EDITOR: Longbing Cao, [Link]@[Link].

au

COLUMN: AI FOCUS

Data-Driven Predictive Maintenance

W
ith the growth of 5G networks, the Inter- advances, we have tools that can process real-
net-of-Things is becoming a reality. The time data acquired from different equipment
advances in networking, machine learn- parts to predict any sign of failure.
ing, data analytics, and robotics largely improve indus-
trial processes. Industry 4.0 is a term for the fourth PdM differs from preventive maintenance because
industrial revolution: the digitization and automation it relies on the actual condition of equipment, rather
of manufacturing.1 Predictive maintenance is one of than average or expected life statistics, to predict
the techniques with high impact in today’s industry. when maintenance will be required. Typically, machine
learning approaches are adopted for the definition of
the actual condition of the system and for forecasting
INDUSTRY 4.0 AND MAINTENANCE its future states.
Maintenance is the process that deals with the health of
equipment and system components to ensure their nor- DATA-DRIVEN PDM
mal functioning under any circumstances. Over the years,
There are two main approaches for PdM: 1) Model
and due to technological advances, different mainte- based, and 2) data driven. In the model-based PdM,
nance strategies have been developed. Three main main- the equipment is described by physical equations, a
tenance approaches can be classified in the following.2 digital representation of the equipment. Digital twins3
are the most popular model-based approach. Data-
› Corrective maintenance: It means run-to-failure, driven PdM monitors the mechanical conditions or
which is the simplest and the oldest method. other health indicators of the equipment, and uses
Actions are taken only after a machine or equip- advanced statistical methods, such as machine learn-
ment fails. It would almost always lead to high ing, to find a pattern of functioning and dynamically
(unexpected) downtime. This method usually determine the operating conditions over time.
generates a critical situation that will demand a Unlike the model-based maintenance approaches
great cost for companies. (e.g., preventive maintenance) that rely on forecasting
› Preventive maintenance: It provides planning the performance degradation using stochastic models,
of regular replacement of components and/or data-driven PdM practices are based on data without
equipment. It considers historical failure data prior knowledge of degradation conditions. Its perfor-
and/or the data provided by the equipment mance strictly depends on the analysis of signals and
manufacturer. Although this approach prevents data. While model-based solutions can be expensive and
unexpected shutdown, it usually needs additional inaccurate for complex systems, data-driven diagnosis
costs and an increased unexploited lifetime. methods are a promising alternative to fault/anomaly
› Predictive maintenance (PdM): It needs direct detection and isolation. Machine learning and deep
monitoring of the mechanical condition and learning algorithms and tools are naturally relevant to
other parameters to determine the operating con- the PdM practices, mainly due to the large amount of
ditions over time. Given the recent technological data available (specifically the unlabelled ones).4

Failure Prediction
1541-1672 ß 2022 IEEE
Digital Object Identifier 10.1109/MIS.2022.3167561 Fault detection is one of the most critical components
Date of publication 25 May 2022; date of current version 12 of PdM. Nevertheless, PdM goes much behind pre-
September 2022. dicting a failure. It is essential to understand the

Authorized licensed
July/August 2022 use limited to: Hochschule
PublishedHeilbronn. Downloaded
by the IEEE Computeron September
Society 22,2025 at 09:53:17
IEEEUTC from IEEE
Intelligent 27
Xplore. Restrictions apply.
Systems
AI FOCUS

consequences and what will be the collateral dam- (corrective maintenance) and time-to-failure (prog-
ages of the failure. One of the approaches in data- nostics) strategies. A categorization of methods
driven PdM is to learn the normal operating condition and techniques for RUL can be found in Nguyen
of a system such that any significant deviation from et al.’s work.6
that operating condition could be spotted as a poten-
tial failure. The most popular techniques include Root Cause Analysis
autoencoders,5 trained with data from the normal This dimension looks for the causes and explanations
operating condition of the system. During the model about what could have happened when an anomaly
exploitation, the system signals an alarm if the recon- was detected. The main concept that allowed this
struction error of the autoencoder exceeds a delimitation is inference, a fundamental task of AI.
threshold. There exist many available techniques with different
PdM is generally employed based on the health implementations. However, it is not common due to
status of critical elements. To avoid possible interrup- limitations to obtain any explanation from data.
tions or even more severe damage, based on the oper-
ational history of different components, this strategy OPEN ISSUES AND FUTURE
can predict failures over time, minimizing costs, and DIRECTIONS
extending the useful life of the components. With the development of the Internet of Things,
manufacturing technology, and mass production, the
Remain Useful Life—RUL methodology of maintenance scheduling and manage-
Different maintenance management policies can be ment has become an essential topic in the industry. In
employed using anomaly detection, diagnostics, and these contexts, there is an urgent need for under-
prognostics. The RUL is strongly related to prognos- standing and trust models and their results.7 Under-
tics, which provides the amount of time equipment standing a predicted failure requires explaining where,
will be operational before it requires any repair or why, and when a failure occurs. This is a relevant
replacement. Prognostic is directly related to the likeli- research topic aligned with the current trends in AI.8
hood of system failure occurrence. It can be regarded Another critical issue in PdM applications is the
as a forecasting process given the current machine design of the maintenance plan after a fault is
conditions and its historical record. detected or predicted. To elaborate the recovery plan,
A well-known technique to decide on the mainte- it is crucial to know the causes of the problem (root
nance approach is P–F curve analysis (cf. Figure 1), cause analysis), which component is affected, and the
which allows understanding of the condition of expected remain useful life of the equipment. Few
equipment over time. During the time between the works in PdM address this issue.
detection of the potential failure and the actual Given the growth in the digital virtualization of the
failure, it is crucial to perform a maintenance action world, the recent technology of digital twins3 is the
to address the problem before a functional failure future for PdM. A digital twin is a virtual model
occurs. As a fundamental task for RUL, prediction designed to reflect a physical machine accurately.
clearly defines the difference between run-to-failure Vital areas of the machine are monitored using sen-
sors. The data collected by the sensors is used to feed
the digital copy of the machine. The digital twin can
simulate scenarios, what-if analysis, etc.

ACKNOWLEDGMENTS
This work was supported in part by CHIST-ERA under
Grant CHIST-ERA-19-XAI-012 and in part by project
CHIST-ERA/0004/2019 funded by FCT.

REFERENCES
1. A. Rojko, “Industry 4.0 concept: Background and
FIGURE 1. P–F curve is a graph that shows the health of overview,” Int. J. Inf. Manage., vol. 11, pp. 77–90, 2017.
equipment over time to identify the interval between poten- 2. R. K. Mobley, An Introduction to Predictive
Maintenance, 2nd ed. Amsterdam, The Netherlands:
tial failure and functional failure.
Elsevier, 2002.

28
Authorized licensed use limited
IEEE to: Hochschule
Intelligent Systems Heilbronn. Downloaded on September 22,2025 at 09:53:17 UTC from IEEE Xplore. Restrictions
July/August apply.
2022
 AI FOCUS

€ ntner, “I-maintenance:
3. F. Strohmeier, C. Schranz, and G. Gu ~ GAMA is a full professor at University of Porto, Porto,
JOAO
A digital twin for smart maintenance,” ERCIM News, Portugal. His main research topics are learning from data
vol. 115, pp. 12–14, 2018. [Online]. Available: [Link]
streams and learning from network data. He is an EurIA fel-
[Link]/en115/special/2095-i-maintenance-a-
low, IEEE fellow, fellow of the Asia-Pacific AI Association, and
digital-twin-for-smart-maintenance
member of the board of directors of the LIAAD, a group
4. N. Davari, B. Veloso, G. de A. Costa, P. M. Pereira, R. P.
Ribeiro, and J. Gama, “A survey on data-driven belonging to INESC-TEC. Contact him at jgama@[Link].
predictive maintenance for the railway industry,”
Sensors, vol. 21, no. 17, 2021, Art. no. 5739, doi: 10.3390/ RITA P. RIBEIRO is an assistant professor at the Department of
s21175739. Computer Science of the University of Porto, Porto, Portugal.
5. L. Wang, X. Zhao, J. Pei, and G. Tang, “Transformer fault
She is the director of the master on data sciences and the
diagnosis using continuous sparse autoencoder,”
senior researcher of LIAAD, a group belonging to INESC-TEC.
SpringerPlus, vol. 5, no. 1, pp. 1–13, 2016.
Her research interests include imbalanced domain learning,
6. D. N. Nguyen, L. Dieulle, and A. Grall, “Remaining useful
lifetime prognosis of controlled systems: A case of outlier detection, evaluation issues on learning tasks, and prob-
stochastically deteriorating actuator,” Math. Problems lems related to social good. Contact her at rpribeiro@[Link].
Eng., vol. 2015, 2015, Art. no. 356916.
7. C. Molnar, Interpretable Machine Learning, 2nd ed., BRUNO VELOSO is an assistant professor at University
Munich, German, 2022. [Online]. Available: https://
Portucalense, Porto, Portugal and an invited assistant profes-
[Link]/interpretable-ml-book
sor at University of Porto. His research interests include data
8. R. Guidotti, A. Monreale, S. Ruggieri, F. Turini,
streams and automated machine learning. He is a member of
F. Giannotti, and D. Pedreschi, “A survey of methods
for explaining black box models,” ACM Spanish Association for Artificial Intelligence AEPIA, and
Comput. Surv., vol. 51, no. 5, pp. 93:1–93:42, 2019, Senior Researcher of LIAAD, a group belonging to INESC-
doi: 10.1145/3236009. TEC. Contact him at [Link]@[Link].

Authorized licensed
July/August 2022 use limited to: Hochschule Heilbronn. Downloaded on September 22,2025 at 09:53:17
IEEEUTC from IEEE
Intelligent 29
Xplore. Restrictions apply.
Systems