See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327334242

Machine Learning approach for Predictive Maintenance in Industry 4.0

Conference Paper · July 2018

DOI: 10.1109/MESA.2018.8449150

CITATIONS READS
59 8,203

6 authors, including:

Marina Paolanti Luca Romeo

University of Macerata Università Politecnica delle Marche
83 PUBLICATIONS   877 CITATIONS    66 PUBLICATIONS   662 CITATIONS   

SEE PROFILE SEE PROFILE

Adriano Mancini Emanuele Frontoni

Università Politecnica delle Marche Università Politecnica delle Marche
160 PUBLICATIONS   2,066 CITATIONS    296 PUBLICATIONS   3,190 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Artificial Intelligence View project

IEEE Italy Section Consumer Electronics Society Chapter View project

All content following this page was uploaded by Luca Romeo on 08 January 2020.

The user has requested enhancement of the downloaded file.

 2018 14th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA)
Draft

Machine Learning approach for Predictive

Maintenance in Industry 4.0
Marina Paolanti, Luca Romeo Jelena Loncarski
Andrea Felicetti, Adriano Mancini Department of Engineering Sciences
Emanuele Frontoni Division for Electricity Research
Department of Information Engineering Uppsala University
Università Politecnica delle Marche Box 534, S-751 21 Uppsala, Sweden
Ancona, Italy Email: jelena.loncarski83@gmail.com
Email: {m.paolanti, l.romeo}@univpm.it
a.felicetti@pm.univpm.it
{a.mancini, e.frontoni}@univpm.it

Abstract—Condition monitoring together with predictive main- been much interest on ways to better diagnose the wellness
tenance of electric motors and other equipment used by the condition of these motors [4]. The bearing failure is identified
industry avoids severe economic losses resulting from unexpected as the most frequent cause of motor failure and most common
motor failures and greatly improves the system reliability. This
paper describes a Machine Learning architecture for Predictive maintenance problem. Accordingly, predictive maintenance
Maintenance, based on Random Forest approach. The system mainly focuses on two aspects: energy efficiency improvement
was tested on a real industry example, by developing the data (key of energy saving) and unscheduled downtime reduction.
collection and data system analysis, applying the Machine Learn- The algorithms developed around these two can also be
ing approach and comparing it to the simulation tool analysis. generally divided into two categories:
Data has been collected by various sensors, machine PLCs and
communication protocols and made available to Data Analysis 1) energy and efficiency, there have been many evaluation
Tool on the Azure Cloud architecture. Preliminary results show a methods and devices developed, as in [5],
proper behavior of the approach on predicting different machine 2) system condition monitoring, including the detection of
states with high accuracy.
motor faults, with various fault-detection techniques de-
I. I NTRODUCTION veloped, as for instance in [6].
Some initial concepts of intelligent predictive decision support
Predictive maintenance (PdM) or sometimes called “on-line
systems have been introduced in [7]. Algorithms are essential
monitoring”, “risk-based maintenance”, or “condition-based
for effective predictive maintenance. There are various kinds
maintenance”, is a subject to many recent research papers with
of techniques to be applied in various phases of PdM imple-
the long history behind it. It refers to the intelligent monitoring
mentation, i.e. data processing, diagnostics, and prognostics,
of equipment to avoid future failures. Predictive maintenance
as given in [8]. In PdM three kinds of approach can be
has evolved from the first method that is visual inspection
distinguished:
to automated methods using the advanced signal processing
techniques based on pattern recognition and machine learning, 1) Data-driven approach;
neural networks, fuzzy logic, etc. The automated methods pro- 2) Model-based approach;
vide viable solution to many industries detecting and collecting 3) Hybrid approach.
sensitive information from the equipment which are mainly The data-driven approach is also known as the data mining
motors, where human eyes or ears can cease to do so [1]. approach or the machine learning approach, which uses his-
Together with integrated sensors, predictive maintenance can torical data to learn a model of system behavior. Model-based
avoid unnecessary equipment replacement, reduce machine approach has the ability to incorporate physical understanding
downtime, find the root cause the fault, and in this way of the target product, relying on the analytical model to repre-
save costs and improve efficiency. Predictive maintenance sent the behavior of the system. Machine learning approaches
overlaps with the scope of preventive maintenance in terms are viably used in the areas where the availability of data
of scheduling the maintenance activity in advance to avoid is increasing [9], [10], [11], [12] such as the maintenance in
machine failures. In contrast to conventional preventive main- industry sector. It is increasingly providing effective solutions,
tenance, predictive maintenance schedules activities are based cloud-based solutions, and newly introduced algorithms [13].
on collected data from sensors and analysis algorithms [2], [3]. Machine Learning-based PdM can be divided into the fol-
In process industries, induction motors make up approximately lowing two main classes: Supervised, where information on
70% of all driven electrical loads. In this regard, there has the occurrence of failures is present in the modelling dataset

978-1-5386-4643-4/18/$31.00 ©2018 IEEE

 2018 14th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA)
Draft

and Unsupervised - where logistic and/or process information II. T HE DATA A NALYSIS
is available, but no maintenance data exists [13], [14]. The Predictive Maintenance is an important maintenance tool
availability of maintenance information mostly depends on that is based on the possibility of estimating the future values
the nature of the existing maintenance management policy. of some quantities that characterize a system (typically a
When possible, supervised solutions are preferable. From the machine, a plant, or a production process) through particular
Machine Learning perspective, depending on the output of mathematical models in order to identify in advance the
the data set, two classes of supervised problem are possible: anomalies and potential faults [16], [17]. The basic scheme
regression problem (if output assumes continuous values), of PdM is as follows:
and classification problem (if output assumes categorical val-
• Measurement of physical quantities in real time.
ues) [15].
• Estimation of measurable (or non-measurable) parameters
In this paper a new PdM methodology based on PdM
at time t + dt.
machine learning approach on a cutting machine is presented.
• Identification of the system status considered anomalous
PdM is a strategy viable adopted when dealing with mainte-
or faulty.
nance issues given the increasing need to minimize downtime
• Planning of preventive and corrective activities before the
and associated costs. The methodology has been implemented
system reaches the critical condition.
in the experimental environment on the example of a real
industrial group, producing accurate estimations. An example Examples of predictive maintenance are the following:
of the machine adopted in our study is depicted in figure 1. • Machine vibrations can signal bearing deterioration or
deformation of particular mechanical parts.
• The temperature of a motor and its drown current, may
indicate that friction and possible mechanical malfunction
are degrading the functionality.
• The measurement of particles in a lubricant indicates the
degradation of rubbing contact parts. With appropriate
sensors it is possible to measure the composition of the
lubricating oil and check the health of the machine.
The first phase of these processes is based on the estimation
of the parameters. On the basis of PdM, the technology is
capable of making reliable forecasts. If the prediction algo-
rithms produce incorrect estimates or with too large reliability
intervals, it will be difficult to identify the anomalies and
make the decisions on maintenance and correction. Broadly
speaking, the forecasts can be of two main types:
Figure 1. Example of the machine adopted in this study.
• Cross-Sectional Forecasting;
Data has been collected by various sensors, machine PLCs • Time Series Forecasting.
and communication protocols and made available to Data
Analysis Tool. The proposed PdM methodology allows dy- A. Cross-Sectional Forecasting
namical decision rules to be adopted for maintenance man- Cross-Sectional Forecasting is the estimation of parameters
agement, which is achieved by training a Random Forest of which there exist no measurements, by using measures on
approach on Azure Machine Learning Studio. Preliminary other variables that have been observed. As an example, it can
results show a proper behavior of the approach on predicting be possible to predict the life of an electronic component used
different machine states with high accuracy (95%) on a data in particular conditions, by measuring the electric current that
set of 530731 data readings on 15 different machine features passes through it [18].
collected in real time from the tested cutting machine.
B. Time Series Forecasting
Paper contributions are mainly on the overall cloud archi-
tecture for Industry 4.0, on the application of ML approaches Time Series Forecasting is referring to the estimation of
to a real data set from machines on the field, on the high level parameters that change over time, being measured until time
of accuracy on predicting the state of the main spindle rotor. instant t and the value to be predicted is at the time instant t
The paper is organized as follows: Section II gives an + dt. Typically, the measurements of the variable of interest
overview on the data analysis, including the approach used can be obtained at regular intervals, and then it is possible
in the experimental verification to analyze the residuals. Sec- to predict the future values. The simplest example is the
tion III gives the Machine learning approaches for PdM, ex- prediction of the minutes of residual charge of our mobile
plaining in deep the regression and the classification methods. phones, which is estimated based on the consumption that we
Section IV gives the experimental results on the real use have made, and the way we use the phone [19]. In time series,
case where Machine Learning approach for PdM has been it is usual to identify:
implemented and tested. • Trends, or a long-term increase (or decrease) in values.
 2018 14th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA)

• The seasonal phenomena, or the phenomena that deter-

mine changes in values over a period of time that is
always repeating the same duration.
• Cyclical phenomena that cause increases and decreases
in values with fluctuations that do not always have the
same duration, i.e. they are not periodic.
One of the most important things to understand when
analyzing data is the type of relationship between the measured
quantities. To do this, graphic visualization with scattered plots
can be viably used to identify how the data are dependent.
The hypothesis of linear regression is that the estimating
phenomenon has a linear behavior. A simple way to check
if the signal we want to predict is linear, or contains other
information that is not contained in our forecast model, is to
analyze the residuals [20].
C. The analysis of the residuals
The residuals are the difference between the measured
values of the size and the fitting values obtained with the Figure 2. Scheme of the activities to be undertaken in PdM.
prediction line

ei = yi − y0i (1) for the algorithm to learn the normal operating scheme,
where ei is the residual of the i-th measurement, yi is in addition to the failure scheme through the training
the i-th measured value and y0i = m0ti + q0 is the estimated process. Consequently, it is essential that the training
value at time ti . Coefficients m and q are calculated with the data contain a sufficient number of examples in both
method of least squares [21]. A practical method to verify categories.
whether the calculated regression model satisfies the signal 2) Maintenance/repair history: An essential source of data
to be predicted is the autocorrelation. The autocorrelation of for predictive maintenance solutions is the detailed asset
the residues above calculated can be calculated, counting how of maintenance history, which contains information about
many of the residuals have a value included in the interval: replaced components, preventive maintenance tasks per-
formed, and so on.
2 3) Machine conditions: to estimate how many days (or
Int = ± √ (2)
N hours, kilometers, etc.) a machine lasts before a failure
occurs, it is assumed that its health status decreases
where N is the number of the measurements. Usually it
with time. It is therefore necessary that the data contain
is considered that if >95% of the residuals are within the
time-varying functions that acquire aging patterns or any
± Int, they are not correlated among them and the noise can be
anomaly that could cause performance reduction.
considered as white noise. This approach allows to construct
algorithms that use linear regression for signal prediction, In order to prepare the data in the format required to
useful for predictive maintenance, and automatically verifies create the functions to be included in the Machine Learning
in real time any variations of the system under observation. algorithm, it is necessary to perform some pre-processing
It is also useful to use the test of residuals to understand if steps. The first is to divide the duration of data collection
linear regression is the best choice for the predictive model into units of time in which each record is a unit of time for
we want to build. an asset. The generic data schema could be:
• Maintenance records: these are the records of the main-
III. M ACHINE L EARNING A PPROACHES FOR P D M
tenance actions performed. The unprocessed maintenance
Figure 2 gives a scheme, showing the actions and technical data is usually associated with an asset ID and a times-
activities to be undertaken to implement PdM. Machine Learn- tamp with information on maintenance activities per-
ing phase of the PdM can be implemented on the basis of the formed at that time. In the case of unprocessed data,
data collected by the I4.0 and MES machine layers (Figure 2). maintenance activities must be transformed into category
When qualifying the suitability of a problem for a predictive columns, where each category corresponds to a type of
maintenance solution (result of data analysis), three essential maintenance action. The base data schema for mainte-
data sources must be found: nance records will include columns for asset IDs, time
1) Fault history: Generally, error events are very rare in pre- and maintenance actions.
dictive maintenance applications. However, when compil- • Fault records: these are records that belong to the target of
ing predictive models that estimate failures, it is necessary the estimate, i.e. failures or reason for the fault. They can
 2018 14th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA)

be specific error codes or error events defined by specific averages, standard deviations, outliers based on standard de-
operating conditions. In some cases the data includes viations, cumulative sum measures, minimum and maximum
multiple error codes, some of which correspond to faults window values, or trend changes, peaks and level changes
of interest. Not all faults are the goal of an estimate, using algorithms that detect anomalies, etc.
so others are typically used to create functions that can
be related to failures. The base data schema for fault A. Machine Learning Algorithms
records will include columns for asset ID, time and fault, The binary classification is viably used for predictive
or reason for failure if a pattern is available. maintenance, being able to estimate the probability that the
• Machine conditions: preferably real-time monitoring data equipment will fail over a future period of time. The time
related to the operating conditions of the data. period is determined and based on business rules and available
• Machine and operator data: these data can be combined data. Some common time periods are the minimum down-
in a scheme to identify the assets managed by a given time or the time required to perform the maintenance routines
operator together with the properties of the assets and necessary to solve the problem that it could occur in that
the operator. period of time. To use the binary classification, it is necessary
to identify two types of examples, which are called positive
and negative. Each example is a record of a unit of time for
Table I
F EATURES DATASET. an asset that conceptually describes the operating conditions
by designing the functions by using historical and other
Features Significance data sources. In the context of the binary classification for
predictive maintenance, the positive types denote the errors,
statoRot Functional spindle rotor status (c)
and the negative ones the normal operations. The aim is to find
Timestamp Event Recorded
a model that identifies the probability that each new example
Machine Running Machine
may fail or work normally within the next unit of time.
Spindle speed Spindle rotation speed
Regression models in predictive maintenance are used to
Spindle power Power absorbed by the spindle
calculate the remaining useful life of an asset, defined as the
Spindle position Spindle angular position
amount of time during which the asset remains operational
X PositionDiff X-axis real-to-nominal position diff
before the next failure occurs. As in the binary classification,
Y PositionDiff Y-axis real-to-nominal position diff
each example is a record that belongs to the unit of time for an
Z PositionDiff Z-axis real-to-nominal position diff
asset. In the context of regression, however, the aim is to find
X Speed X axis speed
a model that calculates the remaining useful life of each new
Y Speed Y axis speed
example as a continuous number. This time period is defined
Z Speed Z axis speed
as a multiple of the unit of time.
X Current Absorbed current X axis
Multi-class classification for predictive maintenance can be
Y Current Absorbed current Y axis
used to estimate two future results. The first is to assign an
Z Current Absorbed current Z axis
asset to one of the different time periods in order to allocate a
time interval for the failure of each asset. The second consists
In Table I a list of data is reported with the statoRot status in identifying the probability of failure in a future period due
variable that will be the machine condition variable in the to one of the multiple root causes. This allows the maintenance
result section. personnel to handle the problem in advance. Another multi-
The first step in modeling is the design of the functions. The class modeling technique focuses on determining the most
idea of generating functions consists conceptually in describ- likely root cause of a given failure. This allows to provide the
ing and abstracting the integrity condition of a machine at a suggestions for the main maintenance actions to be performed
given time using historical data collected up to that instant. The to repair a fault. By having a categorized list of root causes
design methods of the functions described below can however and associated repair actions, technicians can perform the
be used as a baseline for creating functions. In the following, first repair actions after failures more effectively. In predictive
delay (lag) functions need to be created from data sources maintenance, similar to any other area of the solution con-
including timestamps, and also static functions created from taining time-stamp data, the typical training and test routine
static data sources providing examples from the illustrated must consider the variable aspects over time for a better
use cases. In the preventive maintenance the chronological generalization of future unseen data. Many Machine Learning
data generally include timestamps that indicate the time of algorithms depend on a number of hyper parameters that
collection of each single data. For each asset record, a window can significantly modify the performance of the model. The
with dimension “W” is selected, corresponding to the number optimal values of these hyper parameters are not automatically
of units of time for which we want to calculate the chronologi- calculated during the training of the model, but must be
cal aggregations. The aggregate sequencing functions are then specified by the data scientist. Various ways are available to
calculated using the W periods before the date of that record. find the optimal values of the hyper parameters. The most
Examples of sequential aggregations are: incremental counts, common one is the “cross-validation of k sections”, which
 2018 14th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA)

randomly subdivides the examples into “k” sections. For each The main objectives of the proposed test is the evaluation of
set of values of the hyper parameters, the learning algorithm the proposed system for ML predictive maintenance of the
is executed k times. At each iteration, the examples in the cutting machine through FingerPrint analysis: analysis of drive
current section are used as a validation set, while the rest of data for axis monitoring and vibration data analysis to estimate
the examples are used as a training set. The algorithm performs spindle health status. The proposed feature set is reported in
training on training examples and performance metrics are TableI and a data set of 530731 data readings were corrected
calculated on validation examples. At the end of this cycle, on 15 different machine features collected in real time from
for each set of values of the hyper parameters the values of the tested cutting machine.
the performance metrics k are calculated and the values of Different sources of data are taken in consideration in the
the hyper parameters with the best average performances are data collecting process and the following overall architecture
selected. was designed:
In the result section a Decision Forest (DF) classifier is • FlightRecorder: it is an industrial PC placed inside the
used. The DF Classifier is an ensemble learning method MCM machine tool. It is communicating with various
used to improve classification performance. It uses bagging elements inside the machine in order to collect the data
by combining the output of several classifiers [22]. It works necessary for the analysis.
by building multiple decision trees and then it vots on the • IT Accelerometer: is an accelerometer placed on the
common output class. The main goal of ensamble methods is spindle head of the MCM machines that has to collect
that a high number of “weak learner” can be used to create vibration data, which exposes an interface for automatic
a “strong learner”. Generally, in each tree, a sequence of data export.
simple tests is run for each class, increasing the levels of a As for the architecture of the highest level will be intro-
tree structure until a leaf node (decision) is reached. In Azure duced:
Machine Learning Studio, this kind of classifier consists of an • Data analysis unit: it is a PC that hosts the data analysis
ensemble of decision trees. Ensemble models usually provide software that will be implemented.
better coverage and accuracy than single decision trees. • Simulation unit: it is a Azure Cloud architecture based
on Machine Learning Studio that hosts the ML software
that will be implemented (Figure 3).
B. Architecture and data flow
The FlightRecorder system has three different data sources,
as presented in Figure 4:
• Drive data: they are sampled in real time by the CNC
and made available to the FlightRecorder through a
communication protocol developed by D.Electron. This
data is associated with a time reference (timestamp) by
the CNC at the time of sampling.
• Data related to I/O signals: they are sampled in real
time by the machine PLC and made available to the
FlightRecorder the same way as the drive data.
• Vibrational data: vibrational analysis in the frequency
domain is performed to allow efficient monitoring of
the spindle unit. A dedicated sensor/electronic system
is required, which after each episode of fingerprint or
working step makes available to the FlightRecorder both
raw data related to the vibration spectrum trend, and an
evaluation on how much the spectrum detected is deviated
from the average one obtained under good operating
conditions.
Figure 3. General schema of Classification Process on Azure Machine
All the collected data, using the proposed architecture were
Learning Studio. evaluated using the 30% of the data set as a training set and
the rest for the results evaluation. Overall accuracy is reported
IV. E XPERIMENTAL R ESULTS in Table II. Results are accurate for predicting the rotor status
on 4 different classes based on the proposed feature set.
A. Use Case
The Predictive Maintenance approaches discussed above V. C ONCLUSIONS AND FUTURE WORKS
have been implemented on a real cutting machine woodwork- In this paper a new PdM methodology based on PdM
ing machinery that is a machining center on wood industry. machine learning approach on a cutting machine is presented.
 2018 14th IEEE/ASME International Conference on Mechatronic and Embedded Systems and Applications (MESA)

Table II R EFERENCES
C LASSIFICATION R ESULTS .
[1] H. M. Hashemian and W. C. Bean, “State-of-the-art predictive mainte-
nance techniques,” IEEE Transactions on Instrumentation and measure-
Metrics Results ment, vol. 60, no. 10, pp. 3480–3492, 2011.
[2] S.-j. Wu, N. Gebraeel, M. A. Lawley, and Y. Yih, “A neural network
Overall Accuracy 0.95 integrated decision support system for condition-based optimal predic-
Average Accuracy 0.92 tive maintenance policy,” IEEE Transactions on Systems, Man, and
Cybernetics-Part A: Systems and Humans, vol. 37, no. 2, pp. 226–236,
Micro-Averaged Precision 0.94 2007.
Macro-Averaged Precision 0.93 [3] E. Frontoni, R. Pollini, P. Russo, P. Zingaretti, and G. Cerri, “Hdomo:
Micro-Averaged Recall 0.95 Smart sensor integration for an active and independent longevity of the
elderly,” Sensors, vol. 17, no. 11, p. 2610, 2017.
Micro-Averaged Recall 0.94 [4] B. Lu, D. B. Durocher, and P. Stemper, “Predictive maintenance tech-
niques,” IEEE Industry Applications Magazine, vol. 15, no. 6, 2009.
[5] B. Lu, T. G. Habetler, and R. G. Harley, “A survey of efficiency-
estimation methods for in-service induction motors,” IEEE Transactions
on Industry Applications, vol. 42, no. 4, pp. 924–933, 2006.
[6] W. T. Thomson and M. Fenger, “Current signature analysis to detect
induction motor faults,” IEEE Industry Applications Magazine, vol. 7,
no. 4, pp. 26–34, 2001.
[7] R. Yam, P. Tse, L. Li, and P. Tu, “Intelligent predictive decision support
system for condition-based maintenance,” The International Journal of
Advanced Manufacturing Technology, vol. 17, no. 5, pp. 383–391, 2001.
[8] J.-H. Shin and H.-B. Jun, “On condition based maintenance policy,”
Journal of Computational Design and Engineering, vol. 2, no. 2, pp.
119–127, 2015.
[9] M. Paolanti, C. Kaiser, R. Schallner, E. Frontoni, and P. Zingaretti,
“Visual and textual sentiment analysis of brand-related social media
pictures using deep convolutional neural networks,” in International
Conference on Image Analysis and Processing. Springer, 2017, pp.
402–413.
[10] M. Paolanti, E. Frontoni, A. Mancini, R. Pierdicca, and P. Zingaretti,
“Automatic classification for anti mixup events in advanced manu-
facturing system,” in ASME 2015 International Design Engineering
Technical Conferences and Computers and Information in Engineering
Conference. American Society of Mechanical Engineers, 2015, pp.
V009T07A061–V009T07A061.
[11] M. Paolanti, M. Sturari, A. Mancini, P. Zingaretti, and E. Frontoni,
“Mobile robot for retail surveying and inventory using visual and textual
Figure 4. System Data Flow. analysis of monocular pictures based on deep learning,” in Mobile
Robots (ECMR), 2017 European Conference on. IEEE, 2017, pp. 1–6.
[12] S. Naspetti, R. Pierdicca, S. Mandolesi, M. Paolanti, E. Frontoni, and
R. Zanoli, “Automatic analysis of eye-tracking data for augmented real-
ity applications: A prospective outlook,” in International Conference on
PdM is a strategy viable adopted when dealing with mainte- Augmented Reality, Virtual Reality and Computer Graphics. Springer,
nance issues given the increasing need to minimize downtime 2016, pp. 217–230.
and associated costs. The methodology has been implemented [13] G. A. Susto, A. Schirru, S. Pampuri, S. McLoone, and A. Beghi,
“Machine learning for predictive maintenance: A multiple classifier
in the experimental environment on the example of a real approach,” IEEE Transactions on Industrial Informatics, vol. 11, no. 3,
industrial group, producing accurate estimations. Data has pp. 812–820, 2015.
been collected by various sensors, machine PLCs and commu- [14] K. Wang, “Intelligent predictive maintenance (ipdm) system–industry
4.0 scenario,” WIT Transactions on Engineering Sciences, vol. 113, pp.
nication protocols and made available to Data Analysis Tool. 259–268, 2016.
The proposed PdM methodology allows dynamical decision [15] G. Druck, G. Mann, and A. McCallum, “Learning from labeled features
rules to be adopted for maintenance management, which is using generalized expectation criteria,” in Proceedings of the 31st annual
international ACM SIGIR conference on Research and development in
achieved by training a Random Forest approach on Azure information retrieval. ACM, 2008, pp. 595–602.
Machine Learning Studio. Preliminary results show a proper [16] R. K. Mobley, An introduction to predictive maintenance. Butterworth-
behavior of the approach on predicting different machine states Heinemann, 2002.
[17] A. Grall, L. Dieulle, C. Bérenguer, and M. Roussignol, “Continuous-time
with high accuracy (95%) on a data set of 530731 data predictive-maintenance scheduling for a deteriorating system,” IEEE
readings on 15 different machine features collected in real transactions on reliability, vol. 51, no. 2, pp. 141–150, 2002.
time from the tested cutting machine. [18] J. S. Armstrong, Principles of forecasting: a handbook for researchers
and practitioners. Springer Science & Business Media, 2001, vol. 30.
Paper contributions are mainly on the overall cloud archi- [19] C. Chatfield, Time-series forecasting. CRC Press, 2000.
[20] F. E. Tay and L. Cao, “Application of support vector machines in
tecture for Industry 4.0, on the application of ML approaches financial time series forecasting,” Omega, vol. 29, no. 4, pp. 309–317,
to a real data set from machines on the field, on the high level 2001.
of accuracy on predicting the state of the main spindle rotor. [21] S. J. Haberman, “The analysis of residuals in cross-classified tables,”
Biometrics, pp. 205–220, 1973.
Future work will go in the direction of having more robust [22] L. Breiman, “Random forests,” Machine learning, vol. 45, no. 1, pp.
data-set, investigating different fault scenarios, exploring a 5–32, 2001.
different set of features, in particular in the frequency domain.

View publication stats