataset. The next figure shows sequence group distributions as a graph.

Fault
No
f7
f6
f5
f4
f3
f2
f1
Sequence
s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 s11 s12 s13 No

Figure 17- Graphical representation of sequence group distributions according to breakdown orders for failure code
1104 in last five shifts

40
When calculate the detection rate which equals by dividing counts of filled rows in the

table by faults occurring times. For this example, the detection rate is equal to the count of

filled rows in the table, which is 7 divided by all fault counts which is 7, so it comes to a

100% detection rate. This percentage means that the system can detect 10 out of 10

breakdown situations.

The graph below shows all these patterns according to percentages of confidence:

Comparing Confidence and Counts of Pattern Groups

120%
Confiidence Percentages

100%

80%

60%

40%

20%

0%
13 14 15 19
Counts of Sequence Groups

Figure 18- Graphical illustration of patterns and confidence for failure code 1104 in last five shifts

As it can be seen before for the last 3 shifts, the same thing can be seen for the last 5 shifts

as well. For confidence values 100% to 70%, the count of patterns slightly increased

because when decreased the confidence percentages’ threshold, it increased the number of

patterns.

4.2 Results for Failure code 1140

It will be separated each failure code result into two parts according to the last shift numbers

which are the last three and last five shifts before related machine failures occur.

41
4.2.1 Pattern Results for Last 3 Shifts

The next figure belongs to the last three shift patterns and is between two of the same

failure codes called “all counts”, and it was calculated these pattern groups’ confidence

percentages and ordered them from largest to lowest values for 1140 breakdown codes.

42
Pattern Groups Pattern All Last Confidence Sequences
Size Counts 3Shifts Percentage No

0201FFE0,0B010009 2 2 2 100% s1
0201FFE0,0B010009,0B010000 3 2 2 100% s1
0201FFE0,0B010009,0B010000,0B011519 4 2 2 100% s1
0201FFE1,16016465,0201FFE0,0B010009 4 2 2 100% s1
0201FFE1,16016465,0201FFE0,0B010009,0B010000 5 2 2 100% s1
16016464,0201FFE1,16016465,0201FFE0,0B010009 5 2 2 100% s1
16016465,0201FFE0,0B010009 3 2 2 100% s1
16016465,0201FFE0,0B010009,0B010000 4 2 2 100% s1
16016465,0201FFE0,0B010009,0B010000,0B011519 5 2 2 100% s1
0B010011,160169DD,16016A41 3 2 2 100% s2
0B010011,160169DD,16016A41,160169DC 4 2 2 100% s2
0B010011,160169DD,16016A41,160169DC,16016A40 5 2 2 100% s2
160169DD,16016A41,160169DC 3 3 3 100% s2
160169DD,16016A41,160169DC,16016A40 4 3 3 100% s2
16016A41,160169DC 2 3 3 100% s2
16016A41,160169DC,16016A40 3 3 3 100% s2
2010000,0B01000C 2 2 2 100% s3
16016465,160178B5,160178B4,16016464,2010006 5 2 2 100% s4
0201FFE1,0201FFE0 2 4 3 75% s5
160169DC,16016A40 2 4 3 75% s6

Figure 19- Patterns for failure code 1140, comparing last three shifts and confidence percentages

43
 Fault No
f6
Table f5
f4
Graph
f3
Figure 20- Sequence group f2
distributions according to
breakdown orders for failure f1
code 1140 in last three shifts
s1 s2 s3 s4 Sequence No

44
After making groups for these patterns, it will be created a table for these sequence groups

which have a hundred percent confidence and related failure times. For example, groups

of s1 occurred before the third breakdown for 1140.

In this table, it can be seen there are six faults for reason 1140. Before failures occur, there

are four different sequential pattern groups. S2 occurred before the first failure and s1, s3,

and s4 occurred before the third breakdown, s3 before the fourth breakdown, and s2 again

occurred before the fifth breakdowns in last three shifts. It can be seen these information

as a table and graph in the Figure 20.

When calculate detection rate which equals by dividing counts of filled rows in the table

by faults occurring times. For this example, detection rate equals count of filled rows in

the table which is 4 dividing by total faults’ counts (6), so it comes to a 66% detection rate.

This percentage means that system can detect 6 out of 10 breakdowns situation.

For more accuracy, decreased the confidence value to 70%, which might increase the

detection rate. The next figure shows sequence groups between 70% and 100% and a

distribution of these groups as a table and graph.

45
 Seq
FaultNo Groups Fault No
F1 s6 f6
F2 f5
F3 s5,s6 f4
F4 s5 f3
F5 s6 f2
F6 f1
s1 s2 s3 s4 s5 s6 Sequence No
Table Graph

Figure 21- Sequence group distributions according to breakdown orders for failure code 1140 in last three shifts
above 70%

From this table it can be seen s5 and s6 occurred before the same failures, and it does not

change the detection rate. As it can be seen from Figure 20’s Table; s1, s2, s3 and s4

occurred before the same failure numbers.

The graph below of these patterns is created according to percentages of confidence:

Comparing Confidence and Counts of Pattern Groups

120%
Confidence Percentages

100%
80%
60%
40%
20%
0%
4 6 9 11
Counts of Sequence Groups

Figure 22- Graphical illustration of patterns and confidence for failure code 1140 in last three shifts

For confidence values from 100% to 60%, count of patterns slightly increased because

when decreased confidence percentages’ threshold, it increased the number of patterns.

46
4.2.2 Pattern Results for Last 5 Shifts

The next table belongs to the last five shift patterns and is between two of the same failure

codes called “all counts”, and it was calculated these pattern groups’ confidence

percentages and ordered them from largest to lowest values for 1140 breakdown codes.

This time, obviously, there are more pattern groups because of the last shift size. When

increased the shift size, it increased the pattern counts.

This table shows the pattern groups’ counts and confidence values for the last five shifts.

47
Pattern Groups Pattern All Last Confidence Sequences
Size Counts 5Shifts Percentage No
0201FFE0 ,0B010009 2 2 2 100% s1
0201FFE0 ,0B010009 ,0B010000 3 2 2 100% s1
0201FFE0 ,0B010009 ,0B010000 ,0B011519 4 2 2 100% s1
0201FFE1 ,16016465,0201FFE0 ,0B010009 4 2 2 100% s1
0201FFE1,16016465,0201FFE0,0B010009,0B010000 5 2 2 100% s1
16016464,0201FFE1,16016465,0201FFE0,0B010009 5 2 2 100% s1
16016465,0201FFE0 ,0B010009 3 2 2 100% s1
16016465,0201FFE0 ,0B010009 ,0B010000 4 2 2 100% s1
16016465,0201FFE0,0B010009,0B010000 ,0B011519 5 2 2 100% s1
0B010011 ,160169DD ,16016A41 3 2 2 100% s2
0B010011 ,160169DD ,16016A41 ,160169DC 4 2 2 100% s2
0B010011 ,160169DD ,16016A41 ,160169DC ,16016A40 5 2 2 100% s2
160169DD ,16016A41 ,160169DC 3 3 3 100% s2
160169DD ,16016A41 ,160169DC ,16016A40 4 3 3 100% s2
16016A41 ,160169DC 2 3 3 100% s2
16016A41 ,160169DC ,16016A40 3 3 3 100% s2
2010000,0B01000C 2 2 2 100% s3
16016465,160178B5,160178B4,16016464,2010006 5 2 2 100% s4
0201000A,0B01000C,2010002,2010004,0B010014 5 4 4 100% s5
0201FFE1 ,0201FFE0 2 4 4 100% s6
070108CC ,160178B5 2 2 2 100% s7
070108CC ,160178B5 ,160178B4 3 2 2 100% s7
070108CD ,070108CC ,160178B5 3 2 2 100% s7
070108CD ,070108CC ,160178B5 ,160178B4 4 2 2 100% s7
0B011519 ,0B010000 2 2 2 100% s8
0B010009 ,0B011519 ,0B010000 2 2 2 100% s8
0B01000A ,0B010009 ,0B011519 ,0B010000 4 2 2 100% s8
0B011519 ,0B01000C ,0B01000A 3 2 2 100% s9
0B010009 ,0B011519 ,0B01000C 3 2 2 100% s9
0B010009 ,0B011519 ,0B01000C ,0B01000A 4 2 2 100% s9
0B01000A ,160165F4 2 2 2 100% s10
0B01000A ,160165F4 ,1601607C 3 2 2 100% s10
2010006,16016465,2010000,160178B5 ,2010005 5 2 2 100% s11
160178B5 ,160178B4 ,2010005,2010006 4 2 2 100% s12
3E010000 ,0B010002 2 4 5 80% s13
160169DC ,16016A40 2 3 4 75% s14
160178B4 ,2010005,2010006 3 3 4 75% s15
2010005,0B010009 ,0B0114B5 3 3 4 75% s16
3E010000 ,0B010002 ,0201FFE1 3 3 4 75% s17
Figure 23- Patterns for failure code 1140, comparing last five shifts and confidence percentages

48
After making groups for these patterns, it is created another table for these sequence groups

which have a hundred percent confidence and related failure times. For example, groups

of s1 occurred before the third breakdown for 1140.

Faults No 100%
F1 s2,s5,s7,s10
F2 s7
F3 s1,s3,s4,s6,s9,s11,s12
F4 s3,s6,s8,s9
F5 s2,s5
F6 s5,s10

Figure 24- Sequence group distributions according to breakdown orders for failure code 1140 in last five shifts

In this table, it can be seen there are six faults for reason code 1140. However, there are 12

different sequential pattern groups because of shift sizes. Comparing the last three shifts

patterns, the number of patterns increased because this time it was used the last five shifts

dataset. Next figure shows sequence group distributions as a graph.

Fault
No
f6
f5
f4
f3
f2
f1
Sequence
s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 s11 s12 No

Figure 25- Graphical representation of sequence group distributions according to breakdown orders for failure code
1140 in last five shifts

49
When calculate detection rate which equals dividing counts of filled rows in the table by

faults occurring times. For this example, detection rate equals count of filled rows in the

table which is 6 dividing by all faults’ counts which is 6, so it equals 100% detection rate.

This percentage means that system can detect 10 out of 10 breakdowns situation.

The graph below shows all these patterns according to percentages of confidence:

Comparing Confidence and Counts of Pattern Groups

120%
Confidence Percentages

100%
80%
60%
40%
20%
0%
12 13 17 27
Counts of Sequence Groups

Figure 26- Graphical illustration of patterns and confidence for failure code 1140 in last five shifts

For confidence values 100% to 60%, the count of patterns slightly increased because when

decreased confidence percentages’ threshold, it increased number of patterns.

4.3 Results for Failure code 1143

It will be separated each failure code result into two parts according to the last shift numbers

which are the last three and last five shifts.

4.3.1 Pattern Results for Last 3 Shifts

The next figure belongs to the last three shift patterns and is between two of the same

failure codes called “all counts”, and it was calculated these pattern groups’ confidence

percentages and ordered them from largest to lowest values for 1143 breakdown codes.

50
 Pattern Groups Pattern All Last Confidence Sequences
Size Counts 3Shifts Percentage No
05010ADD ,05010ADC 2 2 2 100% S1
050114B4 ,050114B9 ,050114B8 3 2 2 100% S2
050114B4 ,050114B9 ,050114B8 ,5010905 4 2 2 100% S2
050114B4,050114B9,050114B8,5010905,5010904 5 2 2 100% S2
050114B8 ,5010905 2 2 2 100% S2
050114B8 ,5010905,5010904 3 2 2 100% S2
050114B9 ,050114B8 ,5010905 3 2 2 100% S2
050114B9 ,050114B8 ,5010905,5010904 4 2 2 100% S2
050116CA ,050114B5 2 2 2 100% S3
050119A0 ,050116CB 2 2 2 100% S4
050119A0 ,050116CB ,050116CA 3 2 2 100% S4
050119A1 ,050119A0 ,050116CB 3 2 2 100% S4
050119A1 ,050119A0 ,050116CB ,050116CA 4 2 2 100% S4
3E010002 ,3E010001 ,3E010002 3 2 2 100% S5
050114B4 ,050114B9 2 3 2 67% s6
050117B8 ,3E010002 2 3 2 67% s7
050117B9 ,050117B8 ,3E010002 3 3 2 67% s7
0501199C ,050119CF 2 3 2 67% s8
0501199C ,050119CF ,050119CE 3 3 2 67% s8
050119CE ,3E010002 2 3 2 67% s9

Figure 27- Patterns for failure code 1143 with comparing last three shifts and confidence percentages

After making groups for these patterns, it will be created a table for these sequence groups

which have a hundred percent confidence and related faults time. For example, groups of

s1 occurred before the fifth and ninth breakdowns for 1143.

51
In this table (Figure 28), it can be seen there are 17 faults for reason 1143. Before failures

occur, there are four different sequential pattern groups. S3 occurred before the second

failure, s1 occurred before the fifth, s4 occurred before the sixth, s2 and s3 occurred before

the eighth, s1 again occurred before the ninth, and lastly s5 occurred before the twelfth and

fourteenth breakdown and within the last three shifts. This is provided in the table and

graph in Figure 28.

It was calculated the detection rate by dividing counts of filled rows in the table by the fault

occurring times. For this example, the detection rate is equal to the count of filled rows in

52
the table, which is 7 divided by all fault counts (17), so it comes to about 41% detection

rate. This percentage means that system can detect 4 out of 10 breakdown situations.

For more accuracy, decreased the confidence value to 60%, and it might be also increased

the detection rate. The next figure shows sequence groups between 60% and 100% and the

distribution of these groups as a table and graph are illustrated in the next figure.

Increasing accuracy is crucial because it means catching more failures’ patterns before

machine breakdowns occur. However, for more accuracy it is necessary to have more wide

and clean dataset. Nine months and 11 trucks information would not be enough for making

exact decision, but this thesis would be a good example for predictive maintenance.

Above Fault
FaultNo 60% No
F1 f17
F2 s8,s9 f16
F3 f15
F4 s7,s8 f14
F5 f13
F6 s7 f12
F7 f11
F8 s6 f10
F9 f9
F10 f8
F11 s9 f7
F12 f6
F13 f5
F14 f4
F15 f3
F16 f2
f17 f1
Sequen
s1 s2 s3 s4 s5 s6 s7 s8 s9
No
Table Graph

Figure 28- Sequence group distributions according to breakdown orders for failure code 1143 in last three shifts
above 60% confidence

53
From this table it can be seen s7 and s8 occurred before the fourth breakdown, and s9

occurred before the eleventh breakdown. These fault numbers are different from the table

in Figure 28 and increased the detection rate. The detection rate is equal to the count of

filled rows in the table (9) totally, after decreasing confidence, it increased the filled rows

and divided the fault counts (17), so coming to about 52% detection rate. This percentage

means that the system can detect 5 out of 10 breakdown situations.

The graph below shows these all patterns according to percentages of confidence:

Comparing Confidence and Counts of Pattern Groups

120%
Confiidence Percentages

100%

80%

60%

40%

20%

0%
4 6 9 11
Counts of Sequence Groups

Figure 29- Graphical illustration of patterns and confidence for failure code 1143 in last three shifts

As it can be seen before for last 3 shifts, same thing can be seen for last 5 shifts as well.

For confidence values from 100% to 60%, count of patterns slightly increased because

when decreased confidence percentages’ threshold, it increased the number of patterns.

54
4.3.2 Pattern Results for Last 5 Shifts

The next table belongs to the last five shift patterns and is between two of the same failure

codes called “all counts”, and it was calculated these pattern groups’ confidence

percentages and ordered them from largest to lowest values for 1143 breakdown codes.

This time, obviously, there are more pattern groups because of the last shift size. When

increased the shift size, it increased the pattern counts.

This table shows the pattern groups’ counts and confidence values for the last five shifts.

55
Pattern Groups Pattern All Last Confidence Sequences
Size Counts 5Shifts Percentage No
05010ADD ,05010ADC 2 2 2 100% s1
050114B4 ,050114B9 2 3 3 100% s2
050114B4 ,050114B9 ,050114B8 3 2 2 100% s2
050114B4 ,050114B9 ,050114B8 ,5010905 4 2 2 100% s2
050114B4 ,050114B9 ,050114B8 ,5010905,5010904 5 2 2 100% s2
050114B5 ,050114B4 ,050114B9 3 2 2 100% s2
050114B8 ,5010905 2 2 2 100% s2
050114B8 ,5010905,5010904 3 2 2 100% s2
050114B9 ,050114B8 ,5010905 3 2 2 100% s2
050114B9 ,050114B8 ,5010905,5010904 4 2 2 100% s2
050116CA ,050114B5 2 2 2 100% s3
050119A0 ,050116CB 2 2 2 100% s4
050119A0 ,050116CB ,050116CA 3 2 2 100% s4
050119A1 ,050119A0 ,050116CB 3 2 2 100% s4
050119A1 ,050119A0 ,050116CB ,050116CA 4 2 2 100% s4
050119CA,050116CB ,050116CA ,050119CD ,050119D1 5 2 2 100% s4
0501199B ,0501199A ,3E010002 3 2 2 100% s5
050114B5 ,050114B9 ,050114B7 ,050114B6 4 2 2 100% s6
050114B5 ,050114B9 ,050114B7 ,050114B6 ,050114B4 5 2 2 100% s6
050114B9 ,050114B7 ,050114B6 ,050114B4 4 2 2 100% s6
0501199C ,5010249 2 3 3 100% s7
0501199C ,5010249,5010248 3 3 3 100% s7
0501199C ,5010249,5010248,050114B9 4 2 2 100% s7
0501199C ,5010249,5010248,050114B9 ,050114B8 5 2 2 100% s7
0501199D ,0501199C ,5010249 3 2 2 100% s7
0501199D ,0501199C ,5010249,5010248 4 2 2 100% s7
050119CE ,050119CB 2 2 2 100% s8
050119CF ,050119CE ,050119CB 3 2 2 100% s8
3E010001 ,050119CB ,050119CA ,0501177F 4 2 2 100% s9
3E010001 ,050119CB ,050119CA ,0501177F ,0501177E 5 2 2 100% s9
3E010002 ,3E010001 ,3E010002 3 4 4 100% s10
3E010002 ,3E010001 ,3E010000 ,0501177F 4 2 2 100% s11
3E010002 ,3E010001 ,3E010000 ,0501177F ,0501177E 5 2 2 100% s11
5010248,050114B9 ,050114B8 3 2 2 100% s12
5010249,5010248,050114B9 ,050114B8 4 2 2 100% s12
050114B9 ,050114B7 ,050114B6 3 3 4 75% s13
0501015A ,050119CB 2 3 4 75% s14

Figure 30- Patterns for failure code 1143, comparing last five shifts and confidence percentages

56
After making groups for these patterns, it was created another table for these sequence

groups which have a hundred percent confidence and related failure times. For example,

groups of s1 occurred before the fifth and ninth breakdown for 1143.

Faults 100%
No
F1 s7
F2 s3,s6,s7,s12
F3 s7,s12
F4 s8
F5 s1,s5
F6 s4,s5
F7
F8 s2,s3
F9 s1
F10 s8
F11 s11
F12 s9,s10
F13 s9
F14 s10
F15
F16 s10
F17 s10

Figure 31- Sequence group distributions according to breakdown orders for failure code 1143 in last five shifts

In this table, it can be seen there are 17 faults for reason code 1104. However, there are 12

different sequential pattern groups because of shift sizes. Comparing the last three shifts

patterns, the number of patterns increased, because this time it was used the last five shifts

dataset. Next figure shows sequence group distributions as a graph.

57
 Fault
No
f17
f16
f15
f14
f13
f12
f11
f10
f9
f8
f7
f6
f5
f4
f3
f2
f1
Sequence
s1 s2 s3 s4 s5 s6 s7 s8 s9 s10 s11 s12 No

Figure 32- Graphical representation of sequence group distributions according to breakdown orders for failure code
1143 in last five shifts

When calculate detection rate which equals dividing counts of filled rows in the table by

faults occurring times. For this example, detection rate equals count of filled rows in the

table which is 15 dividing by all faults’ counts which is 17, so it equals about 90% detection

rate. This percentage means that system can detect 9 out of 10 breakdowns situation.

58
The graph below shows all these patterns according to percentages of confidence:

Comparing Confidence and Counts of Pattern Groups

120%

Confidence Percentages
100%
80%
60%
40%
20%
0%
12 13 17 27
Counts of Sequence Groups

Figure 33- Graphical illustration of patterns and confidence for failure code 1143 in last five shifts

For confidence values 100% to 70% count of patterns slightly increased because when

decrease confidence percentages’ threshold, it increases number of patterns.

59
 5 CONCLUSION

Nowadays, with the increasing presence of technological devices, every company tries to

follow their vehicles, especially those that have expensive prices and maintenance costs.

The mining industry has several expensive vehicles, and companies record their statuses

regularly. In this study, there is one big dataset belonging to one North American mining

company. The dataset includes records from eleven trucks for nine months, coming to more

than three million rows, and more than a hundred columns.

It is focused on the sequences of alarms and signals, which might be related to reasons for

breakdowns. However, there are several other features which may have an impact on the

breakdown of machines, such as driver mistakes, machine loading status, speed, or road

conditions. For example, there are several threshold speed limits for mining trucks

depending on road slopes, and the dataset does not show these kinds of thresholds. Another

example is the various loading rules for these mining trucks. When trucks are loaded more

than load limits allow, it may have an impact on the machine’s health status, but there are

not this kind of information and did not calculate for those factors. These features need

more in-depth research and expert knowledge for a future research project.

In this thesis, it was selected three various kinds of breakdown codes which cause more

expensive maintenance costs and mechanical failures. Mechanical breakdown reasons are

more related to alarms and signals than other machine failures’ reasons. It was selected

three of them, but for more accurate results, it must implement more than three breakdown

reasons. According to the three machine failure codes, there are specific relationships

between events and breakdown codes. First, it was discovered several patterns between

60
two same failure codes and counted them. Afterwards, it was discovered various patterns

in the last three and five shifts before breakdowns occur. Lastly, it was calculated

confidence values for the last three and five shifts and illustrated them in tables and graphs.

The results showed that when implement the sequential pattern algorithm before machine

failures, it is possible to discover several patterns which may indicate breakdowns.

However, for more accurate results, it is necessary to have cleaner and larger datasets, and

additionally, time records beyond nine months. Despite these missing values, results

indicate a detection rate of more than 90% in the last five shift events, which shows several

specific and identifiable groups of patterns before machine breakdowns occur. However,

the results do not show high detection rates for the last three shifts’ alarms and signals

before machine breakdowns occur. For future work, a more wide and clean dataset would

be more accurate in discovering mining trucks failure reasons.

61
 REFERENCES

Adolfsson, E., Dahlström, T. (2011). Efficiency in corrective maintenance (Master’s Thesis),

Department of Technology Management and Economics Division of Logistics and Transportation
CHALMERS UNIVERSITY OF TECHNOLOGY Gothenburg, Sweden, Report No. E2011:096
Bastos, P., Lopes, I., & Pires, L. (2013). Application of data mining in a maintenance system for
failure prediction. Safety, Reliability and Risk Analysis, 933-940. doi:10.1201/b15938-138
Bastos, P., Lopes, R., & Pires, L. (2012). A Maintenance Prediction System using Data Mining
Techniques. 2012 Proceedings of the World Congress on Engineering.

Cartella, F., Lemeire, J., Dimiccoli, L., & Sahli, H. (2015). Hidden Semi-Markov Models for
Predictive Maintenance. Mathematical Problems in Engineering, 2015, 1-23.
doi:10.1155/2015/278120
Chen, L., Han, J., Lei, W., Cui, Y., & Guan, Z. (2016). Full-Vector Signal Acquisition and
Information Fusion for the Fault Prediction. International Journal of Rotating Machinery, 2016, 1-
7. doi:10.1155/2016/5980802
Chueh, H. (2010). Mining Target-Oriented Sequential Patterns With Time-Intervals. International
Journal of Computer Science and Information Technology, 2(4), 113-123.
doi:10.5121/ijcsit.2010.2410
Esmaeili, M., & Fazekas, G. (2010). Finding Sequential Patterns from Large Sequence Data.
International Journal of Computer Science Issues, (7).
Fekete J.A. (2015). Big Data in Mining Operations (Master’s Thesis), MSc Business
Administration and Information Systems(IT Management and Business Economics), Copenhagen
Business School, Copenhagen, Denmark.
Kantardzic, M. (2011). Data Mining: Concepts, Models, Methods, and Algorithms. Wiley-IEEE
Press.

Liu, B. (2011). Association Rules and Sequential Patterns. In: Web Data Mining. Data-Centric
Systems and Applications. doi:https://doi.org/10.1007/978-3-642-19460-3_2
Marinelli, M., Lambropoulos, S., & Petroutsatou, K. (2014). Earthmoving trucks condition level
prediction using neural networks. Journal of Quality in Maintenance Engineering, 20(2), 182-192.
doi:10.1108/jqme-09-2012-0031
Murakami, T., Saigo, T., Ohkura, Y., Okawa, Y., & Taninaga, T. (2002). Development of Vehicle
Health Monitoring System (VHMS/WebCARE) for Large-Sized Construction Machine (Vol. 48,
Tech.). Komatsu.

Peng, S., & Vayenas, N. (2014). Maintainability Analysis of Underground Mining Equipment
Using Genetic Algorithms: Case Studies with an LHD Vehicle. Journal of Mining, 2014, 1-10.
doi:10.1155/2014/528414

62
Rouse, M., (2014). What is machine data? - Definition from WhatIs.com. (n.d.). Retrieved March
28, 2018, from http://internetofthingsagenda.techtarget.com/definition/machine-data
Slimani, T., & Lazzez, A. (2014). SEQUENTIAL MINING: PATTERNS AND ALGORITHMS
ANALYSIS. International Journal of Information Technology and Computer Science, 6(3).
doi:10.5815/ijitcs.2014.03.09
Ubaidulla, D., Sushmitha, B. S., & Vanitha, T. (2017). A STUDYONMINING SEQUENTIAL
PATTERN IN TIME SERIES DATA. International Journal of Latest Trends in Engineering and
Technology. Retrieved from https://www.ijltet.org/journal/151065801484.pdf.
Ullah, I., Yang, F., Khan, R., Liu, L., Yang, H., Gao, B., & Sun, K. (2017). Predictive Maintenance
of Power Substation Equipment by Infrared Thermography Using a Machine-Learning Approach.
Energies, 10(12), 1987. doi:10.3390/en10121987
US Census Bureau. (2017, May 03). Library. Retrieved March 28, 2018, from
https://www.census.gov/library/publications/2017/econ/2017-csr.html
Viger, P.F. (2017). An Introduction to Sequential Pattern Mining. (2018, January 02). Retrieved
March 28, 2018, from http://data-mining.philippe-fournier-viger.com/introduction-sequential-
pattern-mining/
Yildirim, M., & Dessureault, S. (2007). Truck assignment performance evaluation by using data
mining techniques. Data Mining Mine Data: Truck-shovel Fleet Management Systems.
doi:10.1201/9781439833407.ch12
Zhao, Q., & Bhowmick, S. S. (2003). Sequential Pattern Mining: A Survey. Technical Report,
CAIS, Nanyang Technological University, Singapore, No. 2003118.

www.highservice.com , FleetSafety® – Highservice. (n.d.). Retrieved April 07, 2018, from

http://www.highservice.com/highservice/en/highservice-technology-eng/fleetsafety-collision-
avoidance-system/

www.uky.edu, Coal Mining, Kentucky Geological Survey. (n.d.). Retrieved April 04, 2018, from
http://www.uky.edu/KGS/coal/coal-mining.php

www.modularmining.com , Modular Mining Systems | Mine Management Solutions. (n.d.).

Retrieved April 02, 2018, from http://www.modularmining.com/wp-content/uploads/Mining-
Magazine__Healthy-and-Wise.pdf
https://www.emaint.com , What is Predictive Maintenance & How to Get Started. (2017, December
19). Retrieved April 02, 2018, from https://www.emaint.com/what-is-predictive-maintenance/

www.generalkinematics.com , A Brief History of Mining: The Advancement of Mining

Techniques and Technology. (2015, December 22). Retrieved March 28, 2018, from
https://www.generalkinematics.com/blog/a-brief-history-of-mining-and-the-advancement-of-
mining-technology/

63
www.revolvy.com , "Preventive maintenance" on Revolvy.com. (n.d.). Retrieved March 28, 2018,
from https://www.revolvy.com/main/index.php?s=Preventive maintenance
www.ni.com , The Cost of Mining Equipment Mismanagement. (n.d.). Retrieved March 28, 2018,
from http://www.ni.com/white-paper/52608/en/

www.cat.com , Mining Trucks Mining Trucks. (n.d.). Retrieved March 28, 2018, from
https://www.cat.com/en_US/products/new/equipment/off-highway-trucks/mining-trucks.html

64
 CURRICULUM VITAE

NAME: Abdulgani Kahraman

ADDRESS: Department of Computer Engineering and Computer Science

University of Louisville
Louisville, KY 40292

EDUCATION: M.S. Computer Science

University of Louisville
2015-2017
B.S. Computer Engineering
Sakarya University
2007-2011

EXPERIENCE: Software Developer

Innova IT Solutions, Istanbul, Turkey
2011-2014

AWARDS: Study Abroad Scholarship

Ministry of Education, Turkey
2014

65