PROVA DE LÍNGUA INGLESA – ÁREA DE CIÊNCIAS EXATAS

MACHINE LEARNING APPROACH FOR PREDICTIVE MAINTENANCE IN INDUSTRY

4.0

INTRODUCTION

Predictive maintenance (PdM) or sometimes called “on-line monitoring”, “risk-based

maintenance”, or “condition-based maintenance”, is a subject to many recent research papers with the
long history behind it. It refers to the intelligent monitoring of equipment to avoid future failures.
Predictive maintenance has evolved from the first method that is visual inspection to automated
methods using the advanced signal processing techniques based on pattern recognition and machine
learning, neural networks, fuzzy logic, etc. The automated methods provide viable solution to many
industries detecting and collecting sensitive information from the equipment which are mainly motors,
where human eyes or ears can cease to do so. Together with integrated sensors, predictive maintenance
can avoid unnecessary equipment replacement, reduce machine downtime, find the root cause the fault,
and in this way save costs and improve efficiency. Predictive maintenance overlaps with the scope of
preventive maintenance in terms of scheduling the maintenance activity in advance to avoid machine
failures. In contrast to conventional preventive maintenance, predictive maintenance schedules
activities are based on collected data from sensors and analysis algorithms. In process industries,
induction motors make up approximately 70% of all driven electrical loads. In this regard, there has
been much interest on ways to better diagnose the wellness condition of these motors. The bearing
failure is identified as the most frequent cause of motor failure and most common maintenance
problem. Accordingly, predictive maintenance mainly focuses on two aspects: energy efficiency
improvement (key of energy saving) and unscheduled downtime reduction. The algorithms developed
around these two can also be generally divided into two categories:

I. energy and efficiency, there have been many evaluation methods and devices developed,
II. system condition monitoring, including the detection of motor faults, with various fault-
detection techniques developed.

Some initial concepts of intelligent predictive decision support systems have been introduced in
previous articles. Algorithms are essential for effective predictive maintenance. There are various kinds
of techniques to be applied in various phases of PdM implementation, i.e. data processing, diagnostics,
and prognostics, as given in. In PdM three kinds of approach can be distinguished:

1) Data-driven approach;
2) Model-based approach;
3) Hybrid approach.

The data-driven approach is also known as the data mining approach or the machine learning
approach, which uses historical data to learn a model of system behavior. Model-based approach has
the ability to incorporate physical understanding of the target product, relying on the analytical model
to represent the behavior of the system. Machine learning approaches are viably used in the areas
where the availability of data is increasing such as the maintenance in industry sector. It is increasingly
providing effective solutions, cloud-based solutions, and newly introduced algorithms. Machine
Learning-based PdM can be divided into the following two main classes: Supervised, where
information on the occurrence of failures is present in the modelling dataset and Unsupervised - where
logistic and/or process information is available, but no maintenance data exists. The availability of
maintenance information mostly depends on the nature of the existing maintenance management
policy. When possible, supervised solutions are preferable. From the Machine Learning perspective,
depending on the output of the data set, two classes of supervised problem are possible: regression
problem (if output assumes continuous values), and classification problem (if output assumes
categorical values).
In this paper a new PdM methodology based on PdM machine learning approach on a cutting
machine is presented. PdM is a strategy viable adopted when dealing with maintenance issues given the
increasing need to minimize downtime and associated costs. The methodology has been implemented
in the experimental environment on the example of a real industrial group, producing accurate
estimations.
Data has been collected by various sensors, machine PLCs and communication protocols and made
available to Data Analysis Tool. The proposed PdM methodology allows dynamical decision rules to
be adopted for maintenance management, which is achieved by training a Random Forest approach on
Azure Machine Learning Studio. Preliminary results show a proper behavior of the approach on
predicting different machine states with high accuracy (95%) on a data set of 530731 data readings on
15 different machine features collected in real time from the tested cutting machine.
Paper contributions are mainly on the overall cloud architecture for Industry 4.0, on the application
of ML approaches to a real data set from machines on the field, on the high level of accuracy on
predicting the state of the main spindle rotor.

II. THE DATA ANALYSIS

Predictive Maintenance is an important maintenance tool that is based on the possibility of

estimating the future values of some quantities that characterize a system (typically a machine, a plant,
or a production process) through particular mathematical models in order to identify in advance the
anomalies and potential faults. The basic scheme of PdM is as follows:
 Measurement of physical quantities in real time.
 Estimation of measurable (or non-measurable) parameters at time t + dt.
 Identification of the system status considered anomalous or faulty.
 Planning of preventive and corrective activities before the system reaches the critical condition.

Examples of predictive maintenance are the following:

 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 algorithms 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:
 Cross-Sectional Forecasting;
 Time Series Forecasting.
A. Cross-Sectional Forecasting
Cross-Sectional Forecasting is the estimation of parameters of which there exist no measurements,
by using measures on other variables that have been observed. As an example, it can be possible to
predict the life of an electronic component used in particular conditions, by measuring the electric
current that passes through it.

B. Time Series Forecasting

Time Series Forecasting is referring to the estimation of parameters that change over time, being
measured until time instant t and the value to be predicted is at the time instant t + dt. Typically, the
measurements of the variable of interest can be obtained at regular intervals, and then it is possible to
predict the future values. The simplest example is the prediction of the minutes of residual charge of
our mobile phones, which is estimated based on the consumption that we have made, and the way we
use the phone. In time series, it is usual to identify:
 Trends, or a long-term increase (or decrease) in values.
 The seasonal phenomena, or the phenomena that determine 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.

III. MACHINE LEARNING APPROACHES FOR PDM

When qualifying the suitability of a problem for a predictive maintenance solution (result of data
analysis), three essential data sources must be found:
 Fault history: Generally, error events are very rare in predictive maintenance applications.
However, when compiling predictive models that estimate failures, it is necessary for the
algorithm to learn the normal operating scheme, in addition to the failure scheme through the
training process. Consequently, it is essential that the training data contain a sufficient number
of examples in both categories.
 Maintenance/repair history: An essential source of data for predictive maintenance solutions is
the detailed asset of maintenance history, which contains information about replaced
components, preventive maintenance tasks performed, and so on.
 Machine conditions: to estimate how many days (or hours, kilometers, etc.) a machine lasts
before a failure occurs, it is assumed that its health status decreases with time. It is therefore
necessary that the data contain time-varying functions that acquire aging patterns or any
anomaly that could cause performance reduction.

In order to prepare the data in the format required to create the functions to be included in the
Machine Learning algorithm, it is necessary to perform some pre-processing steps. The first is to divide
the duration of data collection into units of time in which each record is a unit of time for an asset. The
generic data schema could be:
  Maintenance records: these are the records of the maintenance actions performed. The
unprocessed maintenance data is usually associated with an asset ID and a timestamp with
information on maintenance activities performed at that time. In the case of unprocessed data,
maintenance activities must be transformed into category columns, where each category
corresponds to a type of maintenance action. The base data schema for maintenance records
will include columns for asset IDs, time and maintenance actions.
 Fault records: these are records that belong to the target of the estimate, i.e. failures or reason
for the fault. They can be specific error codes or error events defined by specific operating
conditions. In some cases the data includes multiple error codes, some of which correspond to
faults of interest. Not all faults are the goal of an estimate, so others are typically used to create
functions that can be related to failures. The base data schema for fault records will include
columns for asset ID, time and fault, or reason for failure if a pattern is available.
 Machine conditions: preferably real-time monitoring data related to the operating conditions of
the data.
 Machine and operator data: these data can be combined in a scheme to identify the assets
managed by a given operator together with the properties of the assets and the operator.

The first step in modeling is the design of the functions. The idea of generating functions consists
conceptually in describing and abstracting the integrity condition of a machine at a given time using
historical data collected up to that instant. The design methods of the functions described below can
however be used as a baseline for creating functions. In the following, delay (lag) functions need to be
created from data sources including timestamps, and also static functions created from static data
sources providing examples from the illustrated use cases. In the preventive maintenance the
chronological data generally include timestamps that indicate the time of collection of each single data.
For each asset record, a window with dimension “W” is selected, corresponding to the number of units
of time for which we want to calculate the chronological aggregations. The aggregate sequencing
functions are then calculated using the W periods before the date of that record. Examples of sequential
aggregations are: incremental counts, averages, standard deviations, outliers based on standard
deviations, cumulative sum measures, minimum and maximum window values, or trend changes, peaks
and level changes using algorithms that detect anomalies, etc.

V. CONCLUSIONS AND FUTURE WORKS

In this paper a new PdM methodology based on PdM machine learning approach on a cutting
machine is presented.
PdM is a strategy viable adopted when dealing with maintenance issues given the increasing need
to minimize downtime and associated costs. The methodology has been implemented in the
experimental environment on the example of a real industrial group, producing accurate estimations.
Data has been collected by various sensors, machine PLCs and communication protocols and made
available to Data Analysis Tool. The proposed PdM methodology allows dynamical decision rules to
be adopted for maintenance management, which is achieved by training a Random Forest approach on
Azure Machine Learning Studio. Preliminary results show a proper behavior of the approach on
predicting different machine states with high accuracy (95%) on a data set of 530731 data readings on
15 different machine features collected in real time from the tested cutting machine.
Paper contributions are mainly on the overall cloud architecture for Industry 4.0, on the application
of ML approaches to a real data set from machines on the field, on the high level of accuracy on
predicting the state of the main spindle rotor.
 Future work will go in the direction of having more robust data-set, investigating different fault
scenarios, exploring a different set of features, in particular in the frequency domain.

 Com base no texto “MACHINE LEARNING APPROACH FOR PREDICTIVE

MAINTENANCE IN INDUSTRY 4.0”, responda, em língua portuguesa, as seguintes questões:

QUESTÃO 01 – Explique, com suas próprias palavras, o que é manutenção preditiva (PdM) e
como ela difere da manutenção preventiva tradicional. Cite pelo menos dois benefícios específicos
da manutenção preditiva em um ambiente industrial. (2,00)

QUESTÃO 02 – O artigo menciona três abordagens principais para a implementação da

manutenção preditiva: "data-driven", "model-based" e "hybrid approach". Descreva a abordagem
"data-driven" e forneça um exemplo de como ela poderia ser aplicada na análise de dados de uma
máquina. (2,00)

QUESTÃO 03 – No contexto de "cross-sectional forecasting" e "time series forecasting"

mencionados no artigo, explique a diferença entre essas duas técnicas de previsão e forneça um
exemplo prático de aplicação de cada uma delas em manutenção preditiva. (2,00)

QUESTÃO 04 – Quais são as três fontes de dados essenciais para implementar uma solução de
manutenção preditiva? Explique brevemente a função de cada uma delas na construção de modelos
preditivos. (2,00)

QUESTÃO 05 – Com base nos resultados experimentais apresentados no artigo, quais foram as
principais conclusões sobre a eficácia da metodologia proposta? Comente sobre a precisão da
metodologia e seu impacto na indústria. (2,00)