Anomaly Detection in Industrial Machinery using IoT

Devices and Machine Learning: a Systematic Mapping

Sérgio F. Chevtchenko1,∗, Elisson da Silva Rochaa , Monalisa Cristina Moura
Dos Santosa , Ricardo Lins Motaa , Diego Moura Vieiraa , Ermeson Carneiro
de Andradeb , Danilo Ricardo Barbosa de Araújob
arXiv:2307.15807v1 [cs.LG] 28 Jul 2023

a
SENAI Institute of Innovation for Information and Communication Technologies
(ISI-TICs), Recife, Brazil
b
Department of Computing at the Federal Rural University of Pernambuco (UFRPE),
Recife, Brazil

Abstract
Anomaly detection is critical in the smart industry for preventing equip-
ment failure, reducing downtime, and improving safety. Internet of Things
(IoT) has enabled the collection of large volumes of data from industrial
machinery, providing a rich source of information for Anomaly Detection.
However, the volume and complexity of data generated by the Internet of
Things ecosystems make it difficult for humans to detect anomalies man-
ually. Machine learning (ML) algorithms can automate anomaly detection
in industrial machinery by analyzing generated data. Besides, each tech-
nique has specific strengths and weaknesses based on the data nature and its
corresponding systems. However, the current systematic mapping studies on
Anomaly Detection primarily focus on addressing network and cybersecurity-
related problems, with limited attention given to the industrial sector. Ad-
ditionally, these studies do not cover the challenges involved in using ML for
Anomaly Detection in industrial machinery within the context of the IoT

∗
Corresponding author
Email addresses: [email protected] (Sérgio F.
Chevtchenko), [email protected] (Elisson da Silva Rocha),
[email protected] (Monalisa Cristina Moura Dos Santos),
[email protected] (Ricardo Lins Mota),
[email protected] (Diego Moura Vieira),
[email protected] (Ermeson Carneiro de Andrade),
[email protected] (Danilo Ricardo Barbosa de Araújo)

Preprint submitted to TBD August 1, 2023

ecosystems. This paper presents a systematic mapping study on Anomaly
Detection for industrial machinery using IoT devices and ML algorithms to
address this gap. The study comprehensively evaluates 84 relevant stud-
ies spanning from 2016 to 2023, providing an extensive review of Anomaly
Detection research. Our findings identify the most commonly used algo-
rithms, preprocessing techniques, and sensor types. Additionally, this review
identifies application areas and points to future challenges and research op-
portunities.
Keywords: Anomaly detection, IoT ecosystems, Machine learning,
Mapping study

1. Introduction
In recent years, many tasks previously performed by humans have been
automated through smart technologies. In the Smart Industry or Indus-
try 4.0, for example, these technologies have made it possible to monitor
production lines and detect potential problems before they become serious,
leading to fewer delays and increasing productivity [1, 2]. These industries
are constantly present in our daily lives and disruptions can negatively af-
fect productivity and cause economical losses. Generally, these disruptions
come from various types of unexpected and non-standard behaviors that
can occur in these environments, known as anomalies [3]. In industrial ma-
chinery, anomalies can occur due to equipment malfunctions, environmental
conditions, and changes in operating conditions. Therefore, detecting these
anomalies as early as possible is essential to prevent equipment failure, reduce
downtime, and minimize repair costs [3].
IoT ecosystems have enabled the collection of large volumes of data from
industrial machinery, providing a rich source of information for Anomaly De-
tection (AD) [4]. These ecosystems are composed of sensors and monitoring
systems that can collect data such as temperature, pressure, vibration, and
power consumption. This data enables engineers and specialists to monitor
the health of industrial machinery in real-time and take immediate action to
diagnose and address the underlying problem before it causes a significant
disruption to the production process. However, the volume and complexity
of data generated by IoT ecosystems make it difficult to detect anomalies
manually, increasing the number of corrective and preventive maintenance.
ML algorithms can help automate the process of AD by analyzing the data

2
generated by these environments [5]. Despite the fact that there are several
ML techniques used for AD, each technique has its strengths and weaknesses
and can be used depending on the nature of the data and the specifics of the
industrial contexts.
Most of the current systematic mapping study on AD is in Network
Anomaly Detection (NAD), considering the cybersecurity context [6, 7, 8].
They are widely used for fraud detection, network security breaches, and
environmental monitoring. Very few of them have focused on AD in the in-
dustry [9, 10]. In [9], the authors applied a systematic literature review to
identify frameworks, architectures, and tools in the area of predictive main-
tenance. Similarly, Zonta et al. [10] present a survey, discussing the current
challenges and limitations in predictive maintenance and proposing a new
taxonomy to classify this research area considering Industry 4.0. However,
to the best of our knowledge, no related work presents and discusses the
challenges in using ML for the detection of anomalies in industrial machines
using IoT.
Considering these issues and challenges described above, and the grow-
ing acknowledgment of the necessity for AD solutions, a systematic mapping
study on the AD topic is highly necessary. This paper presents a systematic
mapping of AD for industrial machinery using IoT devices and ML algo-
rithms conducted between October 2022 and January 2023. In this study,
84 papers dating from 2016 to 2023 are evaluated. This systematic mapping
aims to provide a comprehensive review of AD research, including an anal-
ysis of current methodologies, a synthesis of evidence, an identification of
applications, a discussion of research issues, and an identification of future
challenges.
This mapping study provides insights into key aspects of AD research in
the context of industrial machinery, addressing the following questions: (i)
What type of machinery is most commonly monitored and why? (ii) What
are the types of sensors and variables employed for detecting anomalies?
(iii) What are the types of machine learning techniques used for anomaly
detection in industrial machinery? (iv) How is the anomaly detection method
computed and evaluated? In this way, this work can help to direct future
efforts regarding AD solutions, as well as promote a discussion about AD
studies, their implications, and challenges for the future.
The remainder of the paper is organized as follows. Section 2 presents
existing reviews regarding the AD field. Section 3 introduces the protocol
used to conduct the research. Section 4 presents the findings for the estab-

3
lished research questions. Section 5 presents the open challenges. Section 6
presents the limitation of this mapping study. Finally, Section 7 presents the
conclusions, the limitations, and the future work for this study.

2. Related Work
In recent years, there has been considerable interest in AD for industrial
machinery using IoT devices and ML. However, the majority of studies have
primarily focused on network anomaly detection. To position our paper and
emphasize its contributions, we start by presenting a concise introduction
to mapping studies that concentrate on network anomaly detection. Subse-
quently, we provide a summary of previous research on surveys conducted on
AD and predictive maintenance in the industry, employing ML techniques.
Finally, we present a comparative analysis of our work in relation to these
studies, considering factors such as scope and evaluation methods.
There is a significant body of literature comprising systematic mapping
studies that specifically concentrate on NAD [11, 12].Ahmed et al. [6] explore
various NAD techniques and elaborate on the use of different categories of
detection methods as solutions for this problem. The authors also discuss
the limitations of using publicly available intrusion detection datasets and
provide a comparison of the effectiveness of presented AD techniques based
on specific criteria. In more recent work, Eltanbouly et al. [7] examine ML
algorithms applied to NAD and analyze the performance of the surveyed
papers. The study provides an overview of their main positive and negative
characteristics, along with numerical analyses of algorithms learned on the
same datasets. Wang et al. [8] discuss different types of ML approaches in
the context of AD, comparing their merits within the scope of NAD. The
authors also address the key challenges faced in detecting various anomalies
in networks with varying complexities. Additionally, they present use cases
for different types of networks, considering IoT networks, though with limited
scope.
Very few mapping studies have focused on AD and predictive maintenance
in the industry. Gaddam et al. [13] investigate AD techniques to identify
sensor faults and outliers in the IoT context, without specific emphasis on
industrial equipment. Their survey comprehensively describes the primary
sources of sensor outliers within the IoT environment and discusses detection
models that are suitable for IoT systems. The authors further provide a com-
parative analysis of these techniques, highlighting their respective strengths

4
and weaknesses. With a focus on Industry 4.0, Kamat and Sugandhi [14]
provide a discussion about the main challenges associated with traditional
equipment maintenance strategies in the manufacturing industry. As a way
to address this problem, the study suggests the adoption of AD in predic-
tive maintenance as a more suitable approach for handling the specific data
characteristics present in this environment. Additionally, the paper briefly
explores alternative approaches to AD found in the literature and discusses
publicly available datasets relevant to the manufacturing scenario.
Dalzochio et al. [9] present important considerations regarding the ap-
plication of predictive maintenance, such as the abundance of data, the crit-
icality of the equipment, the need for redundancy and the availability of
failure-related data for modeling purposes. The researchers address several
key research questions, including the challenges associated with applying ML
to predictive maintenance, commonly employed ML techniques in this con-
text, and the utilization of ontologies in predictive maintenance scenarios. In
terms of ML techniques, the analyzed works were grouped into three main
categories: (i) based on artificial neural networks (ANN); (ii) based on deep
learning (DL); and (iii) based on other ML approaches, such as k-nearest
neighbors (kNN), support vector machines (SVM), and random forest (RF).
The authors suggest that different ML models are better suited for AD/fault
classification and prognostics. While classification or clustering ML models
can address AD and fault classification, regression models such as autoregres-
sive integrated moving averages (ARIMA) are better suited for prognostics
tasks. Furthermore, the paper provides a concise exploration of the use of
ontologies in predictive maintenance, aiming to facilitate decision-making
processes.
Zonta et al. [10] present a systematic literature review that explores ini-
tiatives related to predictive maintenance in the context of Industry 4.0. The
authors analyze 47 articles and, as the main contribution, propose a taxon-
omy for predictive maintenance in the context of Industry 4.0. Schwende-
mann et al. [15] conduct a survey that specifically explores ML techniques
for predictive maintenance of bearings in grinding machines. The selected
works encompass a range of approaches, including ANNs, Hidden Markov
Models (HMMs), and SVMs.
Kang et al. [16] conduct a systematic literature review that specifically
examines the applications of ML in production lines and their components.
Their analysis included 39 primary studies that predominantly concentrated
on quality control within the production lines. The most frequently utilized

5
dependent variables in the ML methods were identified as Fail/Pass indica-
tors, physical properties of materials, and object dimensions.
With a focus on the railway industry, Davari et al. [17] present a survey
on data-driven predictive maintenance. Within this context, they emphasize
real-time AD from time-series data and the need for automatic reasoning
capabilities to explain causality as major challenges faced by the industry.
Nor et al. [18] present an analysis with a specific focus on explainable artificial
intelligence (XAI) applied to prognostics and health management (PHM)
of industrial assets. The authors adopt the preferred reporting items for
systematic reviews and meta-analyses (PRISMA) methodology in their study.
They analyzed a total of 35 selected works, with 13% of them incorporating
an AD approach. It is worth highlighting that the analyzed studies primarily
feature case studies based on real industrial data. This demonstrates the
practical application of AI models in industrial settings and contributes to
an increasing level of confidence in adopting such models within the industry.
The present work distinguishes itself from the related literature above for
the following reasons. Firstly, AD-based predictive maintenance is recognized
as a promising area in several of the related literature reviews with broader
scopes. Thus, the present work offers a state-of-the-art review with a focused
approach on the use of ML techniques for detecting anomalies in industrial
machinery. Secondly, unlike previous studies, we do not limit our analysis
to specific industrial machinery, aiming to provide a more comprehensive
perspective on existing industrial practices. As a result, our work seeks to
offer an up-to-date review of the state-of-the-art, emphasizing the application
of machine learning methods for AD in the industry. Finally, the present work
incorporates the most recent findings from the literature, ensuring that our
review is current and reflective of the latest advancements in the field.

3. Systematic Mapping Study Process

Our Systematic Mapping (SM) was guided by the methodology proposed
by Petersen et al. [19], which we employed to identify articles related to the
use of ML techniques for detecting anomalies in industrial machinery with
the aid of IoT devices, as illustrated in Figure 1.
Step 1. Given the constantly growing volume of data generated by in-
dustrial machinery, the use of ML models has become increasingly significant
in detecting anomalies and enabling preventive maintenance, as opposed to
corrective maintenance. The primary objective of this SM is to investigate

6
 Figure 1: Steps of the adopted protocol.

the application of ML models for predicting anomalies in industrial machin-

ery in the context of IoT ecosystems. Specifically, this study aims to identify
the most commonly adopted machinery, explore the types of data and sen-
sors utilized for AD, identify the computation and evaluation methods used
for AD, and analyze the challenges and opportunities associated with imple-
menting ML-based AD systems in industrial settings. The insights gained
from this research will be critical in assessing the current state of this field
and providing guidance for future research endeavors in this domain.
Step 2. Our SM research began by identifying key Research Questions
(RQs) aimed at exploring the current state of AD in industrial machinery
using ML algorithms and sensors. The following RQs were adopted for this
SM:

• RQ1: What type of machinery is most commonly monitored and why?

• RQ2: What are the types of sensors and variables employed for detect-
ing anomalies?

• RQ3: What are the types of machine learning techniques used for

7
 anomaly detection in industrial machinery?
• RQ4: How is the anomaly detection method computed and evaluated?
Step 3. To compile primary research, we employed automatic database
searches utilizing a search string. The following procedures were undertaken
to elaborate the search phrase for the automatic search [20, 21]: (i) extract
phrases from the study questions; (ii) find alternate spellings and synonyms;
(iii) validate the keywords; and IV. combine search strings using Boolean
operators (OR, AND). The search string was (("Anomaly Detection" OR
"Anomalous Behavior") AND ("IoT" OR "Internet of Things" OR "Sen-
sors") AND ("Machine Learning" OR "Deep Learning" OR "Artificial Intel-
ligence")).
Step 4. The following digital libraries were considered as the primary
sources for our research: IEEE Xplore1 , ACM Digital Library2 , Science Di-
rect3 and Web of Science4 .
Step 5. As numerous papers unrelated to our research topics may be
found, we have established specific inclusion and exclusion criteria. These
criteria are intended to narrow down our search and ensure that the identified
literature is relevant to our assessment research. For the inclusion criteria, we
focused on papers published within the last ten years that explicitly address
the AD using ML and IoT sensors in their abstracts. We excluded, on the
other hand, duplicate, unavailable, or foreign-language articles, as well as
editorials, posters, tutorials, and secondary or tertiary articles.
Step 6. After applying the inclusion and exclusion criteria and utilizing
the search string in the digital libraries, we successfully located the primary
studies.
Step 7. We extracted pertinent information from the primary studies
by thoroughly reading the entire paper and answering the RQs.
Step 8. To classify and organize the articles in accordance with our
research questions stated in Step 2, an overview of all articles is provided in
this stage (see Subsection 4.1).
Step 9. Finally, in Section 4, we present the responses we found to
the RQs posed in Step 2, which details the current state of the literature

1
IEEExplore.ieee.org/Xplore/home.jsp
2
https://dl.acm.org
3
https://www.sciencedirect.com/
4
https://www.webofscience.com

8
regarding the utilization of ML and sensors to identify anomalies in industrial
machinery.

4. Results and Discussion

To provide a comprehensive understanding of AD techniques in industrial
machinery, we begin with an overview of the primary studies conducted in
this field. We then present specific issues related to AD, such as the types of
machinery commonly monitored, the sensors and variables employed for de-
tecting anomalies, the types of ML techniques used for AD, and how the AD
methods are computed and evaluated. This approach provides a solid under-
standing of the techniques and challenges involved in detecting anomalies in
industrial machinery.

4.1. Overview of the Primary Studies

The number of studies identified by the search string in the search sources,
as well as the number of studies following the application of the inclusion
and exclusion criteria, are presented in Table 1. Initially, 8966 articles were
retrieved from the 4 databases. After removing duplicates, the number de-
creased to 8037 articles. Through the application of the inclusion and exclu-
sion criteria, 84 articles were selected for thorough reading and to address the
research questions. This represents only 0.936% of all articles found during
the search, emphasizing the importance of applying rigorous selection criteria
to identify relevant studies.

Table 1: Search results obtained before and after applying the inclusion and exclusion
criteria.

Database Original Search After Primary Studies Identification

ACM DL 1713 6
IEEE Xplore 979 11
Web of Science 1249 27
Science Direct 5025 40

Among the databases searched Science Direct yielded the highest number
of selected works, with 40 articles. This was followed by Web of Science with
27, IEEE Xplore with 11, and finally ACM DL with only 6 articles. In terms
of the proportion between the number of selected articles and the original

9
search, Web of Science achieved the best result with approximately 2.16%
(27/1249) of the selected base, followed by IEEE Xplore with 1.12% (11/979).
Based on the selected articles, Figure 2 presents the number of primary
studies in relation to the year of publication. Although our search aimed to
include articles from 2012 onwards, the first selected study was from 2016.
Notably, no articles were selected for 2017 and 2018. However, we observed
a growing trend in the literature regarding the detection of anomalies in
industrial machinery using sensors and ML from 2019 onwards. In 2019, 9
studies were selected, increasing to 21 in 2020, 28 in 2021, and reaching 39
in 2022. It is worth mentioning that even though our search was conducted
up until November 2022, some studies have already been published in 2023,
indicating that research in this area is ongoing.

Figure 2: Number of primary studies by publication year.

Based on the type of publication, journals represent the majority, com-

prising 80% of primary studies, while conferences account for the remaining
20%. We can also examine the trends in publication venues over the years
(see Figure 3). The first publication, in 2016, was presented at a conference.
By 2019, the proportion of journal publications had increased significantly,
with 55% of the publications (5 papers) being in journals, and 45% (4 papers)
in conferences. In the subsequent years, the number of journal publications
continued to rise, with 12 papers in 2020, 18 in 2021, and 28 in 2022. This
trend could be attributed to multiple factors, including the pandemic, which
prompted researchers to shift their focus to publishing in journals rather than
presenting at conferences. Additionally, the natural progression and devel-

10
opment of research studies over time may have contributed to the increase
in journal publications.

Figure 3: Number of primary studies by year and publishing type.

Several journals have published a significant number of works on the topic

of this MS, with SENSORS being the most productive with 9 publications,
followed by Procedia CIRP with 5 publications. Expert Systems with Ap-
plications, IEEE Access, Journal of Manufacturing Systems, and Mechanical
Systems and Signal Processing also had a substantial output, each with 4
publications. All of these journals have a notably high impact factor. It is
worth mentioning that none of the publications were presented at the same
conference, highlighting the diversity of the research presented and the scope
of the academic community.

4.2. What type of machinery is most commonly monitored and why?

To answer the above research question, we listed the articles that clearly
indicate the type of machinery under evaluation. From this selection, we
created 11 groups based on categorical proximity and their use in the articles
(e.g., articles that mentioned pumps also mentioned valves and hydraulic
systems. In these cases, we considered hydraulic systems). The number of
works gathered in each group is shown in Figure 4. The machinery that did
not fit into any existing group was not used more than once, were placed
in a group called “others”. For this analysis, studies in which the monitored
machinery could not be identified were excluded. Based on the number of

11
occurrences in each group, we ranked the groups to observe the types of
machinery most monitored in the primary studies.

Figure 4: Number of primary studies by the type of industrial machinery.

As shown in Figure 4, the three primary types of machinery used for

anomaly detection are milling and cutting tools, hydraulic systems, and bear-
ings. Several studies conducted across these equipment groups have imple-
mented predictive maintenance techniques with the goal of achieving various
benefits. These benefits include reducing the amount of lost production,
improving the quality of products, minimizing the costs associated with un-
expected equipment downtime, and increasing overall production efficiency.
Furthermore, the implementation of predictive maintenance can also lead to
an improvement in the lifespan of the machinery. [22, 23, 24, 25]
In the category of milling and cutting tools, numerous studies concen-
trate on assessing the effects of wear and tear on their surfaces. These
studies are particularly relevant for machinery such as Computer Numeri-
cal Control (CNC) routers, milling machines, lathes, and hot rolling mills,
which frequently handle heavy-duty tasks involving cutting metal, ceram-
ics, and other hard materials. The nature of these tasks makes these tools
susceptible to significant wear, which is a primary reason for continuous mon-
itoring [22, 23, 24, 25, 26, 27, 28, 29]. As these milling and cutting tools wear

12
out due to intensive usage, they can directly impact the quality of manufac-
tured products and even cause damage to the machinery itself. In some cases,
routine maintenance involves replacing these worn-out parts, even when it
might still be possible to continue using them [22].
Hydraulic systems consist of several components, such as accumulators,
coolers, valves, pumps, compressors, and more. Depending on the applica-
tion, these systems can operate either as standalone units or as part of larger
systems. For example, in aviation, hydraulic systems play a crucial role in
controlling mechanical parts of aircraft, while in other applications, they may
be integrated into cooling systems [30]. It is important to note that some hy-
draulic systems operate under high pressure and deal with toxic or flammable
fluids, emphasizing the criticality of proper maintenance for safety. Tradi-
tionally, manual monitoring of these hydraulic systems has been employed,
relying on the extensive knowledge and judgment of operators. However,
this approach can be subjective and may lead to inefficiencies. Therefore,
researchers advocate for the adoption of automatic monitoring techniques
using IA models to detect anomalies in hydraulic systems [31, 32, 33, 34].
Bearings are the third most monitored component in industrial machinery.
Due to their critical role in supporting rotating parts and reducing friction,
monitoring the condition of bearings is of utmost importance to ensure the
smooth operation and longevity of various machines and equipment [35, 36].
For instance, a recurrent issue arises when wind turbines operate at low
speeds, subjecting the bearings to significant stress due to the considerable
weight of the turbine’s components [37]. Predicting abnormal behavior in
bearing wear is essential in such scenarios. Furthermore, in this mapping
study, several other machinery were identified, including electric motors and
industrial robots. However, milling and cutting tools, hydraulic systems, and
bearings were found to be prevalent.
Given their extensive use in almost all industrial environments, it is im-
perative for these machines to operate smoothly and consistently even un-
der demanding conditions for extended periods. Consequently, predicting
anomalous behavior in industrial machinery is essential to avoid negative
impacts on productivity, as it allows for the early detection of potential
faults and the implementation of preventive maintenance measures. It is
worth highlighting that many of the analyzed studies do not focus on evalu-
ating specific machinery but rather on exploring new ML and DL approaches
using datasets containing data related to various machinery types. Nonethe-
less, they contribute to the overall understanding of AD techniques and their

13
applicability in the context of industrial machinery.

4.3. What are the types of sensors and variables employed for detecting anoma-
lies?
The majority of the primary studies focused on AD modeling and, as a
result, lacked detailed technical information on the types of sensors used for
machinery monitoring. This focus on AD over the whole system monitor-
ing pipeline, which includes data acquisition, processing, and information
delivery, limits the depth of our analysis. Moreover, many studies vali-
dated the proposed methodologies using well-established datasets, such as
C-MAPSS5 [38] and the NASA Ames Prognostic Data Repository6 [24, 39,
40, 22]. While these datasets serve as benchmarks for experimental purposes,
it is important to test these methodologies in real-world industrial settings
for practical applications. Despite these limitations, the majority of the stud-
ies did reveal the variables used for detecting AD. We have categorized these
variables based on the type of sensor used, such as vibration, electric cur-
rent, temperature, noise, pressure, among others. Table 2 lists the studies
that employed each sensor type, either individually or in combination with
each other.

Table 2: Sensors used in selected works.

Electric
Article Vibration Temperature Noise Pressure Others
current

[41, 42, 43, 44, 45, 46, 47]

[35, 48, 25, 39, 49, 50, 51] ✓

[52, 53, 54, 34, 31, 32, 33, 55] ✓

[56, 57, 29, 58, 59] ✓

[60, 61, 62, 63, 30] ✓ ✓ ✓ ✓ ✓

[64, 65, 66, 67] ✓ ✓

[68, 69, 70] ✓ ✓

Continued on next page

5
https://data.nasa.gov/dataset/C-MAPSS-Aircraft-Engine-Simulator-Data/
xaut-bemq
6
https://www.nasa.gov/content/prognostics-center-of-excellence-data-set-repository

14
 Table 2: Sensors used in selected works. (Continued)
Electric
Article Vibration Temperature Noise Pressure Others
current

[71, 72, 73] ✓ ✓ ✓ ✓

[74, 75, 23] ✓ ✓ ✓

[76, 77] ✓

[27, 40] ✓ ✓ ✓ ✓

[78, 26] ✓ ✓

[28, 22] ✓ ✓ ✓

[79, 80] ✓ ✓ ✓ ✓

[81, 82] ✓

[83, 84] ✓ ✓

[85, 86] ✓ ✓ ✓ ✓

[36] ✓ ✓ ✓

[37] ✓ ✓ ✓

[87] ✓ ✓ ✓ ✓ ✓

[88] ✓ ✓

[89] ✓ ✓

[90] ✓ ✓

[91] ✓ ✓

[92] ✓ ✓ ✓ ✓

[93] ✓ ✓

[94] ✓ ✓ ✓ ✓

[95] ✓ ✓ ✓ ✓

[24] ✓ ✓ ✓

[96, 97, 98, 99, 100, 101, 102] ✓

Out of the 84 examined papers, 24 relied on a single type of sensor for

detecting AD. Ten of these studies exclusively used vibration sensors [41, 42,
43, 44, 45, 46, 47, 35, 48, 25, 39, 49, 50, 51]. Seven studies depended solely
on noise sensors [52, 53, 54, 34, 31, 32, 33, 55]. Five papers utilized only

15
electric current sensors [60, 61, 62, 63, 30], while two used only temperature
sensors [76, 77]. Finally, two papers solely relied on pressure sensors [81, 82].
This suggests that roughly 28.57% of the selected studies focus on a single
type of sensor.
While there are certain cases where a single sensor may be sufficient, there
are potential benefits in using multiple sensor types in the field of AD [103].
As an example, multiple sensors can be used to infer new information about
the physical system under observation. Through the combined analysis of
data from these varied sensors, one can extrapolate novel information not
directly observed. This technique is referred to as the “virtual sensors” in
the literature [104]. Nevertheless, many studies employed a combination of
sensors, with five works opting to use the vibration, electric current, tem-
perature and pressure sensors, along with an additional unclassified sensor
type [60, 61, 62, 63, 30]. In three studies vibration, electric current and
temperature sensors are combined [74, 75, 23].
As depicted in Figure 5, the distribution of sensor types across the pri-
mary studies reveals a clear preference for certain sensors. The most fre-
quently employed sensor was the vibration sensor, utilized in 40 studies.
Temperature sensors were the second most common, appearing in 30 stud-
ies, closely followed by electric current sensors, which were used in 29 studies.
Pressure sensors found application in 20 studies, while noise sensors were in-
corporated in 16 studies. A variety of other sensor types, accounting for
variables such as speed, torque, RPM, humidity, viscosity, and proximity,
were collectively used in 38 studies.
Despite noise sensors being the exclusive focus in seven studies as pre-
sented earlier, they were, in fact, the least frequently used when we consider
the overall usage of sensor types across all studies, with just 16 occurrences.
This suggests that a mix of sensor types is proffered to relying solely on noise
sensors or other single sensor type. It is worth mentioning that cameras,
while traditionally used for tasks like segmentation, edge detection, object
classification, recognition, and tracking, also serve as a viable sensor type
for AD. Yet, none of the primary studies in our review employed cameras
for this purpose. Nevertheless, the integration of general-purpose or special-
ized cameras in AD systems could provide valuable insights and broaden the
scope of detectable anomalies. For instance, the use of event-based cameras
for vibration analysis is a promising application, as they provide superior
spatio-temporal resolution in comparison to traditional sensors [105].

16
 Figure 5: Number of primary studies by sensors type.

4.4. What are the types of Machine Learning techniques used for anomaly
detection in industrial machinery?
There are two critical stages in the development of an effective AD tech-
nique for industrial machinery: data preprocessing and the application of AI
algorithms. Both stages are equally important and play fundamental roles in
the ability to accurately and timely detect anomalies. In the following sub-
sections, we provide a detailed description of the preprocessing techniques
and algorithms used in the selected primary studies for AD in industrial
machinery.

4.4.1. Preprocessing techniques

Preprocessing techniques play a crucial role in the field of ML, helping
to clean and prepare data before it is fed into algorithms for analysis. The
process of preprocessing involves a range of techniques such as data clean-
ing, normalization, feature selection, and dimensionality reduction, among
others [106]. These techniques are critical for improving the quality of data,
reducing computational complexity, and enhancing the performance of ML
models. According to the results obtained in the analysis of the selected pri-
mary studies, it was found that approximately 67% of the studies presented

17
some preprocessing technique applied to the data used in the research. The
techniques involved in preprocessing could range from feature selection to
dimensionality reduction, and data transformations aimed at enhancing the
quality of the input data. This reinforces the critical role of preprocessing
in AD for industrial machinery, as the use of appropriate techniques can
significantly improve the efficiency and accuracy of the detection process.
Figure 6 illustrates the most frequently employed preprocessing tech-
niques in the selected primary studies. The most commonly adopted data
preprocessing technique was the Autoencoder, which is a powerful technique
for dimensionality reduction [107]. Autoencoders use neural networks to
learn a compressed representation of the input data. This technique can be
especially useful for nonlinear and high-dimensional data, where traditional
methods such as Principal Component Analysis (PCA) may not be as effec-
tive [108]. Other techniques were categorized together because they served
a similar purpose. For instance, statistical methods such as mean, median,
standard deviation, and variance were used to measure the central tendency
and variability of the data and were grouped together. This group is referred
as “Statistical” and is ranked as the second most commonly utilized tech-
nique for preprocessing. These statistical methods enable researchers to gain
insights into the distribution of data, identify outliers, and detect trends or
patterns [109].
Fast Fourier Transform (FFT), another frequently used technique, and
Short-Time Fourier Transform (STFT) both operate in the frequency do-
main, enabling the time-based analysis of signals. However, FFT is bet-
ter suited for stationary signals, while STFT is more appropriate for non-
stationary signals [110]. Among the identified techniques, FFT was the most
widely used, appearing in 13 studies, while STFT was used in 4 studies. An-
other highly adopted technique for data preprocessing was PCA (Principal
Component Analysis). It identifies the most important variables in a dataset
and reduces the number of dimensions while preserving the most significant
information. By transforming the data into a new coordinate system that
maximizes the variance of the data, PCA helps to eliminate redundant or
noisy features that may hinder the analysis [111]. Normalization is another
widely adopted preprocessing technique that is commonly found in primary
studies. Its importance lies in transforming data into a standardized scale,
allowing for the easy comparison and analysis of data from diverse sources.
Additionally, it helps mitigate any inherent bias in the data that may arise
due to differences in units of measurement or data collection procedures [112].

18
 Figure 6: Number of primary studies by preprocessing techniques.

From the selected primary studies, it was found that PCA was utilized in 10
studies, and normalization was employed in an equal number of studies.
Another preprocessing technique that was moderately used in the primary
studies was Convolutional Neural Network (CNN). It is a deep learning (DL)
architecture that has achieved remarkable success in image and signal pro-
cessing tasks, thanks to its ability to learn hierarchical representations of the
input data. Besides its primary application in image and signal processing,
CNN can also be used for feature extraction in various applications [113].
The popularity of these techniques highlights their importance in data
preprocessing and analysis, as they enable researchers to reduce the dimen-
sionality of complex data and extract relevant features that can improve the
accuracy and efficiency of subsequent analysis.
While less commonly used, several other preprocessing techniques were
also found in the selected studies. These include bandpass and low-pass
filters [65, 47, 49, 29, 56], wavelet transformations [24, 53, 49], sliding win-
dow techniques [99, 114], and others. However, it is worth mentioning that
the selection of the most appropriate preprocessing techniques depends on

19
the nature and specificity of the data being analyzed, and researchers must
consider the advantages and limitations of each technique before making a
decision.
Additionally, some studies combined preprocessing techniques to try to
improve AD. For instance, a notable example is the utilization of autoen-
coders in conjunction with other preprocessing methods. Among the studies
that employed autoencoders, five studies used normalization or statistics in
conjunction with autoencoder, while four studies combined FFT or PCA with
autoencoder, as shown in Figure 7. These results suggest that using multi-
ple preprocessing techniques together can be an effective way to enhance AD
accuracy.

Figure 7: Number of primary studies by preprocessing techniques filtered by autoencoder.

4.4.2. Algorithms
Various ML-based algorithms have been employed to detect anomalies in
industrial machinery. Figure 8 presents these algorithms grouped into differ-
ent levels of categorization. Vertical categories indicate further subdivisions,
while horizontal categories represent the final categorization. The initial
grouping divides the algorithms into three categories: Supervised, Unsuper-
vised, and Heuristics. Supervised models were divided into two categories:
Classification and Regression. Classification models were subdivided into
several types: tree-based models, ensemble models, SVM, neural networks,
distance-based models, time-series models, and others. Regression models,
in turn, were the final categorization. The unsupervised models, on the other
hand, were divided into three final categories: outlier detection, clustering,
and density estimation. Another separate final grouping was the Heuristic.

20
Figure 8: Algorithms categorization.

21
 Figure 9 shows the number of primary studies that used each final al-
gorithm categorization. There are 12 final categorizations in total. This
includes seven groups for classification, one for regression, three for unsuper-
vised learning, and one for heuristic. It is important to note that a single
study might use more than one technique. The results of the studies demon-
strate the effectiveness of various algorithms and approaches for identifying
anomalies in industrial machinery. Heuristic-based approaches were the most
common, used in 28 studies. Neural network methods were also widely used,
found in 24 primary studies. Outlier detection was another common method
for finding anomalies in sensor data, used in 16 primary studies. On the other
hand, time-series-based and density estimate approaches were less frequently
used, with only 2 [76, 83] and 1 [61] primary studies, respectively.

Figure 9: Number of primary studies based on algorithms categorization.

Based on the categorizations that were most prominent in Figure 9, we

are going to analyze which techniques are frequently used in one of these
categories. We start with the methods related to neural networks. After that,
we address the techniques used for outlier detection. Finally, we analyze the
techniques used in the most common categorization, which is Heuristic.

22
 Figure 10 illustrates the most frequently employed neural network tech-
niques in the primary studies related to ML-based AD. Multilayer percep-
tron (MLP) was the most widely used technique, with 11 primary studies,
which involves multiple layers of artificial neurons used to recognize patterns
in data. Long short-term memory (LSTM) was the second most popular
technique with five primary studies [99, 98, 64, 78, 82], which is a type of re-
current neural network (RNN) architecture designed to remember past data
over time. CNN was used in three primary studies [53, 52, 54], which is a
type of feed-forward neural network that was also employed in the prepro-
cessing step, as mentioned previously. One-class neural network (OCNN)
was employed in three primary studies [43, 39, 32], which is a neural network
architecture specifically designed for AD problems. Finally, Generative Ad-
versarial Networks (GANs) were employed in two primary studies [34, 115],
which is a type of neural network that involves two models trained simulta-
neously, where one model generates samples, and the other model evaluates
their authenticity [116]. The less frequent implementation of GANs, rela-
tive to other DL architectures, may not be a reflection of its applicability
for anomaly detection, but rather of its relative novelty in the field, which
suggest that this approach could be explored further.

Figure 10: Number of primary studies by neural network category.

23
 The outlier detection category has also received significant attention in
the literature. Figure 11 shows the specific techniques employed in this cate-
gory. The most commonly used techniques among the primary studies were
One-class support vector machine (OCSVM) with 10 studies, followed by Iso-
lation Forest (Iforest) and Local Outlier Factor (LOF), with 6 studies each.
OCSVM is a well-known method for outlier detection that separates inliers
from outliers by identifying a boundary around the inliers [117]. Iforest is
a tree-based algorithm that isolates outliers by constructing separation trees
[118, 119]. LOF is a density-based method that measures the local devi-
ation of a data point with respect to its neighbors [119]. Although other
techniques such as Histogram-based Outlier Score (HBOS), Dynamic Time
Warping (DTW), and Angle-Based Outlier Detection (ADBOD) were also
adopted, their usage was relatively limited in comparison to the previously
mentioned approaches.

Figure 11: Number of primary studies by outlier detection category.

In the primary studies that utilized heuristics to detect anomalies, the

most prevalent approach involved setting a threshold to identify values that
significantly deviated from normal behavior. This approach, known as rule-
based AD, is widely used in industrial applications. However, using a fixed
value or percentage to set the threshold can lead to false positives or false neg-
atives. To enhance detection accuracy, this technique is typically combined
with preprocessing steps.
Figure 12 illustrates the most commonly used preprocessing techniques
in conjunction with this approach. Among the various preprocessing meth-

24
ods, the autoencoder technique was the most commonly utilized, with 16 in-
stances. Several studies have demonstrated the integration of this technique
with threshold-based rule-based AD, where the threshold is established based
on the autoencoder reconstruction error. Other techniques such as FFT, nor-
malization, and statistics were also widely used to assist in diagnosis using
thresholds. These findings demonstrate that the heuristics-based approach
for AD is promising, but requires careful consideration in selecting an appro-
priate preprocessing technique and defining the threshold for each specific
application.

Figure 12: Number of primary studies categorized by preprocessing techniques and filtered
by the heuristic approach.

4.5. How is the anomaly detection method computed and evaluated?

The choice of an appropriate evaluation method is crucial when aiming
to find the optimal classifier for a given problem, whether it involves classi-
fication, regression, or clustering [120]. Evaluating AD models can be done
using various metrics that align with specific tasks. In our primary studies,
we identified several metrics, as shown in Figure 13. The metric Accuracy was
the most commonly used in the analyzed studies, appearing in 36 instances.
It was closely followed by F1-score, Precision Recall, which appeared in 20,
19, and 18 studies, respectively. These metrics are derived from the confusion
matrix, a tabular representation that summarizes the model’s performance
by indicating correct and incorrect classifications in terms of true positives,
true negatives, false positives, and false negatives.
The Area Under the Receiver Operating Characteristic Curve (AUC-
ROC) is a widely utilized metric in binary classification models. It represents
the relationship between the true positive rate (TPR) and the false positive
rate (FPR) at different threshold levels. Our analysis identified this specific

25
Figure 13: Number of primary studies by metrics.

26
metric in 15 primary studies. Additionally, the TPR and FPR metrics, which
are foundational for calculating the ROC curve, were found in 10 and 8 stud-
ies, respectively. Consequently, it can be deduced that binary classification
is the most frequently performed and evaluated task within the context of
AD.
Other metrics employed in binary classification include Matthews Correla-
tion Coefficient (MCC), frequently used metric for unbalanced classes, which
combines true positive, true negative, false positive, and false negative rates
to evaluate overall model quality, used in 3 studies [93, 85, 115]; False alarm
rate, also known as FPR, which quantifies the proportion of negatives mis-
takenly classified as positives, also used in 3 studies [102, 24, 25]; and Kappa,
which measures the level of agreement between observers or classifiers in a
classification problem, present in 2 studies [93, 45].
In addition to classification-focused metrics, there are also metrics that
specifically target regression models. Mean Absolute Error (MAE), Mean
Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), Mean
Squared Error (MSE), and Coefficient of Determination (R2) are a few of
these. By quantifying the discrepancy between predicted values and actual
values, these metrics seek to assess the precision of numerical predictions and
offer information about how well the model fits the regression data. MAE,
MAPE, and RMSE were the most commonly employed regression metrics,
appearing in 3 studies each [74, 89, 61, 36, 69]. R2 and MSE were present in
2 studies [74, 53, 42].
Most regression-based models are commonly applied in solving Prognos-
tics and Health Management (PHM) problems, as demonstrated in the works
by [121, 122, 123]. These studies primarily focus on predicting the remaining
lifetime of systems using monitoring data as input. An alternative approach
to tackle this scenario is through the application of a time series approach,
as observed in the works by [124, 125]. However, this approach often em-
phasizes predicting the future state of systems rather than real-time state
identification. Within this context, the Health Indicator (HI) metric, as em-
ployed in Zhai et al. (2021) [28] and Guo et al. (2022) [27], aims to evaluate
the health status of monitored equipment.
Metrics concerning time performance, such as training time and predic-
tion time, were also identified in 5 [33, 46, 54, 66, 42] and 4 [33, 59, 54, 66]
studies, respectively. These metrics are not unique to any specific model
type but serve to assess the efficiency and duration of model training and
prediction generation. They offer valuable insights into the scalability and

27
practical feasibility of implementing the model, considering computational
resources and processing time. These metrics are crucial for optimizing re-
source allocation, determining model suitability for real-time applications,
and evaluating overall system performance in time-sensitive scenarios.
Finally, it is clear that there is a lack of research in the literature on the
assessment of inference time, i.e. the time the proposed approach takes to
process the data stream and classify it as normal or anomalous. Importantly,
only 4 of the 84 works in this mapping report the inference time of the pro-
posed model. This emphasizes the necessity for additional study to address
this issue and develop a thorough grasp of the computational needs and ap-
plicability of AD methods in real-time. By examining prediction time along-
side other performance metrics, we can improve decision-making processes
and identify models that strike a balance between accuracy and efficiency in
time-critical domains.

5. Open Challenges
While AD techniques for industry 4.0 are gaining attention in recent years,
several open challenges have been identified by this SM. This section high-
lights these challenges and provides insights into potential areas for future
research.
Anomaly detection on the edge: The rise of edge computing offers
promising potential for real-time AD, particularly for industrial internet of
things (IIoT) sensors. However, edge devices are constrained by limited re-
sources, such as processing power and memory, making complex anomaly
detection tasks challenging to execute on these devices. To address this chal-
lenge, there is a need to develop suitably optimized AD algorithms that can
navigate the trade-off between precision and computational resources. Mul-
tiobjective optimization algorithms can be used to finetune solutions under
simultaneous conflicting objectives, such as power consumption, accuracy
and inference speed [126]. Additionally, spiking neural networks and neuro-
morphic computing are emerging technologies that can be used for real-time
anomaly detection with significantly lower computational resources than tra-
ditional approaches [127].
Usage of low-cost and off-the-shelf sensors: The adoption of IIoT
devices with low-cost sensors could facilitate a wider implementation of pre-
dictive maintenance in the industry. However, a prominent challenge lies in
the quality and reliability of the data collected from these sensors. Low-cost

28
sensors may exhibit higher noise levels, lower accuracy, and reduced stabil-
ity over time compared to their more expensive counterparts. These factors
can significantly impact the performance of AD algorithms, resulting in a
higher rate of false positives. Consequently, AD algorithms would require
additional adjustments, such as implementing an automatic update routine
to compensate for the lower reliability and potential sensory drift over time.
Identification of the anomaly source: While AD techniques can in-
dicate the occurrence of an anomaly based on normal operation data alone,
fault diagnosis remains a challenging next step. Fault diagnosis is crucial for
understanding the root cause of the anomaly and taking appropriate actions.
With the emergence of XAI, there is an increasing demand for models that
not only detect anomalies but also provide interpretable information about
their likely causes. This explanation of the potential causes of the anomaly is
even more valuable in environments where downtime is critical, as it enables
operators to make faster decisions.
Lack of fault data: In industrial settings, it is often the case that there is
an abundance of non-anomalous data, representing healthy machinery, while
the amount of anomalous data is relatively small or even non-existent. This
represents a significant challenge when it comes to using supervised learn-
ing approaches, as these methods heavily rely on labeled anomalous data for
training and generalization of the models. Therefore, addressing this chal-
lenge becomes crucial. One possible solution to tackle the lack of anomalous
data is to leverage a combination of simulation and experimental data. By
utilizing simulated scenarios and conducting controlled experiments, it be-
comes possible to generate synthetic anomalous data. This synthetic data
can then be used for the initial validation and fine-tuning of the AD algorithm
in a more controlled and manageable environment.
Integration with existing infrastructure: One of the key challenges
in AD is the integration of AD systems with the pre-existing infrastructure in
industrial environments. Industrial systems often have well-established data
collection mechanisms and processes in place. Therefore, it is essential for
AD solutions to be compatible with these existing systems and seamlessly
integrate into the infrastructure without causing disruptions or requiring sig-
nificant changes. The integration aspect should be a primary consideration
in research studies, and future works should emphasize how their approaches
can effectively adapt to the already consolidated industrial environments, en-
suring smooth incorporation without major interruptions or modifications.
Retraining of ML models The retraining of ML models is a crucial

29
aspect that requires attention and careful discussion regarding its adaptabil-
ity to the evolution of industries, changes in the environment and new data
collected. Ensuring the continued effectiveness of ML models in detecting
anomalies depends on their ability to adapt to changes that occur over time
in industries. As industries undergo operational changes, these models need
to be capable of adjusting and learning from the new environment. Fur-
thermore, the inclusion of newly collected data and exposure to new failures
is essential to improve the generalizability of the models, enabling them to
identify anomalies with greater precision.

6. Limitations of the Mapping Study

Risks and restrictions apply to systematic mappings like this one [19]. In
this section, we address the most frequent restrictions encountered during
the review process and provide our solutions to overcome them.
Formulation of research questions: The development of research
questions is essential for directing the review process. However, if the ques-
tions are not carefully designed or if they lack specificity, it is possible to
mistakenly exclude pertinent studies or ignore crucial components. By care-
fully crafting and fine-tuning the research questions through consultations
with authors and outside experts, we were able to lessen this constraint.
Searching: Despite using a thorough search method, it is probable that
some pertinent studies were missed. This can be as a result of restrictions
on the databases that were chosen, limitations on language, or the exclu-
sion of specific publishing categories. We maintained the same terms in our
search strings while modifying them for each digital database to lessen this
restriction.
Misclassification or errors in data extraction: These terms allude to
the potential for various reviewers to interpret the data from research in dif-
ferent ways. Although we used our judgment to categorize the studies, there
is still a chance that we did so incorrectly. Multiple author-researchers were
involved in the classification process to help alleviate this possible problem,
and any disagreements were settled by consensus discussions.
We aim to provide a comprehensive and unbiased evaluation of the lit-
erature within the scope of this review, acknowledging these limitations and
employing appropriate methodologies.

30
7. Conclusions and next steps
This SM focused on the topic of AD in industrial machinery using IoT de-
vices and ML algorithms. The study aimed to fill the gap in existing research
by providing a comprehensive review of current methodologies, synthesis of
evidence, areas of application, research questions and future challenges in
this domain. Through the analysis of 84 studies dating from 2016 to 2023,
several important findings emerged. First, it was observed that the types
of machinery most commonly monitored include milling and cutting tools,
hydraulic systems, and bearings. The analysis suggests that these types of
machinery are often subjected to conditions that can lead to anomalies, such
as wear and tear in milling and cutting tools, fluctuations in temperature
and pressure in hydraulic systems, and heat-related failures in bearings.
Second, a wide range of sensors were employed to detect anomalies. The
primary works often did not detail specific characteristics of the sensors, but
we were able to highlight the most used sensors, such as vibration and tem-
perature. A combination of these sensors is observed as a common strategy
to efficiently detect and monitor anomalies in machines. Third, several ML
techniques have been used for AD, including supervised, unsupervised and
heuristic methods. In addition to highlighting preprocessing techniques, we
were able to highlight the most used and also those that normally worked
together. Combination of these techniques, such as FFT and autoencoders,
are found to be effective to enhance AD accuracy.
The review also identified several open challenges in the field of AD.
These challenges include detecting anomalies at the edge using off-the-shelf,
low-cost sensors, identifying the source of anomalies, addressing the lack of
fault data, integrating AD into existing infrastructure, and retraining ML
models to adapt to evolving industrial environments. In summary, this SM
clarified the state of the art of AD in industrial machinery using IoT devices
and ML. It identified key research areas, highlighted current methodologies,
and outlined open challenges. By addressing these challenges, we aim to
contribute to the development of more accurate, efficient, and interpretable
AD systems, thereby enhancing industrial productivity, safety, and efficiency
in the era of smart manufacturing
For future work, our intention is to develop an anomaly detection solu-
tion based on low-cost and off-the-shelf components for rotating machinery.
Additionally, we aim to focus on developing anomaly detection on embedded
devices, with a priority on low power consumption and local processing of

31
most of the data. As our research progresses, we plan to make newly created
anomaly datasets publicly accessible, contributing to the broader research
community’s resources

CRediT authorship contribution statement

Sérgio F. Chevtchenko: Writing - original draft, conceptualization, method-
ology, investigation, data curation, and formal analysis. Elisson da Silva
Rocha: Methodology, Writing - original draft, data curation, and formal
analysis. Monalisa Cristina Moura dos Santos: Writing - original draft and
review, conceptualization, methodology, and data curation. Diego Moura
Vieira: Writing - original draft and review, conceptualization, methodology,
and data curation. Ricardo Lins Mota: Writing - original draft and review,
conceptualization, methodology, and data curation. Ermeson Carneiro de
Andrade: Methodology, Writing - review and editing, Supervision. Danilo
Ricardo Barbosa de Araújo: Methodology, Writing - review and editing, Su-
pervision.

Acknowledgment
The authors would like to thank the ’Brazilian Agency for Industrial
Development - ABDI’ for financial support.

Appendix A. Supplementary material

References
[1] J. Lee, H.-A. Kao, S. Yang, Service innovation and smart analytics for
industry 4.0 and big data environment, Procedia cirp 16 (2014) 3–8.

[2] L. Kaupp, H. Webert, K. Nazemi, B. Humm, S. Simons, Context: An

industry 4.0 dataset of contextual faults in a smart factory, Procedia
Computer Science 180 (2021) 492–501.

[3] V. Chandola, A. Banerjee, V. Kumar, Anomaly detection: A survey,

ACM computing surveys (CSUR) 41 (2009) 1–58.

[4] A. Chatterjee, B. S. Ahmed, Iot anomaly detection methods and ap-

plications: A survey, Internet of Things 19 (2022) 100568.

32
 [5] F. Huch, M. Golagha, A. Petrovska, A. Krauss, Machine learning-
based run-time anomaly detection in software systems: An industrial
evaluation, in: 2018 IEEE Workshop on Machine Learning Techniques
for Software Quality Evaluation (MaLTeSQuE), IEEE, pp. 13–18.
[6] M. Ahmed, A. N. Mahmood, J. Hu, A survey of network anomaly
detection techniques, Journal of Network and Computer Applications
60 (2016) 19–31.
[7] S. Eltanbouly, M. Bashendy, N. AlNaimi, Z. Chkirbene, A. Erbad, Ma-
chine learning techniques for network anomaly detection: A survey, in:
2020 IEEE International Conference on Informatics, IoT, and Enabling
Technologies (ICIoT), IEEE, pp. 156–162.
[8] S. Wang, J. F. Balarezo, S. Kandeepan, A. Al-Hourani, K. G. Chavez,
B. Rubinstein, Machine learning in network anomaly detection: A
survey, IEEE Access 9 (2021) 152379–152396.
[9] J. Dalzochio, R. Kunst, E. Pignaton, A. Binotto, S. Sanyal, J. Favilla,
J. Barbosa, Machine learning and reasoning for predictive maintenance
in industry 4.0: Current status and challenges, Computers in Industry
123 (2020) 103298.
[10] T. Zonta, C. A. Da Costa, R. da Rosa Righi, M. J. de Lima, E. S.
da Trindade, G. P. Li, Predictive maintenance in the industry 4.0: A
systematic literature review, Computers & Industrial Engineering 150
(2020) 106889.
[11] R. Kumar, A. Malik, V. Ranga, An intellectual intrusion detection
system using hybrid hunger games search and remora optimization al-
gorithm for iot wireless networks, Knowledge-Based Systems 256 (2022)
109762.
[12] L. Gupta, T. Salman, A. Ghubaish, D. Unal, A. K. Al-Ali, R. Jain,
Cybersecurity of multi-cloud healthcare systems: A hierarchical deep
learning approach, Applied Soft Computing 118 (2022) 108439.
[13] A. Gaddam, T. Wilkin, M. Angelova, Anomaly detection models for
detecting sensor faults and outliers in the iot-a survey, in: 2019 13th
International Conference on Sensing Technology (ICST), IEEE, pp. 1–
6.

33
[14] P. Kamat, R. Sugandhi, Anomaly detection for predictive maintenance
in industry 4.0-a survey, in: E3S web of conferences, volume 170, EDP
Sciences, p. 02007.

[15] S. Schwendemann, Z. Amjad, A. Sikora, A survey of machine-learning

techniques for condition monitoring and predictive maintenance of
bearings in grinding machines, Computers in Industry 125 (2021)
103380.

[16] Z. Kang, C. Catal, B. Tekinerdogan, Machine learning applications in

production lines: A systematic literature review, Computers & Indus-
trial Engineering 149 (2020) 106773.

[17] N. Davari, B. Veloso, G. d. A. Costa, P. M. Pereira, R. P. Ribeiro,

J. Gama, A survey on data-driven predictive maintenance for the rail-
way industry, Sensors 21 (2021) 5739.

[18] A. K. M. Nor, S. R. Pedapati, M. Muhammad, V. Leiva, Overview of

explainable artificial intelligence for prognostic and health management
of industrial assets based on preferred reporting items for systematic
reviews and meta-analyses, Sensors 21 (2021) 8020.

[19] K. Petersen, R. Feldt, S. Mujtaba, M. Mattsson, Systematic mapping

studies in software engineering, in: 12th International Conference on
Evaluation and Assessment in Software Engineering (EASE) 12, pp.
1–10.

[20] J. Mendonca, E. Andrade, P. T. Endo, R. Lima, Disaster recovery so-

lutions for it systems: A systematic mapping study, Journal of Systems
and Software 149 (2019) 511–530.

[21] B. A. Kitchenham, E. Mendes, G. H. Travassos, Cross versus within-

company cost estimation studies: A systematic review, IEEE Transac-
tions on Software Engineering 33 (2007) 316–329.

[22] J. Jakubowski, P. Stanisz, S. Bobek, G. J. Nalepa, Anomaly detection

in asset degradation process using variational autoencoder and expla-
nations, Sensors 22 (2021) 291.

[23] S. H. Hong, T. Kyzer, J. Cornelius, F. Zahiri, Y. Wang, Intelligent

anomaly detection of robot manipulator based on energy consumption

34
 auditing, in: 2022 IEEE Aerospace Conference (AERO), IEEE, pp.
1–11.

[24] M. Assafo, P. Langend¨"orfer, A topsis-assisted feature selection scheme

and som-based anomaly detection for milling tools under different op-
erating conditions, IEEE Access 9 (2021) 90011–90028.

[25] T. Żabiński, Z. Hajduk, J. Kluska, L. Gniewek, Fpga-embedded

anomaly detection system for milling process, IEEE Access 9 (2021)
124059–124069.

[26] S. Ryu, J. Yim, J. Seo, Y. Yu, H. Seo, Quantile autoencoder with

abnormality accumulation for anomaly detection of multivariate sensor
data, IEEE Access 10 (2022) 70428–70439.

[27] L. Guo, Y. Yu, A. Duan, H. Gao, J. Zhang, An unsupervised feature

learning based health indicator construction method for performance
assessment of machines, Mechanical Systems and Signal Processing
167 (2022) 108573.

[28] S. Zhai, B. Gehring, G. Reinhart, Enabling predictive maintenance

integrated production scheduling by operation-specific health prognos-
tics with generative deep learning, Journal of Manufacturing Systems
61 (2021) 830–855.

[29] T. Watanabe, I. Kono, H. Onozuka, Anomaly detection methods in

turning based on motor data analysis, Procedia Manufacturing 48
(2020) 882–893.

[30] D. Kim, T.-Y. Heo, Anomaly detection with feature extraction based
on machine learning using hydraulic system iot sensor data, Sensors
22 (2022) 2479.

[31] S. Abbasi, M. Famouri, M. J. Shafiee, A. Wong, Outliernets: Highly

compact deep autoencoder network architectures for on-device acoustic
anomaly detection, Sensors 21 (2021) 4805.

[32] B. Boons, M. Verhelst, P. Karsmakers, Low power on-line machine

monitoring at the edge, in: 2021 International Conference on Applied
Artificial Intelligence (ICAPAI), pp. 1–8.

35
[33] E. Di Fiore, A. Ferraro, A. Galli, V. Moscato, G. Sperlì, An anoma-
lous sound detection methodology for predictive maintenance, Expert
Systems with Applications 209 (2022) 118324.

[34] Y. Tagawa, R. Maskeliūnas, R. Damaševičius, Acoustic anomaly de-

tection of mechanical failures in noisy real-life factory environments,
Electronics 10 (2021) 2329.

[35] D. Juodelyte, V. Cheplygina, T. Graversen, P. Bonnet, Predicting

bearings’ degradation stages for predictive maintenance in the phar-
maceutical industry, arXiv preprint arXiv:2203.03259 (2022).

[36] N. Song, X. Hu, N. Li, Anomaly detection of wind turbine generator

based on temporal information, in: Proceedings of the 2019 7th Inter-
national Conference on Information Technology: IoT and Smart City,
pp. 477–482.

[37] F. K¨"onig, C. Sous, A. O. Chaib, G. Jacobs, Machine learning based

anomaly detection and classification of acoustic emission events for
wear monitoring in sliding bearing systems, Tribology International
155 (2021) 106811.

[38] A. Saxena, K. Goebel, D. L. Simon, Damage propagation modeling

for aircraft engine run-to-failure simulation, Aerospace Science and
Technology 12 (2008) 399–408.

[39] S. K. Bose, B. Kar, M. Roy, P. K. Gopalakrishnan, A. Basu, Adepos:

Anomaly detection based power saving for predictive maintenance us-
ing edge computing, in: Proceedings of the 24th asia and south pacific
design automation conference, pp. 597–602.

[40] T. P. Q. Nguyen, P. N. K. Phuc, C.-L. Yang, H. Sutrisno, B.-H. Lu-

ong, T. H. A. Le, T. T. Nguyen, Time-series anomaly detection using
dynamic programming based longest common subsequence on sensor
data, Expert Systems with Applications 213 (2023) 118902.

[41] K. PATRA, R. N. SETHI, D. K. BEHERA, Anomaly detection in

rotating machinery using autoencoders based onbidirectional lstm and
gru neural networks, Turkish Journal of Electrical Engineering and
Computer Sciences 30 (2022) 1637–1653.

36
[42] A. Mostafavi, A. Sadighi, A novel online machine learning approach
for real-time condition monitoring of rotating machines, in: 2021 9th
RSI International Conference on Robotics and Mechatronics (ICRoM),
pp. 267–273.

[43] G. Jiang, S. Nie, P. Xie, Y. Li, X. Li, Multiscale one-class classifica-

tion network for machine health monitoring, IEEE Sensors Journal 22
(2022) 13043–13054.

[44] Y.-C. Yu, S.-W. Chuang, H.-H. Shuai, C.-Y. Lee, Fast adaption for
multi motor anomaly detection via meta learning and deep unsuper-
vised learning, in: 2022 IEEE 31st International Symposium on Indus-
trial Electronics (ISIE), pp. 1186–1189.

[45] K. Vos, Z. Peng, C. Jenkins, M. R. Shahriar, P. Borghesani, W. Wang,

Vibration-based anomaly detection using lstm/svm approaches, Me-
chanical Systems and Signal Processing 169 (2022) 108752.

[46] L. C. Brito, G. A. Susto, J. N. Brito, M. A. Duarte, An explainable ar-

tificial intelligence approach for unsupervised fault detection and diag-
nosis in rotating machinery, Mechanical Systems and Signal Processing
163 (2022) 108105.

[47] S. Jian, S. Ishida, Y. Arakawa, Initial attempt on wi-fi csi based vi-
bration sensing for factory equipment fault detection, in: Adjunct
Proceedings of the 2021 International Conference on Distributed Com-
puting and Networking, pp. 163–168.

[48] R. de Paula Monteiro, M. C. Lozada, D. R. C. Mendieta, R. V. S.

Loja, C. J. A. Bastos Filho, A hybrid prototype selection-based deep
learning approach for anomaly detection in industrial machines, Expert
Systems with Applications (2022) 117528.

[49] C. Li, L. Mo, H. Tang, R. Yan, Lifelong condition monitoring based

on nb-iot for anomaly detection of machinery equipment, Procedia
Manufacturing 49 (2020) 144–149.

[50] F. Pittino, M. Puggl, T. Moldaschl, C. Hirschl, Automatic anomaly

detection on in-production manufacturing machines using statistical
learning methods, Sensors 20 (2020) 2344.

37
[51] A. L. Alfeo, M. G. Cimino, G. Manco, E. Ritacco, G. Vaglini, Using
an autoencoder in the design of an anomaly detector for smart manu-
facturing, Pattern Recognition Letters 136 (2020) 272–278.

[52] H. Yun, H. Kim, E. Kim, M. B. Jun, Development of internal sound

sensor using stethoscope and its applications for machine monitoring,
Procedia Manufacturing 48 (2020) 1072–1078.

[53] Z. Ying, L. Shu, T. Kizaki, M. Iwama, N. Sugita, Hybrid approach for

onsite monitoring and anomaly detection of cutting tool life, Procedia
CIRP 104 (2021) 1541–1546.

[54] M. Meire, P. Karsmakers, Comparison of deep autoencoder architec-

tures for real-time acoustic based anomaly detection in assets, in:
2019 10th IEEE International Conference on Intelligent Data Acquisi-
tion and Advanced Computing Systems: Technology and Applications
(IDAACS), volume 2, IEEE, pp. 786–790.

[55] H. Ahn, I. Yeo, Deep-learning-based approach to anomaly detection

techniques for large acoustic data in machine operation, Sensors 21
(2021) 5446.

[56] T. Gr¨"uner, F. B¨"ollhoff, R. Meisetschl¨"ager, A. Vydrenko, M. Bator,

A. Dicks, A. Theissler, Evaluation of machine learning for sensorless
detection and classification of faults in electromechanical drive systems,
Procedia Computer Science 176 (2020) 1586–1595.

[57] T. Hiruta, K. Maki, T. Kato, Y. Umeda, Unsupervised learning based

diagnosis model for anomaly detection of motor bearing with current
data, Procedia CIRP 98 (2021) 336–341.

[58] S. Givnan, C. Chalmers, P. Fergus, S. Ortega-Martorell, T. Whalley,

Anomaly detection using autoencoder reconstruction upon industrial
motors, Sensors 22 (2022) 3166.

[59] Y. Park, M. J. Kim, Design of cost-effective auto-encoder for electric

motor anomaly detection in resource constrained edge device, in: 2021
IEEE 3rd Eurasia Conference on IOT, Communication and Engineering
(ECICE), IEEE, pp. 241–246.

38
[60] S. Panagou, F. Fruggiero, M. Lerra, C. del Vecchio, F. Menchetti,
L. Piedimonte, O. R. Natale, S. Passariello, Feature investigation with
digital twin for predictive maintenance following a machine learning
approach, IFAC-PapersOnLine 55 (2022) 132–137.

[61] T. Yanabe, H. Nishi, M. Hashimoto, Anomaly detection based on

histogram methodology and factor analysis using lightgbm for cooling
systems, in: 2020 25th IEEE International Conference on Emerging
Technologies and Factory Automation (ETFA), volume 1, IEEE, pp.
952–958.

[62] A. González-Muñiz, I. Díaz, A. A. Cuadrado, D. García-Pérez,

D. Pérez, Two-step residual-error based approach for anomaly detec-
tion in engineering systems using variational autoencoders, Computers
and Electrical Engineering 101 (2022) 108065.

[63] D. Velasquez, E. Perez, X. Oregui, A. Artetxe, J. Manteca, J. E. Man-

silla, M. Toro, M. Maiza, B. Sierra, A hybrid machine-learning en-
semble for anomaly detection in real-time industry 4.0 systems, IEEE
Access 10 (2022) 72024–72036.

[64] M. Canizo, I. Triguero, A. Conde, E. Onieva, Multi-head cnn–rnn for

multi-time series anomaly detection: An industrial case study, Neuro-
computing 363 (2019) 246–260.

[65] J. L. Conradi Hoffmann, L. P. Horstmann, M. Martínez Lucena,

G. Medeiros de Araujo, A. A. Fr¨"ohlich, M. H. N. Nishioka, Anomaly
detection on wind turbines based on a deep learning analysis of vibra-
tion signals, Applied Artificial Intelligence 35 (2021) 893–913.

[66] S. Yan, H. Shao, Y. Xiao, B. Liu, J. Wan, Hybrid robust convo-

lutional autoencoder for unsupervised anomaly detection of machine
tools under noises, Robotics and Computer-Integrated Manufacturing
79 (2023) 102441.

[67] B. Denkena, M.-A. Dittrich, H. Noske, D. Stoppel, D. Lange, Data-

based ensemble approach for semi-supervised anomaly detection in ma-
chine tool condition monitoring, CIRP Journal of Manufacturing Sci-
ence and Technology 35 (2021) 795–802.

39
[68] W.-T. Yang, M. S. Reis, V. Borodin, M. Juge, A. Roussy, An inter-
pretable unsupervised bayesian network model for fault detection and
diagnosis, Control Engineering Practice 127 (2022) 105304.

[69] E.-S. Apostol, C.-O. Truică, F. Pop, C. Esposito, Change point en-
hanced anomaly detection for iot time series data, Water 13 (2021)
1633.

[70] B. Lindemann, F. Fesenmayr, N. Jazdi, M. Weyrich, Anomaly detec-

tion in discrete manufacturing using self-learning approaches, Procedia
CIRP 79 (2019) 313–318.

[71] K. Meyer, V. Mahalec, Anomaly detection methods for infrequent

failures in resistive steel welding, Journal of Manufacturing Processes
75 (2022) 497–513.

[72] K. Kammerer, B. Hoppenstedt, R. Pryss, S. St¨"okler, J. Allgaier,

M. Reichert, Anomaly detections for manufacturing systems based
on sensor data—insights into two challenging real-world production
settings, Sensors 19 (2019) 5370.

[73] R. Langone, A. Cuzzocrea, N. Skantzos, Interpretable anomaly pre-

diction: Predicting anomalous behavior in industry 4.0 settings via
regularized logistic regression tools, Data & Knowledge Engineering
130 (2020) 101850.

[74] S. Ayvaz, K. Alpay, Predictive maintenance system for production

lines in manufacturing: A machine learning approach using iot data in
real-time, Expert Systems with Applications 173 (2021) 114598.

[75] P. Calvo-Bascones, A. Voisin, P. Do, M. A. Sanz-Bobi, A collaborative

network of digital twins for anomaly detection applications of complex
systems. snitch digital twin concept, Computers in Industry 144 (2023)
103767.

[76] F. Abbracciavento, S. Formentin, J. Balocco, A. Rota, V. Manzoni,

S. M. Savaresi, Anomaly detection via distributed sensing: a var mod-
eling approach, IFAC-PapersOnLine 54 (2021) 85–90.

40
[77] X. Wang, J. Lin, N. Patel, M. Braun, A self-learning and online algo-
rithm for time series anomaly detection, with application in cpu man-
ufacturing, in: Proceedings of the 25th ACM International on Confer-
ence on Information and Knowledge Management, pp. 1823–1832.

[78] M. Wielgosz, M. Karwatowski, Mapping neural networks to fpga-based

iot devices for ultra-low latency processing, Sensors 19 (2019) 2981.

[79] A. T. Keleko, B. Kamsu-Foguem, R. H. Ngouna, A. Tongne, Health

condition monitoring of a complex hydraulic system using deep neu-
ral network and deepshap explainable xai, Advances in Engineering
Software 175 (2023) 103339.

[80] D. Coelho, D. Costa, E. M. Rocha, D. Almeida, J. P. Santos, Pre-

dictive maintenance on sensorized stamping presses by time series seg-
mentation, anomaly detection, and classification algorithms, Procedia
Computer Science 200 (2022) 1184–1193.

[81] G. P. Tancredi, G. Vignali, E. Bottani, Integration of digital twin,

machine-learning and industry 4.0 tools for anomaly detection: An
application to a food plant, Sensors 22 (2022) 4143.

[82] A. F. de Galatheau, A.-L. Olteanu, N. Julien, S. Le Garrec, Long short

term memory-based anomaly detection applied to an industrial dosing
pump, IFAC-PapersOnLine 55 (2022) 240–245.

[83] K. Liu, W. Mao, H. Shi, C. Wu, J. Chen, Online anomaly detection

with streaming data based on fine-grained feature forecasting, in: 2021
33rd Chinese Control and Decision Conference (CCDC), pp. 454–459.

[84] M. Antonini, M. Pincheira, M. Vecchio, F. Antonelli, A tinyml ap-

proach to non-repudiable anomaly detection in extreme industrial en-
vironments, in: 2022 IEEE International Workshop on Metrology for
Industry 4.0 & IoT (MetroInd4. 0&IoT), IEEE, pp. 397–402.

[85] M. Züfle, F. Moog, V. Lesch, C. Krupitzer, S. Kounev, A machine

learning-based workflow for automatic detection of anomalies in ma-
chine tools, ISA Transactions 125 (2022) 445–458.

[86] R. Yasaei, F. Hernandez, M. A. A. Faruque, Iot-cad: Context-aware

adaptive anomaly detection in iot systems through sensor association,

41
 in: Proceedings of the 39th International Conference on Computer-
Aided Design, pp. 1–9.

[87] F. De Vita, D. Bruneo, S. K. Das, On the use of a full stack hard-

ware/software infrastructure for sensor data fusion and fault prediction
in industry 4.0, Pattern Recognition Letters 138 (2020) 30–37.

[88] K. Hendrickx, W. Meert, Y. Mollet, J. Gyselinck, B. Cornelis, K. Gryl-

lias, J. Davis, A general anomaly detection framework for fleet-based
condition monitoring of machines, Mechanical Systems and Signal Pro-
cessing 139 (2020) 106585.

[89] C. Kim, I. Joe, D. Jang, E. Kim, S. Nam, Online monitoring automa-

tion using anomaly detection in iot/it environment, in: Computer
Science On-line Conference, Springer, pp. 96–106.

[90] P. Calvo-Bascones, M. A. Sanz-Bobi, T. M. Welte, Anomaly detection

method based on the deep knowledge behind behavior patterns in in-
dustrial components. application to a hydropower plant, Computers in
Industry 125 (2021) 103376.

[91] H. Lu, M. Du, K. Qian, X. He, K. Wang, Gan-based data augmenta-

tion strategy for sensor anomaly detection in industrial robots, IEEE
Sensors Journal 22 (2021) 17464–17474.

[92] H. Shi, W. Mao, G. Wang, K. Liu, Deep multi-task svdd: A new

robust online detection method of bearings early fault, in: 2021 Global
Reliability and Prognostics and Health Management (PHM-Nanjing),
IEEE, pp. 1–7.

[93] B. N. DS, V. Dondeti, S. Balakrishna, Comparative analysis of machine

learning-based algorithms for detection of anomalies in iiot, Interna-
tional Journal of Information Retrieval Research (IJIRR) 12 (2022)
1–55.

[94] Q. Luo, J. Chen, Y. Zi, Y. Chang, Y. Feng, Multi-mode non-gaussian

variational autoencoder network with missing sources for anomaly de-
tection of complex electromechanical equipment, ISA transactions
(2022).

42
 [95] H. Choi, D. Kim, J. Kim, J. Kim, P. Kang, Explainable anomaly detec-
tion framework for predictive maintenance in manufacturing systems,
Applied Soft Computing 125 (2022) 109147.

[96] Q. Qian, X. Fang, J. Xu, M. Li, Multichannel profile-based monitoring

method and its application in the basic oxygen furnace steelmaking
process, Journal of Manufacturing Systems 61 (2021) 375–390.

[97] N. Iftikhar, T. Baattrup-Andersen, F. E. Nordbjerg, K. Jeppesen, Out-

lier detection in sensor data using ensemble learning, Procedia Com-
puter Science 176 (2020) 1160–1169.

[98] B. Wang, Y. Li, Y. Luo, X. Li, T. Freiheit, Early event detection in

a deep-learning driven quality prediction model for ultrasonic welding,
Journal of Manufacturing Systems 60 (2021) 325–336.

[99] D. Wu, Z. Jiang, X. Xie, X. Wei, W. Yu, R. Li, Lstm learning with
bayesian and gaussian processing for anomaly detection in industrial
iot, IEEE Transactions on Industrial Informatics 16 (2019) 5244–5253.

[100] Y. Wang, M. Perry, D. Whitlock, J. W. Sutherland, Detecting anoma-

lies in time series data from a manufacturing system using recurrent
neural networks, Journal of Manufacturing Systems 62 (2022) 823–834.
¨
[101] M. Netzer, J. Bach, A. Puchta, P. G"onnheimer, J. Fleischer, Process
segmented based intelligent anomaly detection in highly flexible pro-
duction machines under low machine data availability, Procedia CIRP
107 (2022) 647–652.

[102] E. Lughofer, A.-C. Zavoianu, R. Pollak, M. Pratama, P. Meyer-Heye,

H. Z¨"orrer, C. Eitzinger, T. Radauer, On-line anomaly detection with
advanced independent component analysis of multi-variate residual sig-
nals from causal relation networks, Information sciences 537 (2020)
425–451.

[103] H. Li, X. Wang, Z. Yang, S. Ali, N. Tong, S. Baseer, Correlation-based

anomaly detection method for multi-sensor system, Computational
Intelligence and Neuroscience 2022 (2022).

[104] S. Kabadayi, A. Pridgen, C. Julien, Virtual sensors: Abstracting data

from physical sensors, in: 2006 International Symposium on a World

43
 of Wireless, Mobile and Multimedia Networks (WoWMoM’06), IEEE,
pp. 6–pp.

[105] Z. Lai, I. Alzugaray, M. Chli, E. Chatzi, Full-field structural moni-

toring using event cameras and physics-informed sparse identification,
Mechanical Systems and Signal Processing 145 (2020) 106905.

[106] A. Géron, Hands-on machine learning with Scikit-Learn, Keras, and

TensorFlow, " O’Reilly Media, Inc.", 2022.

[107] P. Baldi, Autoencoders, unsupervised learning, and deep architectures,

in: Proceedings of ICML workshop on unsupervised and transfer learn-
ing, JMLR Workshop and Conference Proceedings, pp. 37–49.

[108] C. S. Wickramasinghe, D. L. Marino, M. Manic, Resnet autoencoders

for unsupervised feature learning from high-dimensional data: Deep
models resistant to performance degradation, IEEE Access 9 (2021)
40511–40520.

[109] Z. Ali, S. B. Bhaskar, Basic statistical tools in research and data

analysis, Indian journal of anaesthesia 60 (2016) 662.

[110] R. Aggarwal, J. K. Singh, V. K. Gupta, S. Rathore, M. Tiwari,

A. Khare, Noise reduction of speech signal using wavelet transform
with modified universal threshold, International Journal of Computer
Applications 20 (2011) 14–19.

[111] D. Groth, S. Hartmann, S. Klie, J. Selbig, Principal components anal-

ysis, Computational Toxicology: Volume II (2013) 527–547.

[112] A. Smolinska, A.-C. Hauschild, R. Fijten, J. Dallinga, J. Baum-

bach, F. Van Schooten, Current breathomics—a review on data pre-
processing techniques and machine learning in metabolomics breath
analysis, Journal of breath research 8 (2014) 027105.

[113] M. Jogin, M. Madhulika, G. Divya, R. Meghana, S. Apoorva, et al.,

Feature extraction using convolution neural networks (cnn) and deep
learning, in: 2018 3rd IEEE international conference on recent trends
in electronics, information & communication technology (RTEICT),
IEEE, pp. 2319–2323.

44
[114] H. Zhou, K. Yu, X. Zhang, G. Wu, A. Yazidi, Contrastive autoencoder
for anomaly detection in multivariate time series, Information Sciences
610 (2022) 266–280.

[115] H. Liang, L. Song, J. Wang, L. Guo, X. Li, J. Liang, Robust unsu-

pervised anomaly detection via multi-time scale dcgans with forgetting
mechanism for industrial multivariate time series, Neurocomputing 423
(2021) 444–462.

[116] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta,

A. A. Bharath, Generative adversarial networks: An overview, IEEE
signal processing magazine 35 (2018) 53–65.

[117] S. M. Erfani, S. Rajasegarar, S. Karunasekera, C. Leckie, High-

dimensional and large-scale anomaly detection using a linear one-class
svm with deep learning, Pattern Recognition 58 (2016) 121–134.

[118] F. T. Liu, K. M. Ting, Z.-H. Zhou, Isolation forest, in: 2008 eighth
ieee international conference on data mining, IEEE, pp. 413–422.

[119] Z. Cheng, C. Zou, J. Dong, Outlier detection using isolation forest and
local outlier factor, in: Proceedings of the conference on research in
adaptive and convergent systems, pp. 161–168.

[120] M. Hossin, M. N. Sulaiman, A review on evaluation metrics for

data classification evaluations, International journal of data mining
& knowledge management process 5 (2015) 1.

[121] S. Zhao, C. Zhang, Y. Wang, Lithium-ion battery capacity and remain-

ing useful life prediction using board learning system and long short-
term memory neural network, Journal of Energy Storage 52 (2022)
104901.

[122] T. Berghout, M. Benbouzid, A systematic guide for predicting remain-

ing useful life with machine learning, Electronics 11 (2022).

[123] S. Siahpour, X. Li, J. Lee, A novel transfer learning approach in remain-

ing useful life prediction for incomplete dataset, IEEE Transactions on
Instrumentation and Measurement 71 (2022) 1–11.

45
[124] S. Fu, S. Zhong, L. Lin, M. Zhao, A novel time-series memory auto-
encoder with sequentially updated reconstructions for remaining useful
life prediction, IEEE Transactions on Neural Networks and Learning
Systems 33 (2022) 7114–7125.

[125] C.-P. Lin, J. Cabrera, F. Yang, M.-H. Ling, K.-L. Tsui, S.-J. Bae,
Battery state of health modeling and remaining useful life prediction
through time series model, Applied Energy 275 (2020) 115338.

[126] M. T. Emmerich, A. H. Deutz, A tutorial on multiobjective optimiza-

tion: fundamentals and evolutionary methods, Natural computing 17
(2018) 585–609.

[127] F. C. Bauer, D. R. Muir, G. Indiveri, Real-time ultra-low power

ecg anomaly detection using an event-driven neuromorphic processor,
IEEE transactions on biomedical circuits and systems 13 (2019) 1575–
1582.

46