HyPredictor: Hybrid Failure Prognosis Approach Combining
Data-Driven and Knowledge-Based Methods
Miguel Almeida
1 a
, Eliseu Pereira
1,2 b
and Gil Gonc¸alves
1,2 c
1
Faculty of Engineering, University of Porto, Porto, Portugal
2
SYSTEC - ARISE, Faculty of Enginnering of the University of Porto, Porto, Portugal
Keywords:
Failure Prediction, Hybrid Approaches, Knowledge-Based Methods, Data-Driven Methods, Explainable
Artificial Intelligence.
Abstract:
In modern manufacturing, marked by an unprecedented surge in data generation, utilising this wealth of in-
formation to enhance company performance has become essential. Within the industrial landscape, one of the
significant challenges is equipment failures, which can result in substantial financial losses and wasted time
and resources. This work presents the HyPredictor framework, a comprehensive failure prediction and report-
ing system designed to enhance the reliability and efficiency of industrial operations by leveraging advanced
machine learning techniques and domain knowledge. Six machine learning algorithms were evaluated for
failure prediction. The predictions from the algorithms are then refined using rule-based adjustments derived
from domain knowledge. Additionally, Explainable Artificial Intelligence (XAI) techniques were incorpo-
rated, as well as the capability of users to customise the system with their own rules and submit failure reports,
prompting model retraining and continuous improvement. Integrating domain-specific rules improved the per-
formance by up to 28 percentage points in the F1 Score metric in some prediction models, with the best hybrid
approach achieving an F1 Score of 90% and a Recall of 92% in failure prediction. This adaptive, hybrid ap-
proach improves prediction accuracy and fosters proactive maintenance, significantly reducing downtime and
operational costs.
1 INTRODUCTION
In industrial operations, equipment failures cause sig-
nificant downtime and financial losses. Predictive
maintenance is a proactive strategy that anticipates
faults before they occur, minimising disruptions. In
the era of Industry 4.0, technologies like the Inter-
net of Things, Artificial Intelligence (AI), and Ma-
chine Learning enhance productivity and operational
intelligence by leveraging real-time data and combin-
ing data-driven and knowledge-driven methods, such
as analysing sensor data and incorporating expert in-
sights.
Integrating these advanced technologies and data-
driven insights promises to revolutionise industrial
operations but also presents challenges. The vast
quantity and complexity of data requires sophisticated
tools to avoid information overload and extract rel-
a
https://orcid.org/0009-0003-6559-6271
b
https://orcid.org/0000-0003-3893-3845
c
https://orcid.org/0000-0001-7757-7308
evant insights. Traditional machine learning meth-
ods may struggle with this complexity, and static
models may fail to capture the dynamic nature of
industrial processes. Relying solely on data-driven
or knowledge-driven approaches has inherent limita-
tions. Combining both methodologies offers a more
comprehensive understanding of equipment health
and enhances decision-making processes. Address-
ing these challenges is crucial for effectively utilising
the abundant data available.
By integrating data-driven and knowledge-driven
methodologies, this work aims to address these chal-
lenges and provide a comprehensive understanding
of the equipment’s health. By developing accurate
predictive models, this work also seeks to minimise
downtime, reduce maintenance costs, and promote
sustainable industrial practices by mitigating the en-
vironmental impact of equipment failures.
Data-driven methods process various sensor data,
extracting valuable insights from the information gen-
erated by industrial equipment. Concurrently, expert
knowledge will be integrated through knowledge-
Almeida, M., Pereira, E. and Gonçalves, G.
HyPredictor: Hybrid Failure Prognosis Approach Combining Data-Driven and Knowledge-Based Methods.
DOI: 10.5220/0012924300003822
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 21st International Conference on Informatics in Control, Automation and Robotics (ICINCO 2024) - Volume 1, pages 245-252
ISBN: 978-989-758-717-7; ISSN: 2184-2809
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
245

specific rules, enriching the analytical framework
with insights and domain expertise. Finally, XAI
techniques will be deployed to provide transparent ex-
planations of the model’s predictions. This integrated
approach of data-driven methods, knowledge-driven
rules, and XAI forms a robust analytical framework
that leverages the unique advantages of each method,
aligning with the principles of Industry 4.0.
This paper is organised as follows: Section 2
presents the literature review, Section 3 describes the
dataset, Section 4 outlines the implementation, Sec-
tion 5 presents the results, and Section 6 concludes
and discusses potential future work.
2 LITERATURE REVIEW
This literature review section discusses predictive
maintenance and explores data-driven methodologies,
knowledge-driven strategies, and XAI methods.
2.1 Predictive Maintenance
In the realm of maintenance, as outlined by (Zonta
et al., 2020), four primary categories have been iden-
tified: corrective, preventive, predictive, and prescrip-
tive strategies. Predictive maintenance represents a
transformative shift in industrial asset management,
employing advanced data analytics, machine learn-
ing, and sensor technologies to anticipate equipment
failures proactively. By utilising extensive datasets
from sensors and other industrial sources, data-driven
methodologies uncover intricate patterns and anoma-
lies, forming the foundation of predictive mainte-
nance. Model-driven approaches utilise mathemati-
cal or computational models to simulate equipment
behaviour and optimise maintenance strategies. In
contrast, knowledge-driven methods integrate human
expertise, providing qualitative insights and contex-
tual understanding to enhance predictive accuracy, es-
pecially in scenarios requiring careful considerations
and predictions of rare events.
2.2 Data-Driven Methods
Data-driven methods in predictive maintenance utilise
the abundance of data generated in industrial en-
vironments to identify patterns, correlations, and
anomalies, facilitating the prediction of equipment
health and potential failures. This approach encom-
passes both supervised and unsupervised learning
techniques, with supervised methods like classifica-
tion and regression being particularly prevalent (An-
gelopoulos et al., 2020). Notably, machine learning
algorithms such as Random Forest (RF), Support Vec-
tor Machines (SVMs), Gradient Boosting (GB), Arti-
ficial Neural Networks (ANNs), and Logistic Regres-
sion (LR) are among the most widely utilised in pre-
dictive maintenance (Leukel et al., 2021).
Unsupervised methods operate without labelled
data, making them particularly valuable in scenarios
where the outcomes are not well-defined or unknown.
Clustering and anomaly detection are prevalent unsu-
pervised techniques in predictive maintenance, with
Principal Component Analysis (PCA) and K-Means
clustering being among the most widely used meth-
ods. Moreover, PCA was utilised to reduce the di-
mensionality of features (Canizo et al., 2017; Li et al.,
2014). In the study by (Bekar et al., 2020), a K-Means
clustering technique is employed to obtain diagnostic
information, which can be utilised for labelling the
data and supporting practitioners in predictive main-
tenance decision-making.
2.3 Knowledge-Drive Methods
Knowledge-driven methods in artificial intelligence
and machine learning play a pivotal role in decision-
making by utilising expert knowledge and domain
expertise. These methods contrast data-driven ap-
proaches, as they involve incorporating explicit rules,
ontologies, and logical reasoning into the decision-
making process.
One popular example of knowledge-based meth-
ods is the development of Rule-Based Systems, where
predefined rules guide decision-making (Sun and Ge,
2021). These rules can be derived from expert knowl-
edge, logical reasoning, or established guidelines
within a specific domain.
Ontologies, structured representations of knowl-
edge defining relationships and entities within a do-
main, are frequently used in knowledge-based meth-
ods (Chi et al., 2022). Expert Systems are a spe-
cialised category of knowledge-based methods de-
signed to emulate the decision-making capabilities of
human experts. However, traditional knowledge rep-
resentations built upon expert systems demand a spe-
cific data structure design, and most of these systems
possess intricate architectures, restricting the ease of
knowledge sharing and reuse (Chi et al., 2022).
2.4 Explainable Artificial Intelligence
XAI plays a crucial role in enhancing the transparency
and trustworthiness of failure prediction systems in
industrial settings. XAI techniques provide insights
into how and why specific predictions are made,
enabling humans to understand the underlying fac-
ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics
246

tors influencing the model’s decisions (Ahmed et al.,
2022; Barredo Arrieta et al., 2020).
Techniques such as Local Interpretable Model-
agnostic Explanations (LIME) and Shapley Additive
Explanations (SHAP) are commonly used to break
down model predictions into understandable compo-
nents, thus closing the gap between complex ma-
chine learning algorithms and practical, applicable in-
sights. Integrating XAI into failure prediction sys-
tems not only enhances their reliability but also em-
powers users to leverage these systems more effec-
tively, encouraging a proactive approach to main-
tenance and operational efficiency (Barredo Arrieta
et al., 2020).
2.5 Gap Analysis
Current predictive maintenance methodologies have
notable gaps. While data-driven approaches are
prevalent, they often overlook the valuable insights
domain experts provide. Moreover, the lack of trans-
parency in machine learning models poses a signif-
icant challenge, hindering trust and interpretability.
Additionally, many existing systems lack mechanisms
for continuous improvement based on real-time feed-
back, resulting in models that struggle to adapt to
changing operational environments. These gaps high-
light the need for innovative hybrid solutions that in-
tegrate data-driven methods with domain expertise
while also prioritising transparency and adaptability.
3 DATASET
This work utilised a public dataset from a metro
train of Porto in an operational context (Davari et al.,
2021). The dataset includes 15 signals, such as pres-
sure readings, temperature, motor current, and air in-
take valves, collected from a compressor’s Air Pro-
duction Unit between February and August 2020. The
data was logged at a frequency of 1Hz by an onboard
embedded device, resulting in 1,516,948 instances.
4 IMPLEMENTATION
The HyPredictor methodology encompasses several
critical stages, including data reception and pre-
processing, model development, rule-based adjust-
ments, explainability through XAI, user-implemented
rules, and failure reporting with model retraining, as
can be seen in Figure 1. Each stage ensures robust,
accurate, and transparent predictions, promoting con-
tinuous improvement and user engagement.
1
Figure 1: Proposed implementation structure.
Given that the dataset was initially unlabeled, the
process began by labelling it using the failure reports
provided by the company. This involved correlat-
ing the collected sensor data with documented failure
events to create a labelled dataset. To allow operators
to address issues before an actual failure occurred, la-
bels were assigned to indicate a failure beginning two
hours before the reported failure time. This preemp-
tive labelling strategy aimed to provide a sufficient
lead time for operators to intervene and potentially
prevent failure, thereby enhancing the system’s effec-
tiveness in facilitating proactive maintenance. To en-
hance the model’s predictive capabilities, feature en-
gineering was performed. This process involved ag-
gregating the data into 15-minute intervals, comput-
ing the median value within each interval, and inte-
grating these median values into the original dataset.
Moreover, a thorough analysis was conducted to
enhance the model’s predictive capabilities for fail-
ure prediction. Initially, the target variable was iden-
tified as ’Failure’, and correlation coefficients be-
tween this variable and all other features in the dataset
were calculated. The features most strongly asso-
ciated with failure occurrences were determined by
ordering these correlation coefficients by their abso-
lute values. Subsequently, a correlation threshold of
|
0.2
|
was applied to filter out features with low cor-
1
To access the code for this
project, please visit the repository at:
https://github.com/miguelalmeida8/HyPredictor-
Framework
HyPredictor: Hybrid Failure Prognosis Approach Combining Data-Driven and Knowledge-Based Methods
247

relation to the target variable. These low-correlation
features were then dropped from the dataset to opti-
mise the model’s training process. Additionally, to
mitigate multicollinearity issues, the correlation ma-
trix of the remaining features was visualised through
a heatmap, as can be seen in Figure 2, and pairs of
features with high correlation coefficients were iden-
tified using a threshold of
|
0.8
|
to denote high correla-
tion. Finally, the features that exhibit a high correla-
tion with others were removed. This feature selection
and multicollinearity mitigation process aimed to op-
timise the model’s predictive performance and ensure
the robustness of the failure prediction system for in-
dustrial applications.
This step involved dividing the dataset into train-
ing and testing sets using a temporal split. The test-
ing data, comprising 638,486 samples from June 4,
2020, onward, represented approximately 42% of the
dataset. Six machine learning algorithms were eval-
uated to determine the most suitable method for ac-
curate failure prediction: Random Forest, XGBoost,
CatBoost, Gradient Boosting Machine (GBM), Light-
GBM and a voting algorithm. The voting algorithm
was used to combine the strengths of multiple mod-
els, aiming to enhance overall prediction reliability.
Due to their strong overall performance, the four al-
gorithms chosen for the ensemble were XGBoost,
CatBoost, GBM, and LightGBM. The models were
trained using the pre-processed data, and k-fold cross-
validation was implemented to ensure that the model
generalises well to unseen data. Additionally, grid
search was employed to optimise the model param-
eters, enhancing its performance. Recall was used as
the primary metric to evaluate the model’s effective-
ness in accurately identifying failures, ensuring a fo-
cus on minimising false negatives in the predictions.
For the knowledge-driven approach, rules were
defined based on expertise and experience to adjust
the model’s predictions. These rules are crucial for
improving accuracy and handling edge cases, en-
suring that rare but critical scenarios are correctly
addressed. The system applies these rules post-
prediction, refining the initial outputs. The imple-
mented rules are designed to enhance the model’s pre-
dictive accuracy by incorporating domain-specific in-
sights, particularly concerning the median oil temper-
ature and pressure readings. These rules adjust the
model’s initial predictions based on predefined con-
ditions indicative of potential failures. The specific
rules applied are as follows:
• Rule 1: If the initial prediction is 0 (no failure)
but the median oil temperature exceeds 83°C, the
prediction is adjusted to 1 (failure).
• Rule 2: If the initial prediction is 1 (failure) but
the median oil temperature is below 67.25°C, the
prediction is adjusted to 0 (no failure).
• Rule 3: If the initial prediction is 0 (no fail-
ure) and both the median oil temperature exceeds
75.65°C and the median differential pressure ex-
ceeds -0.02 Bar, the prediction is adjusted to 1
(failure).
Subsequently, XAI methods like LIME reveal the
rules and contributions of individual features in pre-
dictions, as illustrated in Figure 3.
A user interface (UI) displays these explanations,
allowing experts to understand the model’s behaviour
behind specific predictions. The visual representa-
tion of feature contributions provides clear insights
into the factors influencing each prediction, making
the model’s decision-making process transparent and
comprehensible.
The system also empowers experts to implement
new rules based on their observations and insights
from the model explanations. These insights can then
be codified into new rules integrated into the system
to refine its predictive accuracy. This continuous feed-
back loop is crucial in enhancing the system’s perfor-
mance. As experts identify and address new scenar-
ios or anomalies, the rules evolve, making the model
more robust over time.
This dynamic interaction between the model and
the experts promotes a proactive maintenance strat-
egy. The system can predict failures more accu-
rately by preemptively adjusting the model based on
real-world observations, reducing downtime and op-
erational costs. This integration of human expertise
with machine learning not only optimises the pre-
dictive model but also ensures that the system re-
mains aligned with the evolving operational context
and complexities of the industrial environment.
The interface also allows experts to submit a fail-
ure report whenever they identify a failure. This re-
porting mechanism is crucial for maintaining the sys-
tem’s accuracy and responsiveness. When a failure re-
port is submitted, the system triggers a retraining pro-
cess for the model. The model adapts and learns from
recent occurrences by incorporating the new failure
data, continuously improving its predictive accuracy
and reducing future prediction errors. This process
not only updates the model with the latest data but
also enhances its ability to recognise similar patterns
and anomalies in the future. The integration of real-
time feedback ensures that the model remains relevant
and effective in an ever-changing operational environ-
ment. This capability is particularly important in in-
dustrial settings, where conditions and failure modes
can evolve rapidly.
ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics
248

Figure 2: Correlation matrix.
Figure 3: LIME explanation.
The described methodology ensures a comprehen-
sive and adaptative failure prediction and reporting
system. The system delivers accurate, transparent,
and continuously improving predictions by combin-
ing advanced data processing, machine learning, and
domain knowledge in the form of rules, XAI, and user
interaction.
5 RESULTS
This section presents the performance and results of
the failure prediction and reporting system.
HyPredictor: Hybrid Failure Prognosis Approach Combining Data-Driven and Knowledge-Based Methods
249

5.1 Model Performance Metrics
The model’s performance was assessed using sev-
eral key metrics: Accuracy, Precision, Recall, and
F1 Score. The evaluation results for the six differ-
ent models, Random Forest, XGBoost, LightGBM,
GBM, CatBoost and the voting algorithm, are pre-
sented in the ”Before” column of Table 1 and 2. This
comparison highlights the strengths and weaknesses
of each model, allowing for an informed selection
based on the specific requirements of the predictive
maintenance task.
The evaluation of the models showcased signifi-
cant variations in their performance metrics. While
XGBoost, GBM, LGBM, CatBoost and the voting al-
gorithm demonstrated good performance across mul-
tiple metrics, with high recall rates and F1 scores,
Random Forest fell short in terms of precision and F1
score, indicating a higher incidence of false positives.
5.2 Rule-Based Adjustments
Applying rule-based adjustments significantly en-
hanced the model’s performance in specific scenarios.
By integrating domain-specific rules, the system im-
proved its handling of rare but critical failure scenar-
ios. For instance, rules were implemented to address
conditions where certain sensor readings, when com-
bined, indicated an impending failure that the model
alone did not detect. The combination of machine
learning predictions and rule-based adjustments re-
sulted in an overall accuracy improvement.
Table 1 and 2 presents the performance metrics
before and after applying the rule-based adjustments.
The improvements in Recall, Precision, and F1 Score
across different models highlight the added value of
expert knowledge in enhancing predictive accuracy.
The decision to deploy GBM was primarily influ-
enced by its highest recall rate after applying the do-
main knowledge, indicating its effectiveness in cap-
turing a larger proportion of true positive instances.
While high recall rates are important for capturing as
many true positives as possible, the potential conse-
quences of false alarms must be carefully weighed,
especially in industrial settings where false positives
could lead to unnecessary maintenance interventions
and operational disruptions.
This hybrid approach not only improves the
model’s predictive accuracy but also ensures that crit-
ical, rare failure scenarios are effectively captured,
thereby enhancing the reliability and efficiency of the
failure prediction system. The system remains adap-
tive and robust in real-world operational contexts by
continuously refining the rules based on new data and
expert feedback.
5.3 Prediction and Explanation Time
The model’s prediction generation time was measured
to average 24.8 milliseconds per instance. This rapid
prediction time ensures the system can provide timely
alerts, allowing for rapid intervention and mitigation
measures. Similarly, the time required to generate ex-
planations using LIME averaged 233.8 milliseconds
per instance. While explanations took slightly longer
than predictions, this timeframe is considered accept-
able.
The relatively quick prediction and explanation
times contribute to enhanced operational efficiency.
Users can act on predictions and understand the un-
derlying reasons without significant delays, which is
crucial in time-sensitive environments.
5.4 Interface
Transitioning from discussing the technical perfor-
mance of the predictive model, it’s important to high-
light the user interface, which is the primary means
through which users interact with the system. Images
of the interface, showcasing its various features and
capabilities, are provided in Figure 4.
To test the interface, raw data was sent via
Message Queuing Telemetry Transport (MQTT), a
lightweight messaging protocol ideal for real-time
data transmission. The interface efficiently processes
this incoming data and displays the system’s current
status on a user-friendly dashboard. Key features of
the interface include:
• Real-Time Failure Detection. The dashboard
displays whether a failure is predicted based on
the incoming data, allowing for immediate user
intervention.
• Sensor Data Visualization. Alongside the fail-
ure prediction, the values from various sensors,
such as pressure, temperature, motor current, and
air intake valve readings, are shown in real-time.
This gives users a comprehensive view of the op-
erational status and helps diagnose potential is-
sues.
• Failure Report Management. The UI can dis-
play existing failure reports and allows users to
create new failure reports when a failure is iden-
tified. This helps track and document failure in-
stances systematically.
• Rule Management. Users can view the rules in
place, organised by order of priority, ensuring that
the most critical rules are given top priority. Ad-
ditionally, the interface allows users to create new
rules based on their insights and observations.
ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics
250

Table 1: Comparison of model performance before and after rules (Part 1).
Metrics
RF XGBoost LGBM
Before After Before After Before After
Recall (%) 90.685 90.999 91.991 91.991 88.786 90.748
F1 Score (%) 64.655 93.093 90.104 90.471 92.144 92.232
Precision (%) 50.236 95.285 88.293 89.001 95.767 95.855
Accuracy (%) 96.586 99.535 99.305 99.333 99.479 99.547
Table 2: Comparison of model performance before and after rules (Part 2).
Metrics
CatBoost GBM Voting
Before After Before After Before After
Recall (%) 91.413 91.413 91.003 92.241 91.003 91.413
F1 Score (%) 89.303 90.771 89.576 90.245 91.055 91.278
Precision (%) 87.288 90.137 88.193 88.333 91.107 91.143
Accuracy (%) 99.247 99.547 98.886 98.950 99.059 99.081
• Model Explanation. The UI includes functional-
ity to show detailed model explanations, provid-
ing insights into the factors influencing each pre-
diction. This transparency helps users understand
the model’s decision-making process.
The combination of real-time failure alerts, de-
tailed sensor data visualisation, failure report man-
agement, rule management, and model explanations
ensures that users can quickly understand and react to
potential problems.
6 CONCLUSIONS AND FUTURE
WORK
In modern manufacturing, equipment failures can
cause extensive downtime, resulting in considerable
financial losses and wasted resources. This paper in-
troduces a methodology for a failure prediction and
reporting system designed to mitigate such risks by
proactively reducing instances of equipment failure.
The UI enables real-time failure detection, sen-
sor data visualisation, failure report management,
rule management, and model explanations, facilitat-
ing rapid understanding and response to potential is-
sues. Additionally, the integration of domain knowl-
edge in the form of rules significantly enhanced the
system’s performance. Notably, GBM, with the ap-
plied rules, emerged as the most effective approach in
the evaluation. Overall, the HyPredictor methodology
ensures a comprehensive and adaptive failure predic-
tion and reporting system, leveraging advanced data
processing, machine learning, and domain knowledge
to enhance operational efficiency and facilitate proac-
tive maintenance strategies in industrial settings.
Future work should focus on ensuring the scala-
bility of the failure prediction and reporting system to
deploy across various industrial environments.
ACKNOWLEDGEMENTS
This work was partially supported by the HORIZON-
CL4-2021-TWIN-TRANSITION-01 openZDM
project, under Grant Agreement No. 101058673.
REFERENCES
Ahmed, I., Jeon, G., and Piccialli, F. (2022). From artificial
intelligence to explainable artificial intelligence in in-
dustry 4.0: a survey on what, how, and where. IEEE
Transactions on Industrial Informatics, 18(8):5031–
5042.
Angelopoulos, A., Michailidis, E. T., Nomikos, N.,
Trakadas, P., Hatziefremidis, A., Voliotis, S., and Za-
hariadis, T. (2020). Tackling faults in the industry 4.0
era—a survey of machine-learning solutions and key
aspects.
Barredo Arrieta, A., D
´
ıaz-Rodr
´
ıguez, N., Del Ser, J., Ben-
netot, A., Tabik, S., Barbado, A., Garcia, S., Gil-
Lopez, S., Molina, D., Benjamins, R., Chatila, R.,
and Herrera, F. (2020). Explainable artificial intelli-
gence (xai): Concepts, taxonomies, opportunities and
challenges toward responsible ai. Information Fusion,
58:82–115.
Bekar, E. T., Nyqvist, P., and Skoogh, A. (2020). An intelli-
gent approach for data pre-processing and analysis in
HyPredictor: Hybrid Failure Prognosis Approach Combining Data-Driven and Knowledge-Based Methods
251

Figure 4: Images of the interface.
predictive maintenance with an industrial case study.
Advances in Mechanical Engineering, 12.
Canizo, M., Onieva, E., Conde, A., Charramendieta, S.,
and Trujillo, S. (2017). Real-time predictive mainte-
nance for wind turbines using big data frameworks. In
2017 IEEE International Conference on Prognostics
and Health Management (ICPHM), pages 70–77.
Chi, Y., Dong, Y., Wang, Z. J., Yu, F. R., and Leung, V. C.
(2022). Knowledge-based fault diagnosis in industrial
internet of things: A survey. IEEE Internet of Things
Journal, 9:12886–12900.
Davari, N., Veloso, B., Ribeiro, R. P., Pereira, P. M., and
Gama, J. (2021). Predictive maintenance based on
anomaly detection using deep learning for air produc-
tion unit in the railway industry. In 2021 IEEE 8th
International Conference on Data Science and Ad-
vanced Analytics (DSAA), pages 1–10. IEEE.
Leukel, J., Gonz
´
alez, J., and Riekert, M. (2021). Adoption
of machine learning technology for failure prediction
in industrial maintenance: A systematic review.
Li, H., Parikh, D., He, Q., Qian, B., Li, Z., Fang, D., and
Hampapur, A. (2014). Improving rail network veloc-
ity: A machine learning approach to predictive main-
tenance. Transportation Research Part C: Emerging
Technologies, 45:17–26.
Sun, Q. and Ge, Z. (2021). A survey on deep learning for
data-driven soft sensors. IEEE Transactions on Indus-
trial Informatics, 17:5853–5866.
Zonta, T., da Costa, C. A., da Rosa Righi, R., de Lima,
M. J., da Trindade, E. S., and Li, G. P. (2020). Pre-
dictive maintenance in the industry 4.0: A systematic
literature review. Computers and Industrial Engineer-
ing, 150.
ICINCO 2024 - 21st International Conference on Informatics in Control, Automation and Robotics
252