Machine Learning and Optimization for Predictive Maintenance based

on Predicting Failure in the Next Five Days

Eman Ouda

, Maher Maalouf

and Andrei Sleptchenko

Research Center of Digital Supply Chain and Operations, Department of Industrial and System Engineering,

Khalifa University, Abu Dhabi, U.A.E.

Keywords:

Condition-based Maintenance, Predictive Maintenance, Machine Learning, Optimization.

Abstract:

This study proposes a framework to predict machine failures using sensor data and optimize predic-

tive/corrective maintenance schedule. Using historical data, machine learning (ML) models are trained to

predict the failure probabilities for the next ﬁve days. Multiple algorithms, including feature extraction tech-

niques, selections, and ML models (both regression and classiﬁcation based) are compared. The machine

learning models’ output is fed to an optimization model to propose an optimized maintenance policy, and we

demonstrate how prediction models can help increase system reliability at lower costs.

1 INTRODUCTION

With the rise of digitization, Artiﬁcial Intelligence

(AI) implementation in companies can help them be-

come more efﬁcient and competitive. One applica-

tion of AI is through the integration of machine learn-

ing and maintenance. Companies rely on three types

of maintenance for their machines (Carvalho et al.,

2019; Lee and Scott, 2009):

1. Run to failure, also called corrective maintenance,

is when maintenance is conducted only when the

machine fails. This causes a longer downtime pe-

riod and has a high cost because it causes lags in

the processes and the tasks of the machines.

2. Preventive maintenance, also called time-based

maintenance and scheduled maintenance, is when

maintenance follows a schedule periodically. Al-

though failures are prevented, maintenance is

done before the machine fails and it may be un-

necessary.

3. Predictive maintenance uses predictive tools to es-

timate when maintenance is necessary. It mon-

itors the machine health continuously over time

and allows for early detection of failures based on

historic data.

https://orcid.org/0000-0002-8631-5577

https://orcid.org/0000-0003-0516-6870

https://orcid.org/0000-0001-7188-5692

The signiﬁcance of predictive maintenance has in-

creased in the last decades with advances in sensoring

and Internet of Things (IoT) technologies. The main

advantage of the predictive maintenance that it can

minimize the downtime and related costs, can help to

increase the lifespan of the equipment.

There are multiple techniques that can facilitate

predictive maintenance. One technique is relying on

historical data by analyzing it using machine learning

(ML) methods and tools. ML methods nowadays are

faster when compared to ordinary statistical survival

analysis, due to the greater availability of computing

power for sensors and failure data collected in the last

years.

This study applies feature engineering techniques

and machine learning models on sensor data to pre-

dict the failure probabilities up to ﬁve days in ad-

vance. Different machine learning models from the

literature are compared to ﬁnd the highest accuracy.

The obtained failure probabilities are used further in

the proposed optimization model to create the opti-

mum maintenance schedule. In this way, the machine

learning methods, big data, and optimization models

are put together to create an efﬁcient predictive main-

tenance policy. To the best of the authors’ knowl-

edge, a simpliﬁed optimization model that uses the

failure probabilities of multiple machines from ma-

chine learning algorithms has not been formulated.

The remainder of the paper is organized as fol-

lows. Section 2 includes a literature review on ma-

chine learning methods for predictive maintenance

192

Ouda, E., Maalouf, M. and Sleptchenko, A.

Machine Learning and Optimization for Predictive Maintenance based on Predicting Failure in the Next Five Days.

DOI: 10.5220/0010247401920199

In Proceedings of the 10th International Conference on Operations Research and Enterprise Systems (ICORES 2021), pages 192-199

ISBN: 978-989-758-485-5

and mathematical models used. Section 3 introduces

a methodology to deﬁne the libraries and models uti-

lized in this paper. Section 4 presents the results, and

the conclusion demonstrating the ﬁndings is given in

Section 5.

2 MACHINE LEARNING IN

PREDICTIVE MAINTENANCE

Researchers categorize machine learning models into

three types (Cai et al., 2018): supervised learning, un-

supervised learning, and reinforcement learning. Su-

pervised learning includes both regression and classi-

ﬁcation problems. On the other hand, unsupervised

learning uses clustering and association to ﬁnd the in-

herent groupings in the data and detect rules to de-

scribe it. This part of the review covers the different

methods to use machine learning for predictive main-

tenance in literature. The utilization of ML in pre-

dictive maintenance can help in using different strate-

gies. These include regression models to predict the

remaining useful life (RUL) and classiﬁcation models

to predict failure in a given time window.

Regression is used when problems are needed to

predict results based on continuous input and contin-

uous output. Hence, various studies use it to predict

the RUL of machines. One of the studies is a compar-

ative analysis for regression is done for predicting the

RUL (Yurek and Birant, 2019). Two methods are used

to calculate the RUL from The Commercial Modu-

lar Aero-Propulsion System Simulation (C-MAPSS)

dataset before running the regression models. The

ﬁrst approach uses the difference between the current

time of the machine and the fault time is calculated.

In the second approach, the running time is calculated

which is the maximum value of RUL for each fail-

ure. The author used Azure Learning Machine to ap-

ply multiple feature selection methods and different

machine learning algorithms. The study included 72

different models with different feature selection meth-

ods and machine learning models combinations. The

author compared the results and concluded that the

Decision Forest Regression achieves the least predic-

tion error.

Classiﬁcation is used in machine learning to clas-

sify faults and predict their probability. One study

uses the Azure Machine Learning tool to train a Mul-

ticlass Decision Forest model (Paolanti et al., 2018).

The authors achieved an accuracy of 95% on data

from a cutting machine. The proposed method esti-

mated parameters ahead of time. The predicted value

is at time t +dt where t is the time of the readings and

dt is the time ahead. Another research also uses this

method for vibration signals (Amihai et al., 2018).

The authors form a supervised learning problem by

having features and creating labels. Moreover, the

values are of readings are at certain points in time

and the labels are ahead of time. Further, the au-

thors used Python and R for data processing, anal-

ysis, and modeling. Additionally, another research

also uses Python to classify faults in wind turbines

using machine learning models (Hsu et al., 2020).

With 92.7% accuracy for the decision tree model and

91.9% for the random forest model.The authors test it

using K-fold cross-validation to verify the model and

decrease the probability of false alarms. Moreover,

Kusiak and Verma also used condition-based mon-

itoring tools to identify potential faults (Kusiak and

Verma, 2012). Additionally, another study performs a

comparative analysis of the classiﬁcation of four dif-

ferent classes. It uses ﬁve different machine learning

algorithms. (Neural Network, Support Vector Ma-

chine, Random Forest, Boosting Tree, and General

Chi-square Automatic Interaction Detector). The data

used is from large wind farms with 17 wind turbines.

It proved that Random forest algorithm model pro-

duces the best accuracy of 98%. Another study also

generates models using the Random forest approach

with number of trees = 40 and maximum depth = 25

(Canizo et al., 2017). The values were chosen based

on the trial and error of several predictive models and

methods to ﬁnd the best values. The study uses Big

Data and processes it to generate predictive models.

The Pearson correlation for feature selection is used

and the accuracy produced reached 82.04%. Using

Python to process the data and build a model is fea-

sible. When it comes to predicting faults and clas-

sifying them, the decision tree models produced the

highest accuracy in the studies reviewed.

Although the majority of the aforementioned pa-

pers concluded that the decision tree algorithm pro-

vides high accuracy, other feature engineering meth-

ods can be used to boost accuracy. One is a di-

mension reduction method done for classiﬁcation

(Aremu et al., 2020). The authors used the Machine

Learning Dimension Reduction framework (MLDR-

framework) and produced higher accuracy. This ex-

plains how researchers can use feature selection meth-

ods to increase the accuracy of results. Another

study uses AutoRegressive Integrated Moving Aver-

age (ARIMA) to predict failures and classify faults

(Kanawaday and Sane, 2018). The ARIMA model

was used to predict future data points. They use data

collected from a slitting machine. The four mod-

els used are Na

ıve Bayes, Support Vector Machine,

CART, and Deep Neural Network. Accuracies were

96%, 95%, 94%, and 98% respectively. Moreover,

Machine Learning and Optimization for Predictive Maintenance based on Predicting Failure in the Next Five Days

193

other studies have implemented unsupervised learn-

ing algorithms. The authors in (Amruthnath and

Gupta, 2018) use unsupervised learning on exhaust

fan vibration data for predictive maintenance and pro-

pose a method using R programming. They conclude

that clustering algorithms are best for fault detection

under different levels. Overall, when studies use the

right techniques, unsupervised models may perform

better than decision trees. However, due to its sim-

plicity, only supervised models are studied in this re-

search.

Condition-based monitoring that depends on op-

timization models is widely available in the litera-

ture. Literature reviews usually focus on single com-

ponents (Alaswad and Xiang, 2017; Sakib and Wuest,

2018). Due to the limitations of complexity, the study

of multiple component systems in predictive main-

tenance in literature is not widely covered. The de-

pendence of the components and their level makes it

harder to construct an effective model. For illustra-

tion, a single component mathematical optimization

model is studied for condition-based maintenance

(Tian et al., 2012). For single-component models, as

the reliability increases, the cost increases. Reliability

is the probability of preventive maintenance. Hence,

the objective function is to minimize the cost for opti-

mum maintenance. Moreover, another research stud-

ies multiple components into consideration (Einabadi

et al., 2019). The indices were the number of parts,

periods, and machines. Different costs were consid-

ered such as the cost of repair, renewing, and purchas-

ing. The decision variables include timings of when

a component must be replaced and the time of main-

tenance. These additional features helped in this re-

search to put together an optimal maintenance sched-

ule.

Another type of research refers to a threshold

to compare the probability of failure (Nichenametla

et al., 2017). Once a component exceeds the thresh-

old, maintenance is required. The higher the probabil-

ity of failure of the machine, the higher the priority of

inspection. The authors establish threshold targets by

probability plots and reliability comparisons. Further-

more, another study presents a unique model called

Preventive Maintenance Scheduling Problem with In-

terval Costs (PMSPIC) with an objective to minimize

maintenance costs. (Bangalore and Patriksson, 2018).

It takes into consideration both age-based and cost-

based failure rates. A different cost model takes into

consideration repair costs, downtime cost, and set-up

costs performed an opportunistic PdM strategy (Hu

et al., 2012). However, the maintenance strategy is

to put the machines in groups. The strategy suggests

that maintenance may be executed in groups and it de-

creases costs as low as possible. Additionally, another

study also groups the machines (Vu et al., 2020). The

authors in the research use Genetic Algorithm with

Memory (GAM) to group the machines into groups in

certain time slots to decrease the costs. Figures 1 and

2 show the original maintenance plan and the modi-

ﬁed one. Since maintenance is implemented groups

in the same time period, it produces a lower cost. In

conclusion, the method of grouping machines to per-

form maintenance is an efﬁcient method to decrease

the cost as much as possible.

Figure 1: Original schedule before grouping components.

Figure 2: Schedule after grouping components.

3 PROPOSED METHODOLOGY

The methodology section focuses on two parts. First,

the machine learning phase which consists of the fol-

lowing steps:

1. Historical data selection: Selection of sufﬁcient

historical data to process.

2. Data preprocessing: The data is modiﬁed through

feature engineering to be suitable to be fed into

the model.

3. Model selection: A suitable model is chosen to

based on the given dataset and the application.

4. Model training and model validation: The data

is fed into the machine learning model and vali-

dated.

ICORES 2021 - 10th International Conference on Operations Research and Enterprise Systems

194

The second phase consists of the mathematical model

to minimize the cost of the overall maintenance of the

machines. The full model diagram is shown in Figure

3. Sensor data is the input to the model and the ﬁnal

output is a decision of the number of machines to be

maintained.

Figure 3: Proposed model diagram.

3.1 Proposed Machine Learning Model

3.1.1 Data Preprocessing - Format and Label

The data used has to be relevant, sufﬁcient, and of

high quality. Therefore, understanding the data is cru-

cial. During the data preprocessing step, feature engi-

neering was used to frame the problem appropriately.

Feature extraction is a part of feature engineering and

is used to generate features from the existing data by

changing and transforming it. Moreover, feature se-

lection is useful to eliminate unnecessary data from

the dataset (Cai et al., 2018). It reduces the dimen-

sions which can give us better models for machine

learning. In supervised learning, this is usually done

manually.

The dataset in this study is obtained from the

datasets used by Fidan Boylu Uz in “Predictive Main-

tenance Modelling Guide” (Uz, 2016). It contains

sensor readings from 100 machines. Originally, ﬁve

different datasets are given and a sample of each one

is shown in Table 1. Moreover, these are merged into

one big dataset and the columns are considered the

features. However, feature engineering plays a role

here and other features are extracted from the original

dataset. The important additional features are shown

in Table 2. The ﬁnal features are then used to feed

the machine learning model and predict failures in ad-

vance.

From the features, it is important to have sufﬁ-

cient data to answer the question the machine learn-

ing model output is expected to answer. The data is

measured in time stamps. Hence, some of the times-

tamps may not have recorded data. Therefore, han-

dling the missing data is important. The average of

each over 24 hours is taken, this feature engineering

step eliminates the problem of having missing data.

Moreover, decreases the data points from 876,100 to

36,600 readings eases data handling. The last feature

in Table 2 represents the labels of the dataset. In ev-

ery model, a different column is targeted as the label

feature. Parameter dt ranges from 1 to 5 and each

number represents the number of look-ahead days to

be predicted.

3.1.2 Machine Learning Algorithms

In this study, three machine learning algorithms are

used to predict the failures, namely, Logistic Regres-

sion, Random Forest and Gradient Boosting Classi-

ﬁer. The Scikit-learn library was used for predictive

analysis and machine learning models. The machine

learning model is used to predict the failure days in

advance. Depending on dt, the number of days in ad-

vance, the label is predicted. Different machine learn-

ing models are trained every time to predict the differ-

ent days. The label is either a 0 or 1; 1 is a prediction

that failure will happen after dt days. The probabil-

ity of every class can also be extracted. Hence, we

achieve the probability that a failure will occur.

The data is then split into testing and training. The

dataset consists of data for one whole year. The ﬁrst

three quarters of the year are used to train the model.

The last quarter was used for testing. The lowest test-

ing accuracy of the two classes is recorded.

Machine Learning and Optimization for Predictive Maintenance based on Predicting Failure in the Next Five Days

195

Table 1: The given datasets.

Dataset Given Columns/Features

Telemetry datetime machineID volt

1/1/2015

6:00

1 176.21

vibration pressure rotate

45.087 113.078 418.50

Errors datetime machineID error ID

1/3/2015

7:00

1 error1

Machines machineID model age

1 model3 18

Maint datetime machineID comp

6/1/2014

6:00

1 comp2

Failures datetime machineID failure

1/5/2015

6:00

1 comp

3.2 Optimization Model

The proposed mathematical model reduces mainte-

nance costs. Costs are split into different parts. In-

cluding costs of labor, repairing, renewing, or replac-

ing. The costs depend on what state the machine is

at. At each state s = 0, . . . , S, the probability of failure

decreases until it reaches state 0 which is the failure

mode. At failure mode, it is assumed that the machine

would need replacement and the cost would be high-

est.

Indices:

t – Days Index (t = 1, ..., T ).

s – State Index (s = 0, ..., S).

Parameters:

M – Minimum number of working machines in

day t.

– Cost of ﬁxed maintenance for one machine

from state s.

– Probability of moving to state s to s

Decision variables:

– Number of machines maintained at state s in

day t.

– Number of machines in state s at day t.

Table 2: Additional features added to the dataset.

Feature Explanation

Volt 24hr mean

The Average of

the voltage over

the past 24 hours.

Volt 24hr std

The standard deviation

of the voltage over the

past 24 hours.

Volt Moving Av

The average voltage

within a time window.

The time window is two

days used for this study.

Volt Moving Std

The average voltage

within a time window.

The time window is two

days used for this study.

Volt Expanding Av

The average of all voltage

readings since

last maintained.

Volt Expanding Std

The standard deviation of

all voltage readings

since last maintained.

Total Errors

The total number of errors

in the past 24 hours.

Last failure

Number of days since

the last failure occurred.

Last maintenance

Number of days since the

last maintenance occurred.

Failure (T + dt)

Binary

1= Failure

0= No failure

dt is time that needs to

predicted ahead of time.

Optimization Model formulation:

min

∑

t=1

∑

s=1

(1)

s.t. Y

= (Y

st−1

− X

st−1

)

1 −

s−1

∑

=s+1

t−1

− X

t−1

, ∀s < S (2)

= Y

St−1

1 −

S−1

∑

S−1

∑

t−1

(3)

∑

s=1

≥ M (4)

< Y

(5)

≥ 0 (6)

In this formulation, objective (1) is set to mini-

ICORES 2021 - 10th International Conference on Operations Research and Enterprise Systems

196

mize the cost of maintenance, where C

is the cost of

repairing machines in state s to the fully operational

state.

Constraints (5) ensure that the number of ma-

chines repaired in a certain state must be less than

the total number of machines in that state. Constraints

(6) ensure that the number of machines remaining in a

certain state (Y

) or being repaired from a certain state

) must be non-negative. Constraints (2) ensure

that the total number of machines at each state at a

certain point in time considers the previously repaired

machines and the probabilities of failure. Constraints

(3) are similar to constraints (2). However, they only

apply to the ﬁnal state and include all the previously

repaired machines. Constraints (4) ensure that the

number of machines working at a certain point in time

is greater or equal to M. M is the minimum number

of operating machines required.

The number of machines that should be main-

tained from different states on the next t number of

days is decided. This is the predictive maintenance

decision. We can consider the costs to be corrective

maintenance if we wait till the machine reaches fail-

ure mode and then repair or replace it. Hence, com-

paring both approaches is the methodology adopted

in this research.

4 CASE STUDY

4.1 Input Data

In this section of the results, the input data to the ma-

chine learning model and the optimization model are

deﬁned. First, the data used to train and test the ma-

chine learning model is the same dataset analyzed in

Section 3. Then, the obtained predicted probabilities

are extracted and used in the optimization model.

In this study case, there are initially 100 machines

at the fully operational state and are then classiﬁed

into three states:

1. Failure state

2. Suspicious Failure state

3. Fully Operational state

At each state, the average of all the machines’ prob-

abilities is calculated. The Suspicious Failure state

probability to the Failure state is extracted from ma-

chines with a probability of failure greater than 80%.

The Fully Operational state to the Suspicious Fail-

ure state probability is resulted from machines with

a probability of failure between 20% and 80%. The

Fully Operational state to the Failure state probability

is assumed to be 10%, because of the uncertainty in

the model and that there is always a chance of sudden

failure.

Furthermore, the distribution of the maintenance

costs in this study case shown in Figure 4 is based on

justiﬁable assumptions. The costs of repair at Fully

Operational state and Suspicious Failure state are as-

sumed to be 1 unit and 3 units respectively. As men-

tioned in Section 3, we expect the Failure state to cost

more than other states an this why it assumed to cost

6 units per machine. The cost of repair at a fully oper-

ational state and is considered low because it includes

inspection costs.

Additionally, we assume that M, the minimum

number of machines operating, is 80. The mainte-

nance decision of the next f ive days is made.

Figure 4: (a) Predictive maintenance cost distribution.

(b) Corrective maintenance cost distribution.

4.2 Results

In this section, the results of both the machine learn-

ing model and the optimization model are illustrated.

Three machine learning models, Logistic Regres-

sion, Random Forest Classiﬁer, and Gradient Boost-

ing Classiﬁer, are used to calculate the probability of

failure. Since the output is classiﬁed to be either 0

or 1, the class with the lowest accuracy is recorded.

The accuracy also corresponds to the recall score and

the lowest recall score is noted. Figure 5 shows the

results of the three models in predicting failure in the

next ﬁve days. Gradient Boosting classiﬁer appears

to produce the best result in all ﬁve days. The proba-

bility of failure of each machine is predicted from the

gradient boosting classiﬁer. Using the assumptions

mentioned in Section 4.1, the machines are split into

three different states based on the probability values.

At each state, the average of the probabilities is cal-

culated and the results are listed below.

• Fully Operational state to the Suspicious Failure

= 0.3

• Suspicious Failure state probability to the Failure

Machine Learning and Optimization for Predictive Maintenance based on Predicting Failure in the Next Five Days

197

state = 0.9

The probabilities of failure are then fed into the opti-

mization model presented in 3.2 along with the inputs

discussed in Section 4.1. A linear program solver is

used and the model maintenance policy suggests that

within the next 5 days, the total number of machines

inspected or repaired at each state is as follows:

1. Failure state – 17 machines.

2. Suspicious Failure state – 87 machines.

3. Fully Operational state – 0 machines.

Given this predictive maintenance approach, the total

cost of keeping 80 machines operating is 363 units.

However, when using the corrective maintenance ap-

proach, the machines are repaired only at Failure state

and inspected at the Fully Operational state. Hence,

the Suspicious Failure state is not accounted for and

all machines in this state are assumed to reach the

Failure state. This further justiﬁes Figure 4. Accord-

ing to the predicted probabilities of failure, in total,

104 machines should be maintained. Further, as men-

tioned in Section 4.1, the corrective costs are 6 units.

The addition of all costs outputs the costs of correc-

tive maintenance; 104 × 6 = 624 units. In compar-

ison with the 363 units calculated by the proposed

maintenance policy, this is close to a 50% increase in

cost. The prediction model calculated a lower cost be-

cause 87 out of 104 machines could be repaired early

and avoid high corrective maintenance costs. Hence,

the predictive maintenance model achieves the same

goals at lower costs and is solved within a few sec-

onds, on a consumer’s laptop. The proposed model

can be applied to more than three states and achieve

efﬁcient maintenance policies.

Figure 5: Accuracy comparison between different machine

learning models.

5 CONCLUSION

In conclusion, this study develops a framework to re-

duce maintenance costs using both a machine learn-

ing model and an optimization model. Three differ-

ent machine learning models were compared on the

given dataset and the gradient boosting classiﬁer pro-

duced the highest prediction accuracy reaching 99%

for one day in advance prediction. Moreover, the sug-

gested maintenance policy was applied and compared

to a conventional corrective policy. It displayed how

having a predictive maintenance policy can increase

system reliability and decrease costs.

This study faces two main limitations. First, the

dataset is imbalanced. The number of failures is the

minority in the dataset. Hence, feature engineering

techniques must be used to balance out the dataset

before feeding it into the machine learning models.

Another limitation in this study was the costs of main-

tenance. They were given assumed values due to not

having sufﬁcient and exact data about the cost of re-

pair at each state.

For the prediction model, future works may in-

clude other feature engineering techniques to produce

higher accuracy. Big data may also be used along with

other regression models to predict the remaining use-

ful life of the machine. Moreover, deep learning can

be implemented on the dataset used in this research to

avoid feature selection. For the mathematical model,

future works may include maintenance decisions per

machine rather than taking an average and expand the

schedule to suggest maintenance decisions for more

days in advance.

ACKNOWLEDGEMENTS

This publication is based upon work supported by the

Khalifa University of Science and Technology under

Award No. RC2 DSO.

REFERENCES

Alaswad, S. and Xiang, Y. (2017). A review on condition-

based maintenance optimization models for stochas-

tically deteriorating system. Reliability Engineering

and System Safety.

Amihai, I., Gitzel, R., Kotriwala, A. M., Pareschi, D., Sub-

biah, S., and Sosale, G. (2018). An industrial case

study using vibration data and machine learning to

predict asset health. Proceeding - 2018 20th IEEE In-

ternational Conference on Business Informatics, CBI

2018, 1:178–185.

ICORES 2021 - 10th International Conference on Operations Research and Enterprise Systems

198

Amruthnath, N. and Gupta, T. (2018). A research study on

unsupervised machine learning algorithms for early

fault detection in predictive maintenance. In 2018

5th International Conference on Industrial Engineer-

ing and Applications, ICIEA 2018, pages 355–361. In-

stitute of Electrical and Electronics Engineers Inc.

Aremu, O. O., Hyland-Wood, D., and McAree, P. R. (2020).

A machine learning approach to circumventing the

curse of dimensionality in discontinuous time series

machine data. Reliability Engineering and System

Safety, 195(October 2019):106706.

Bangalore, P. and Patriksson, M. (2018). Analysis of

SCADA data for early fault detection, with applica-

tion to the maintenance management of wind turbines.

Renewable Energy.

Cai, J., Luo, J., Wang, S., and Yang, S. (2018). Feature se-

lection in machine learning: A new perspective. Neu-

rocomputing.

Canizo, M., Onieva, E., Conde, A., Charramendieta, S.,

and Trujillo, S. (2017). Real-time predictive main-

tenance for wind turbines using Big Data frameworks.

2017 IEEE International Conference on Prognostics

and Health Management, ICPHM 2017, pages 70–77.

Carvalho, T. P., Soares, F. A. A. M. N., Vita, R., and Fran-

cisco, P. (2019). A systematic literature review of ma-

chine learning methods applied to predictive mainte-

nance. Computers & Industrial Engineering, 137(Au-

gust):106024.

Einabadi, B., Baboli, A., and Ebrahimi, M. (2019). Dy-

namic Predictive Maintenance in industry 4.0 based

on real time information: Case study in automotive

industries. IFAC-PapersOnLine, 52(13):1069–1074.

Hsu, J. Y., Wang, Y. F., Lin, K. C., Chen, M. Y., and Hsu, J.

H. Y. (2020). Wind Turbine Fault Diagnosis and Pre-

dictive Maintenance Through Statistical Process Con-

trol and Machine Learning. IEEE Access, 8:23427–

23439.

Hu, J., Zhang, L., and Liang, W. (2012). Opportunistic

predictive maintenance for complex multi-component

systems based on DBN-HAZOP model. Process

Safety and Environmental Protection.

Kanawaday, A. and Sane, A. (2018). Machine learning for

predictive maintenance of industrial machines using

IoT sensor data. Proceedings of the IEEE Interna-

tional Conference on Software Engineering and Ser-

vice Sciences, ICSESS, 2017-Novem(Figure 2):87–90.

Kusiak, A. and Verma, A. (2012). A Data-Mining Approach

to Monitoring. IEEE Transactions on Sustainable En-

ergy, 3(1):150–157.

Lee, H. H. Y. and Scott, D. (2009). Overview of mainte-

nance strategy, acceptable maintenance standard and

resources from a building maintenance operation per-

spective. Journal of Building Appraisal.

Nichenametla, A. N., Nandipati, S., and Waghmare, A. L.

(2017). Optimizing life cycle cost of wind turbine

blades using predictive analytics in effective mainte-

nance planning. In Proceedings - Annual Reliability

and Maintainability Symposium.

Paolanti, M., Romeo, L., Felicetti, A., Mancini, A., Fron-

toni, E., and Loncarski, J. (2018). Machine Learn-

ing approach for Predictive Maintenance in Industry

4.0. 2018 14th IEEE/ASME International Conference

on Mechatronic and Embedded Systems and Applica-

tions, MESA 2018, pages 1–6.

Sakib, N. and Wuest, T. (2018). Challenges and Opportu-

nities of Condition-based Predictive Maintenance: A

Review. In Procedia CIRP.

Tian, Z., Lin, D., and Wu, B. (2012). Condition based main-

tenance optimization considering multiple objectives.

Journal of Intelligent Manufacturing.

Uz, F. B. (2016). Predictive Maintenance Modelling Guide

Data Sets.

Vu, H. C., Do, P., Fouladirad, M., and Grall, A. (2020). Dy-

namic opportunistic maintenance planning for multi-

component redundant systems with various types of

opportunities. Reliability Engineering and System

Safety.

Yurek, O. E. and Birant, D. (2019). Remaining Useful Life

Estimation for Predictive Maintenance Using Feature

Engineering. Proceedings - 2019 Innovations in In-

telligent Systems and Applications Conference, ASYU

2019, pages 1–5.

Machine Learning and Optimization for Predictive Maintenance based on Predicting Failure in the Next Five Days

199