Heliyon 9 (2023) e17584

Contents lists available at ScienceDirect

Heliyon
journal homepage: www.cell.com/heliyon

Review article

Machine learning for fault analysis in rotating machinery: A

comprehensive review
Oguzhan Das a, Duygu Bagci Das b, Derya Birant c, *
a
National Defence University, Air NCO Higher Vocational School, Department of Aeronautics Sciences, Izmir, Turkey
b
Ege University, Ege Vocational School, Department of Computer Programming, Izmir, Turkey
c
Dokuz Eylül University, Department of Computer Engineering, Izmir, Turkey

A R T I C L E I N F O A B S T R A C T

Keywords: As the concept of Industry 4.0 is introduced, artificial intelligence-based fault analysis is attracted
Intelligent fault diagnosis the corresponding community to develop effective intelligent fault diagnosis and prognosis (IFDP)
Rotating machine models for rotating machinery. Hence, various challenges arise regarding model assessment,
Transfer learning
suitability for real-world applications, fault-specific model development, compound fault exis­
Machine learning
tence, domain adaptability, data source, data acquisition, data fusion, algorithm selection, and
Deep learning
Challenges and future directions optimization. It is essential to resolve those challenges for each component of the rotating ma­
chinery since each issue of each part has a unique impact on the vital indicators of a machine.
Based on these major obstacles, this study proposes a comprehensive review regarding IFDP
procedures of rotating machinery by minding all the challenges given above for the first time. In
this study, the developed IFDP approaches are reviewed regarding the pursued fault analysis
strategies, considered data sources, data types, data fusion techniques, machine learning tech­
niques within the frame of the fault type, and compound faults that occurred in components such
as bearings, gear, rotor, stator, shaft, and other parts. The challenges and future directions are
presented from the perspective of recent literature and the necessities concerning the IFDP of
rotating machinery.

1. Introduction

The concept of Machine Learning (ML) is an essential branch of artificial intelligence that helps to solve numerous problems in
various disciplines including engineering, health, finance, education, and military in which optimization, prediction, and assessment
are required. ML techniques can be categorized into supervised, unsupervised, semi-supervised, and reinforcement learning. In gen­
eral, supervised learning addresses regression and classification procedures while unsupervised learning is typically employed in
association and cluster problems. Various ML techniques are practiced for different kinds of problems. The most commonly used al­
gorithms are Deep Neural Networks (DNN) algorithms [1–5], Artificial Neural Networks (ANN) [6–8], AdaBoost (AB) [8,9], Random
Forest (RF) [10,11], Support Vector Machines (SVM) [12–14], K Nearest Neighbors (KNN) [12,13,15], and Decision Trees (DT) [16,
17]. As the concept of the Fourth Industrial Revolution or Industry 4.0 is introduced, the use of ML becomes even more critical,
especially for the industry itself. Industry 4.0 aims to constitute smart factories where all industrial procedures are conducted by
establishing communication among industrial systems and humans. Apart from the traditional techniques, the communication

* Corresponding author.
E-mail address: [email protected] (D. Birant).

https://doi.org/10.1016/j.heliyon.2023.e17584
Received 18 February 2023; Received in revised form 9 April 2023; Accepted 21 June 2023
Available online 22 June 2023
2405-8440/© 2023 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
O. Das et al. Heliyon 9 (2023) e17584

procedure is conducted by the cooperation between humans and the machines or the elements of the machines through the internet
[18]. Such communication is provided by software and networked sensors to come up with a better business. The adaptation and
adoption of Industry 4.0 in the industrial divisions (e.g. manufacturing, maintenance, quality assurance, logistics, finance, human
resources, etc.) require developing intelligent models that should operate productively and cost-effectively. ML helps to satisfy these
requirements by bringing powerful and effective techniques that can solve complex problems.
In the field of Industry 4.0-based maintenance, the concept of the Industrial Internet of Things (IIoT) and Cyber-Physical Systems
play a major role. These technologies associate data and machine learning to build intelligent techniques for condition monitoring of
the machines. Hence, developing IFDP methods elicits an opportunity to predict, plan, and apply the maintenance procedure as
effectively as possible. On the other hand, there are several challenges that have to be overcome. These challenges can be split into
organizational challenges, architectural challenges, infrastructural challenges, content and contextual challenges, and integration
challenges [19]. Machine learning-related challenges can be summed up into architectural and content and contextual challenges. As it
is anticipated, the foremost challenge is to constitute an implementable effective IFDP model [20]. The selection, tuning, or modifi­
cation of an IFDP technique is compelling in various ways. It is essential to assess the IFDP method based on its success, time con­
sumption, explainability, applicability, and generalizability. Hence, it is first beneficial to understand the fundamentals of a chosen
technique including its mathematical background, applicability to a given problem, tunable parameters, pros, and cons. In addition, it
is critical to pick or constitute an intelligent model that can be operated without requiring a heavy computational load that may result
in an increase in costs due to the necessity of high-end hardware. During selecting or developing a model, it is also needed to check its
compatibility with the data used for monitoring. Some visual or signal data need to be processed to be employable as input to IFDP
models. In addition, a selected IFDP approach may give fallacious results when small data is used. For instance, deep learning ap­
proaches generally give better results when they are fed with bigger amounts of data. An explainable (XAI) IFDP model would be
favorable since it provides users with an in-depth understanding related to the way of learning of the IFDP model. Following the
selection of IFDP, it has to be tuned or modified if necessary to achieve a robust model which successfully analyzes the machine in a
short time. At this step, it is also necessary to pay attention to overfitting where an intelligent method gives accurate results by
memorizing instead of learning. The challenges do not come to an end when an effective IFDP is obtained. The proposed approach has
to be tested in real-world settings if it is not assessed yet. Based on such challenges, this paper provides a comprehensive overview of
ML-based fault diagnosis and the prognosis of rotating machines in industries to present the recent situation related to this field and
address the challenges, shortcomings, and future directions.
Rotating machines have an essential role to accomplish various purposes in the industry. These machines are the backbone of
mechanized and autonomous production that provides fast and easy-to-reach all products that humanity needs. A sudden outage of the
rotating machines due to a fault or a failure not only disrupt the supply chain but also causes enormous amounts of costs. Therefore, it is
essential to determine a maintenance strategy and arrange a schedule meticulously. For this reason, industries adopted many different
maintenance strategies. As the IIoT is introduced, these maintenance strategies evolved and become even more significant. Especially
the impact of digital technologies on Predictive Maintenance (PdM) improved its preferability significantly. Consequently, developing
the IIoT integrated and ML-based PdM models attracted researchers, and therefore, various studies are published in which different
intelligent fault detection, isolation, diagnosis, and prognosis techniques are proposed, examined, and interpreted.
An intelligent model should be effective, accurate, fast, cheap, low-computational cost, and real-world applicable for fault
detection, fault isolation, fault diagnosis, and fault prognosis of rotating machines. To develop Intelligent Fault Diagnosis and Prog­
nosis (IFDP) model, researchers need to consider suitable data sources, preprocessing techniques, machine learning algorithms, and
optimization algorithms. For this purpose, they deemed vibration [11,14,15,21,22], acoustic [11,23,24], thermal [13], current [6,7,9,
25,26], pressure [27], and other characteristic data [27–29] as the main source for IFDP of rotating machines. Afterwards, the
distinctive features are extracted by employing feature extraction methods such as statistical feature extraction [30–32], Fourier
Transform [33–35], Wavelet Transform [36–38], Empirical Mode Decomposition [28,39,40] or other techniques [6,7,41,42]. The
features may also be extracted automatically by employing deep learning approaches including convolutional neural networks,
autoencoders, long-short term machines, etc. [43–45] Following the feature extraction, a suitable machine learning approach is
necessary to be considered. Researchers employed and improved numerous machine learning approaches [46–48] including deep
learning [49–51] and ensemble learning [12,14,52] methods.
The literature comprises several review studies in which the IFDP approaches regarding rotating machines are discussed. Most of
these studies focused on various aspects including intelligent approaches [53–58] and data preparation/processing [54,55,58–60]
regarding the past, present, and future timeline [58]. On the other hand, few reviews considered fault-specificity [53,57], the
component of the rotating machinery [54,58,59,61], performance metrics [53–56], and fault analysis strategies [62]. However, a gap
still exists regarding these concepts because most of those works focused on a specific component or a group of faults related to a single
component [57,61]. Besides, the relationship and the importance among the fault specificity, selected intelligent approach, data
sources, data processing, data fusion, performance metrics, and fault analysis strategy is not discussed yet. Another significant issue is
the compound faults that involve multi-fault occurrences that take place at the same time in the same or different components. Hence,
this study differs from other reviews by addressing the recent IFDP approaches related to rotating machines and proposing the
challenges and future directions within the concept of fault analysis including fault diagnosis (fault detection, localization, and
identification) and fault prognosis (fault severity assessment and remaining useful life) regarding machine learning approaches,
rotating machine components, single and compound fault-specificity, data sources, and fusion techniques. The contributions of this
review are summarized as follows.

2
O. Das et al. Heliyon 9 (2023) e17584

(i) Presenting a recent and comprehensive discussion regarding IFDP in bearing, gear, rotor, stator, shaft, and other rotating
machine components.
(ii) Discussing the studies concerning various performance metrics and fault specificity regarding intelligent fault diagnosis for the
first time.
(iii) Investigating the approaches and performance metrics regarding intelligent fault prognosis of rotating machines for the first
time.
(iv) Examining the literature considering the pursued fault analysis strategies for IFDP of rotating machines including specific faults
and compound faults for the first time.
(v) Reviewing the literature regarding the relationship between the developed intelligent models and compound faults for the first
time.
(vi) Addressing the challenges and the future directions related to IFDP of rotating machines.

This review discusses the following research questions (RQ):

RQ1. Which rotating machine components are mostly considered in IFDP with ML?
RQ2. What kind of faults are detected, located, identified, and subjected to prognosis with ML?
RQ3. What are the most common data sources used in the field of IFDP of rotating machinery?
RQ4. What fusion techniques are considered for different data sources in the IFDP of rotating machinery?
RQ5. What feature extraction techniques are adopted for the IFDP of rotating machinery?
RQ6. What are the most common machine learning approaches used in the field of IFDP of rotating machinery?
RQ7. What are the most common machine learning techniques used in the field of IFDP of rotating machinery?
RQ8. Which approach and performance metrics are commonly considered to assess ML approaches built for the IFDP of rotating
machinery?

Fig. 1. Systematic literature review process.

3
O. Das et al. Heliyon 9 (2023) e17584

RQ9. What are the most suitable maintenance strategies for IFDP of rotating machinery?
RQ10. What are the main challenges in the field of IFDP of rotating machinery?
RQ11. What are the possible research directions for IFDP of rotating machinery?
The rest of the paper is organized as follows. Section 2 presents the considered research methodology including the selection
procedure of the reviewed studies, publication trends over time, and an overview of the process of the rotating machinery fault
analysis. The number of the most relevant studies published in the last decade is presented. Section 3 introduces the maintenance
strategies including their conceptual definitions and their suitability for intelligent fault diagnosis and prognosis procedures. Section 4
proposed fault analysis strategies by explaining the differences among fault detection, fault isolation, fault diagnosis, and fault
prognosis. Besides, the techniques regarding each fault analysis strategy are briefly explained. Section 5 comprises the component-wise
IFDP procedures of rotating machines where the IFDP studies are reviewed based on the considered fault type and component. In
addition, the complexity of the faults for each component in terms of frequency spectrums. Section 6 presents the IFDP of rotating
machine literature from the perspective of data sources, data fusion, feature extraction techniques, and machine learning concepts.
Section 7 summarizes the study and points out the challenges that existed in the field of IFDP of rotating machines. Finally, Section 8
presents the research directions in a wide perspective for future works related to this field.

2. Research methodology

2.1. Systematic literature review methodology

Fig. 1 shows the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) flow diagram that illustrates the
identification, screening, eligibility, and inclusion of studies. The initial search on Web of Science (WOS) and SCOPUS databases
returned a large number of papers, totaling 8775 and 2034, respectively. We selected these databases since they are widely regarded as
the standard and greatest authority for scientific research.
Thereafter, the retrieved documents were filtered by years (2013–2023) to focus on state-of-the-art studies. After excluding con­
ference papers, notes, editorial materials, meeting abstracts, and retracted papers, a total of 5381 and 842 peer-reviewed journal
articles and book chapters were obtained for the review since they usually present a greater level of detail. Following this, a small
minority of papers were specifically excluded since they were not in English. After the abstract screening, approximately 1000 studies
were deemed eligible for the scope of this review. The papers that do not present sufficient information about the study and duplicates
were excluded resulting in 218 articles. The rest of the papers were scanned for relevancy and checked whether they will contribute
valuable knowledge to this review. Furthermore, studies for which full texts were not available were excluded. This resulted in 128
research papers suitable for full-text reading. Upon the full-text reading, 18 papers seemed ineligible due to the lack of considerable
scientific contribution and detailed information about the research questions, and the rest of the 115 papers were included to be used in
the assessment of the review questions.

2.2. Publication volumes

Fig. 2 shows the publication trend in the field of fault analysis over the last decade in terms of the cumulative number of papers
published. As seen in Fig. 2, the number of studies geometrically increases, and that reflects the popularity of fault diagnosis and
prognosis of rotating machines. For example, while approximately 200 relevant papers were published in 2014, that number was
almost 1640 in 2022, which is a significant increase. When analyzing the proportion of publications for each year, continuous growth
can be observed from 2020 to 2022, with high proportions of 15.9%, 20.24%, and 21.45%, respectively. Thus, fault analysis has
recently received increasing attention worldwide due to its extensive applications and has become a field of research increasingly hot
and popular. In the forthcoming years, machine learning will become even more significant due to the presence of a huge volume of
fault analysis-related data.

Fig. 2. The publication trend in the field over time.

4
O. Das et al. Heliyon 9 (2023) e17584

2.3. The process of rotating machinery fault analysis

As shown in Fig. 3, the process of rotating machinery fault analysis consists of the sequence of individual parts, including data
acquisition, data preparation, learning from data, evaluation, prediction, presentation of the results, and decision-making. The data
acquisition phase highly depends on the machinery component, the type of sensor, and the way the data is transferred and stored.
Common mechanical components that have been considered in fault diagnosis systems include bearing, gear, shaft, belt-pulley, and
induction motors, especially rotor and stator. The type of sensor is selected with an awareness of machine failure modes and the related
warning indicators. Typical warning signals in rotating machinery are usually handled with acoustic, vibration, temperature, and
pressure sensors, as well as visual devices. Collected data is transferred via a communication technology (i.e., Wi-Fi, Bluetooth) to the
server and stored in a database in the cloud environment. Signal records are a collection of time-indexed data instances acquired across
a time period. Since this data contains raw values, its current format does not directly give the essence of the data and does not provide
meaningful information about the faults. Therefore, signal processing is usually required to prepare data for further analysis and
process. A typical data preparation pipeline includes normalization, filtering, aggregating, linearization, missing data elimination,
feature extraction, and feature selection. The training phase includes the use of machine learning, ensemble learning, and deep
learning techniques to develop the model that will be used for fault prediction. In the evaluation phase, the trained model is tested in its
ability to predict the output by using various metrics such as accuracy, precision, recall, f-measure, confusion matrix, and area under
the curve of receiver operating characteristics (AUCROC). In the prediction phase, probable machinery fault for previously unseen
input data is produced by using the constructed model. Following that, the predicted output is presented to the users via a web/mobile
application to provide guidance to the decision-makers. Finally, in the decision-making phase, the predicted result is utilized by
managers for performing activities related to fault diagnosis, fault isolation, fault prognosis, fault detection, and exclusions.

3. Maintenance strategies

The word “maintenance” is referred to the procedure taken to preserve and repair a machine, component, structure, or building to
its initial condition or to state where an effective functionality is obtained. The ultimate goal of maintenance is to operate the machine
as expected, and prevent faults, breakdowns, or other problems that may endanger its reliability or performance. The maintenance
procedure comprises various tasks including inspection, lubrication, cleaning, replacements, and repair. Because maintenance may be
an expensive and time-consuming procedure, experts developed numerous maintenance strategies to lower such drawbacks.
Some significant strategies are preventive maintenance (PnM), predictive maintenance (PdM), corrective maintenance (CM),

Fig. 3. The overall process for rotating machinery fault analysis.

5
O. Das et al. Heliyon 9 (2023) e17584

condition-based maintenance (CBM), reliability-centered maintenance (RCM), overhaul maintenance (OM), and run-to-failure
maintenance (RFM). PnM focuses on regular inspections, cleaning, replacing, and other related tasks periodically to prevent failure
[63]. PdM uses information obtained through sensors and other equipment to monitor the machinery and estimate the period when the
replacement or repair is needed [64]. CM is based on fixing the part that is already malfunctioning or broken [63,65]. CBM is similar to
PdM as it also monitors the conditions of the components of the machine. However, CBM aims to give the exact time when the faulty
part is needed to be replaced or fixed whereas PdM gives a future prediction regarding the time that the maintenance should start [66].
RCM focuses on the most essential parts of a machine and constitutes a maintenance schedule and strategy around those parts to
guarantee their reliability [66]. OM is about taking the machine out of the order and conducting maintenance in which the machine
parts are completely disassembled, inspected, cleaned, and reassembled [67]. Finally, the RFM strategy is based on zero proactive
maintenance and waiting for the complete failure of the components [68].
Choosing the most suitable maintenance strategy is another challenging task. The selection of the correct method depends on the
type, age, condition, fault type, and fault severity of the system or its component. In addition, the availability of replacement parts,
safety and regulatory requirements, maintenance costs, skills of the maintenance technicians, and resources needed for maintenance
are other essential parameters to adopting the appropriate maintenance strategy. In the industry, PdM is considered one of the most
common strategies since it is cost-effective and reduces downtime by conducting a routine periodic check, which significantly lowers
the possibility of unexpected repairs or breakdowns. PdM is also one of the most suitable approaches for Industry 4.0 [69,70] because
PdM is a data-driven technique that can be easily implemented with the aspects of Industry 4.0 such as IIoT, AI, ML, and sensor
technologies. Another appropriate strategy may be CBM, which also is a data-driven approach just like PdM. Because both strategies
require data that helps to monitor the condition of the system/machine or its component, it is essential to understand fault diagnosis
and prognosis procedures in monitoring the machines.

4. Machinery fault analysis strategies

Fault diagnosis (FD) and fault prognosis (FP) procedures refer to detecting, identifying, and assessing the faults or issues related to a
machine to comprehend its current and future status. The main purpose of FD is to reveal the root cause of the fault and therefore,
provide one to take the necessary actions to correct the malfunctioning. FD procedure involves detecting the condition of the machine,
locating the faulty component, and finding the fault type if the machine is malfunctioning. These sequential sub-procedures are called
fault detection (FDE), fault isolation or localization (FI), and fault identification (FID). Machine Learning-based fault diagnosis is
generally conducted in six stages called data acquisition, data processing, feature extraction, model training, model assessment, and
decision support. In data acquisition, signal or visual data is collected through sensors or cameras. Afterwards, the data is processed by
signal or image processing techniques in the data processing stage to eliminate noise or any disruptive matter that adversely affects the
characteristics of the data. In the feature extraction procedure, the time, frequency, time-frequency, or any other features of the
processed or unprocessed data is extracted. If the model is untrained, then the feature sets are needed to be labeled for supervised
learning-based and semi-supervised learning-based approaches. Later, the parameters of the intelligent model are tuned and the model
is trained by considering the features as input and the labels as output. The trained model then is assessed by feeding it with unseen
data. After the reliance and the performance of the model are proven, it may be implemented in a real-world setting as a decision-
support mechanism for fault diagnosis. Unsupervised methods pursue a different strategy where the input data is unlabeled and the
decision is made based on clustering. Regarding classification-based deep learning methods, it is not usually mandatory to perform the
hand-crafted feature extraction stage since the model is generally capable of extracting the features automatically.
FDE can be expressed as the disclosure of an abnormal situation that may cause the machine not to work properly or even lead to an
outage. This is also comprehended as the first step of the fault diagnosis procedure since the malfunction is identified yet, the kind,
source, and severity of the malfunction are unknown. Fault detection can be conducted in various ways including visual or auditory
inspections, monitoring the sensor data, statistical pattern control, and machine learning-based methods. Visual or auditory in­
spections are based on mostly work-experience subjective examinations in which the origin of a fault cannot be detected. Although can
be related to fault detection, it may require expertise, training, and time for an effective inspection. Monitoring the sensor data
provides a more in-depth approach to fault detection since the abnormality is detected by machine data such as vibration, temperature,
pressure, etc. The sensor data is generally reliable and provides sufficient information about whether the machine is healthy or not.
However, reading and interpreting the sensor data requires expertise. In addition, the data have to be collected from the healthy
sensors that are properly mounted. Statistical pattern control employs statistical methods to find any changes in the trends or patterns
in the sensor data to detect any abnormalities in the machine. This technique may reflect the issue of the machine in a simpler way than
monitoring solely the sensor data. However, it still requires expertise due to the selection of the statistical metrics and reading such
statistical data. Besides, early-stage abnormalities can easily go unnoticed since they may slightly impact the statistical patterns of a
machine that operates normally. Machine learning-based methods utilize the sensor data or its characteristics obtained throughout
image processing, signal processing, statistical feature extraction, etc. as input and pass them through a specified machine learning
algorithm to find out whether the machine is healthy or not. Therefore, early-stage abnormalities may be detected in a fast and
effective way. Although selecting an appropriate machine learning method and processing the data may cause a struggle, such
techniques have the potential to provide the best and most precise results.
FI is expressed as the detection of a particular subsystem or component within a machine that causes an abnormality or failure.
Following the fault detection procedure, it is one of the most essential steps in the fault diagnosis procedure since the exact location and
component of the malfunctioning part is determined. Fault isolation can be performed by testing each component, tracing the output
signal, and machine learning-based fault isolation methods. The component testing procedure is testing the functionality of the

6
O. Das et al. Heliyon 9 (2023) e17584

components to validate if they are working properly. As it can be estimated, testing each component can be time-consuming and
expensive, which is undesirable in fault diagnosis procedures. To lower such consumption, substitution testing may be adopted in
which the known healthy parts are swapped out with the existing components to observe whether the problem is solved or not.
However, it is still an expensive and time taking procedure that is not generally preferred. Signal tracing is a testing method in which a
signal is given during the test and it is traced to find the location where the signal does not behave normally. By employing the signal
tracing technique the cost is significantly lowered and the malfunctioning component is usually located precisely. However, it requires
expertise and time, especially for complex machines having numerous components. Machine learning-based fault isolation uses the
visual or signal data to identify the abnormality and the abnormal component, including its location if the data acquisition devices are
placed properly. As mentioned before, machine learning-based techniques require an appropriate algorithm selection, choosing and
employing image or signal processing techniques, which are challenging tasks and require expertise.
FID refers to identifying the type of fault in a malfunctioning machine or component. This phase comprises the description of the
fault instead of referring to the erroneous section as “faulty” only. Hence, it is a critical stage of the FD since the root cause of a fault is
determined. By doing so, the necessary precautions and actions can be taken by the experts. FID can be employed by processing the
vibration, acoustic, and temperature data of the machine since each fault type has a characteristic impact on such data. These data may
be interpreted by the experts or provided to an AI-based advanced algorithm for an automated and effective FID procedure.
Fault Prognosis (FP) is known as predicting the severity of the faults that existed in a component of a machine that can be used for
estimating the remaining useful life (RUL) or remaining life (RL) of the component. The procedure is conducted on the question of
“How bad is the fault?”.
Thanks to FP, a possible catastrophic impending failure may be prevented. By providing an early warning, it is also possible to
schedule maintenance or repairs, which may significantly reduce the downtime of the machine during maintenance. To conduct FP,
the same methods as those implemented for FD are used. Apart from the FD, FP may also require a perspective of regression than
classification. Therefore, it is necessary to focus more on the change in the trends of the measured data.
Developing effective fault diagnosis and prognosis approaches is vital in scheduling maintenance. Since each component has a
unique impact on a rotating machine, it is essential to understand each component. Various kinds of faults may occur in the com­
ponents specific to the rotating machines such as bearings, gear, induction motor, belt-drive mechanism, shafts, fans, turbine blades,
etc. Hence, developing component-specific and fault-specific diagnosis and prognosis strategies or methods is a key procedure to take
in time and correct actions.

5. Fault diagnosis and prognosis in rotating machinery

A rotating machine refers to a system where the components rotate around an axis to generate mechanical energy for numerous
purposes. Such machines are utilized in pumping fluids, turbines, generators, fans, and compressors. The fault diagnosis and prognosis
procedures regarding a rotating machine are generally conducted on its components. Fig. 4 shows the parts of the rotating machinery

Fig. 4. The components and faults in the rotating machine.

7
O. Das et al. Heliyon 9 (2023) e17584

including the fault types that may occur.

In the last ten years, numerous studies related to intelligent fault diagnosis and prognosis (IFDP) of rotating machines are conducted
and published in the literature. Fig. 5 shows the percentile distribution regarding the considered published studies and the considered
machine component.
It is seen in Fig. 5 that the majority of the studies are considered bearing faults [71–75] followed by rotor faults [76–80], gear faults
[81–85], shaft faults [21,46,71,72,86], stator [7,25,87], and other component faults [28,88–90]. The pie chart shown in Fig. 5
comprises studies that consider the faults of a single component and the faults of multi-components. As seen in Fig. 6, most studies
developed an intelligent model to find the fault types of a single component. However, there are a considerable amount of studies that
also considered multi-component faults to build IFDP models for rotating machines [91–101]. Fig. 7 shows the percentiles of the
considered fault analysis and intelligent model development strategies regarding FD, FDE, FP, and FI. Among those studies, it is seen
that the majority pursued to constitute an FD-based intelligent model, followed by FDE, FP, and FI models.

5.1. Bearing faults

Bearings are the mechanical parts in rotating machines that are used to guide and support the moving components. They have a key
role in an effective process by reducing friction, supporting the rotating components, and providing a consistent and smooth move­
ment. Faults in bearing may occur due to various reasons including excessive loading, poor bearing selection, faulty installation, lack of
maintenance, excessive vibration, etc. The most common bearing faults are related to the inner race, rolling element, cage, and outer
race of the bearing. Besides, assembly and lubrication issues may cause wear, excessive vibration, and oil leakage [102].
The faults manifest themselves in characteristic frequencies. Ball pass frequency outer (BPFO) refers to the outer race faults
whereas ball pass frequency inner (BPFI) is an indicator for inner race faults. Regarding the rolling element faults, ball spin frequency
(BSF) is used. Finally, the cage faults are referred to as fundamental train frequency (FTF). If the bearings are not replaced when those
frequencies appeared, the following stage will affect the 1× RPM amplitude and make the BFPO, BPFI, BSF, and FTF disappear. At
frequencies close to and higher than 30 k Hz, random vibrations are observed. These characteristic fault frequencies are calculated as
[102].
( )
n dB rpm
BPFO = 1− cos θ (1)
2 dP 60
( )
n dB rpm
BPFI = 1+ cos θ (2)
2 dP 60
( ( )2 )
dP dB rpm
BSF = 1− 2
cos θ (3)
2dB dP 60

( )
1 dB rpm
FTF = 1− cos θ (4)
2 dP 60

Fig. 5. The percentile distribution of the considered rotating machine.

8
O. Das et al. Heliyon 9 (2023) e17584

Fig. 6. The ratio of IFDP methods build for single component faults and multi-component faults.

where n is the number of balls, θ is the contact angle, dB and dP are the ball diameter and pitch diameter, respectively. The frequency
spectrums of those characteristic frequencies are shown in Fig. 8 [102].
Regarding the intelligent fault diagnosis and prognosis (IFDP) of bearing faults, numerous studies are presented in the last decade.
Table 1 refers to the studies related to the intelligent bearing fault diagnosis and prognosis. The abbreviations of the fault diagnosis and
prognosis procedures in Table 1 are as follows. FD: Fault Diagnosis, FDE: Fault Detection, FI: Fault Isolation (Fault Localization), FP:
Fault Prognosis. The abbreviations of the bearing fault types in Table 1 are; RF: Rolling Element Fault, IR: Inner Race Fault, OR: Outer
Race Fault, CAF: Cage Fault, Comp.: Compound Faults, Oth.: Other Faults, (c): Classification. Fig. 9 shows the percentile distribution of
the bearing fault types investigated within the scope of the studies given in Table 1.
As seen from Fig. 9, the majority of those studies are focused on the rolling faults [103–106], inner race [107–118], and outer race
defects [119–127]. On the other hand, cage faults are also examined by the researchers, but not much as the ball, inner race, and outer
race faults. The remaining faults such as house eccentricity or lubrication issues constitute the minority of the literature. Another
significant condition is the compound fault situation where multiple different faults take place at the same time. The literature
comprises a few studies related to this condition regarding bearing faults. Besides, some studies examined compound faults that
include the same or different component faults in addition to the bearing faults [10,40,97,106,107,111].
The studies given in Table 1 mostly used publicly available benchmark datasets. Most of the studies considered benchmark datasets,
especially those in which ball, inner race, and outer race faults include. In addition, the dataset also comprises different fault locations
regarding the outer race faults. The measurements of the dataset are conducted considering four motor loads. Apart from this dataset,
various studies use the data from their setup or the industry [6,7,16,22,23,32,71,74,87,92,94,95,99,107,113,114,116,124].
Some studies related to the intelligent bearing fault diagnosis and prognosis are presented as follows. Li et al. [93] proposed a
support vector classification-based approach to identify bearing faults. For this purpose, they considered two gearbox setups where
different gear and bearing faults exist. Regarding the bearing faults, the setup includes IR, OR, RF, and House Eccentricity (HC) faults.
Although they adopted an FD procedure, the model performance on the fault identification is not clear because only overall model
accuracy is considered. Shoaib et al. [41] employed a deep learning-based approach for bearing FD and FP. They considered RF, IR,
and OR faults. As performance metrics, only overall model accuracy is considered. Although overall accuracy is an essential indicator
for model performance, it is not sufficient for FD procedure because overall accuracy does not include the fault-specific performance of

9
O. Das et al. Heliyon 9 (2023) e17584

Fig. 7. The percentile distribution of the considered strategies regarding IFDP of rotating machines.

a proposed approach, which is essential especially if multiple types of faults are asked to be found. The study also comprises a
classification-based FP approach with an accuracy higher than 96.50% regarding each fault severity. Zhang et al. [48] used a
presumption-based Naive Bayes approach for the FD and FI of bearings. They employed RF, IR, and OR faults of bearings. They
considered both accuracy and confusion matrices for FD and FI. The usage of the confusion matrix enabled them to provide the
fault-specific performance of their proposed model for the FD procedure. Regarding FI, they considered OR faults that exist at the
center at @6:00, orthogonal at @3:00, and opposite at @12:00 locations of the outer race of the bearing. The FI procedure is conducted
in combination with the FD procedure. In other words, they considered each location of OR fault as a distinct labeled case and built the
model with that assumption. Hence, instead of four classes including normal case (NC), IR, RF, OR; six cases (NC, IR, RF, OR@3:00,
OR@6:00, and OR@12:00) are considered. The overall accuracy of their proposed approach is 99.17% whereas the fault-specific
prediction accuracy differs between 97.50% and 100.00%. Sanchez et al. [97] considered five different datasets for the FD of bear­
ings. Four bearing faults, RF, IR, OR, and HC are predicted by two machine learning-based techniques. To measure the FD performance
of their proposed approach, they only considered overall model accuracy for each wavelet and feature type. The highest overall FD
accuracy is obtained by 83.0%. Guo et al. [38] used a deep learning-based approach for FD and FI of bearings. They considered RF, IR,
and OR faults that existed on a simulator setup. According to their experimental results, they only considered only overall model
accuracy values for both FD and FI. They obtained an overall accuracy of 96.04% for FD and 99.93% for FI of bearings. Choudary et al.
[32] employed a deep learning-based method for bearing FD of rotating machines. They considered bearings having RF, IR, OR, CAF,
and lack of lubrication (LB) faults. They presented the model performance of their proposed approach by using overall and
fault-specific accuracy, precision, recall, F-score values, and confusion matrix. According to the experimental results, their proposed
approach makes predictions with an overall accuracy of 99.80%. Considering the fault-specific performance of the proposed model, the
accuracy values change between 97.05% and 100.00%. Presenting precision, recall, and F1 scores indicate the sensitivity and
robustness of their proposed approach. These are also essential metrics that illustrate how well the model makes predictions when it
deals with related unseen data.
As seen in Table 1, most of the studies focused on FD rather than FDE, FI, or FP. In general, a complete FD procedure comprises FDE
and FI. Hence, all FD studies are also related to FDE and FI in some way. On the other hand, FP is about the fault severity and the
remaining useful life (RUL) of the component. There are only a few studies that are interested in FP for bearings [29,39,41,111,125].
Besides, most of those studies adopted a classification-based strategy for FP, which may give a fallacious idea regarding the perfor­
mance of an intelligent model. For FP procedures, confusion matrices should be also considered. Acquiring a high accuracy value for
each severity is important, but considering a classification-based FP procedure may not reflect the true performance of a proposed
model since severity values can vary infinitely unlike labeled fault types. For instance, consider a dataset that comprises three IR crack
lengths, 0.01 mm, 0.02 mm, and 0.03 mm. Building an intelligent fault prognosis model will be labeled with those values if a
classification-based strategy is employed. An accurate model will predict the unseen data that includes 0.01 mm, 0.02 mm, or 0.03 mm
crack length. On the other hand, it is unknown how it will classify the data which includes 0.015 mm, 0.025 mm, etc. The model may

10
O. Das et al. Heliyon 9 (2023) e17584

Fig. 8. The frequency spectrum of (a) outer ring fault, (b) inner ring fault, (c) ball fault, and (d) cage fault.

predict the crack length of 0.015 mm as 0.01 mm or 0.02 mm. Such situations make the classification-based FP strategies generally
unreliable and inappropriate for severity prediction and/or RUL. Regarding performance metrics, numerous studies focused on only
accuracy [31,37,39,41,45,47,50,73,104,112,119] and discarded the other metrics such as confusion matrix, precision, and recall. The
omission of the confusion matrix in the FD procedure makes it unable to interpret fault-specific model performance. Hence, the
prediction ability of a proposed approach remains unknown for a specific type of fault, which endangers the reliability of a model. The
confusion matrix also enables the evaluation of other metrics such as precision and recall which are the indicators of the sensitivity and
robustness of an intelligent model. Those metrics should also be presented regarding the overall and fault-specific performance of the
intelligent FD technique. Considering precision and recall illustrates the performance of the model for unseen related data. Some
studies considered confusion matrices and/or precision-recall [16,26,47,51,55,129], yet they constitute the minority of the literature.
According to the studies given in Tables 1 and it is seen that regression-based strategy for FP procedures is lacking in the current
literature. In addition, confusion matrix, precision, and recall metrics are mostly not considered to reflect the fault-specific perfor­
mance, sensitivity, and robustness of a proposed intelligent FD method. Regarding the fault type percentiles shown in Fig. 9, it is seen
that the majority of the studies focused on only four types of bearing faults whereas other faults such as house eccentricity, aging,
journal bearing faults, oil issues, wear, and clearance problems are barely investigated [6,22,31,72,93,123]. Besides, compound faults
are rarely considered [10,40,74,106,107], yet they are essential in FD and FP procedures of rotating machines.

5.2. Gear faults

Gears are one of the most critical parts of rotating machines because they transmit power and motion between rotating compo­
nents. They are shaped and sized differently in accordance with the needs. Gear faults may occur due to excessive loading, faulty
installation, improper or lack of lubrication, fatigue, corrosion, or contamination by dust, dirt, or any material that can harm. Some
gear faults are tooth wear, broken or cracked tooth, chipped tooth, gear misalignment, and eccentricity. Just like bearing faults, gear
faults show themselves uniquely in the frequency spectrum. Fig. 10 shows the frequency spectrum of some gear faults [102].
Diagnosis and prognosis of the gear faults attracted researchers from the past to the present. In the scope of the IFDP, various

11
O. Das et al. Heliyon 9 (2023) e17584

Table 1
Intelligent fault diagnosis and prognosis studies related to bearing faults.
Studies Year Procedures Fault Types

FD FDE FI FP RF IR OR CAF Comp./Oth.

[16] 2013 X X X
[71] 2014 X X
[6] 2013 X Ageing
[124] 2014 X X(c) X X
[39] 2014 X X(c) X X X
[91] 2014 X X(c) X X X
[92] 2015 X X(c) X X X
[93] 2015 X X X X House Eccentricity
[123] 2015 X X X X Retainer Fault
[72] 2015 X X X Wear, Lack of Bearing
[94] 2016 X X Lack of Lubrication
[7] 2016 X X X
[87] 2016 X X
[41] 2017 X X(c) X X X
[95] 2017 X X X X
[118] 2017 X X X X
[48] 2018 X X X X X
[74] 2018 X X X X IR + OR
[97] 2018 X X X X House Eccentricity
[100] 2018 X X X X X
[47] 2019 X X X
[23] 2019 X X X X
[126] 2019 X X X X
[127] 2019 X X(c) X X X
[22] 2019 X Wear
[40] 2019 X X X Gear + Bearing, Gear + Bearing + Looseness
[38] 2020 X X X X X
[109] 2020 X X X X X
[110] 2020 X X X X
[111] 2020 X X(c) X X X IR + OR, RF + OR
[114] 2020 X X X X X
[115] 2020 X X X X
[37] 2021 X X X X
[32] 2021 X X X X X LB
[35] 2021 X X X X
[106] 2021 X X X X BF + IR + OR
[107] 2021 X X(c) X X IR + OR
[108] 2022 X X X X X
[10] 2022 X X X IR + OR
[45] 2022 X X X X X
[103] 2022 X X X X
[101] 2022 X X X
[75] 2022 X X X X
[105] 2022 X X X X
[15] 2022 X X X X X X
[11] 2023 X X X X Rotor + Bearing
[125] 2023 X X X

approaches are proposed regarding different gear fault types. Table 2 refers to the studies related to the intelligent bearing fault
diagnosis and prognosis. The abbreviations of the gearing fault types in Table 2 are as follows. CT: Chipped Tooth, IR: Inner Race Fault,
BT: Broken Tooth, MT: Missed Tooth, PT: Pitting, CRC: Crack, WT: Worn Tooth, Comp.: Compound Faults, Oth.: Other Faults, (c):
Classification.
Fig. 11 shows the percentile distribution of the bearing fault types investigated within the scope of the studies given in Table 2.
As seen in Fig. 11, an almost equal distribution is observed considering CT, WT, BT, CRC, and PT faults [128–131]. The MT fault is
slightly less investigated than those faults. Gear faults like chafing, surface wear, gnash, eccentricity, and spalling are rarely examined
[9,81,83,84,128]. All these faults constitute only 4% of the total studies. The compound faults among gear faults are also barely
investigated. Among all the studies taken into account, only one study measured the performance of their proposed approach on
compound faults considering gear compound faults and gear-other component compound faults [40].
As seen in Table 2, most of the studies focused on the FD procedure rather than FDE, FI, or FP. On the other hand, conducting a
complete FD also means performing FDE and FI. Only about 17% of the studies are focused on classification-based FP procedures. Some
studies related to IFDP regarding gear faults are presented as follows. Chen et al. [92] adopted a deep learning-based approach for FD
and FP of the gearbox where various gear and bearing faults exist. Regarding the gearing faults, face wear (FW), chafing (CH), BT, PT,
and CT. They evaluated their proposed approach by considering accuracy and confusion matrix metrics. In addition, mean and median
statistics of the accuracy values are obtained from the confusion matrix. FD and FP procedures are considered as a whole, meaning that

12
O. Das et al. Heliyon 9 (2023) e17584

Fig. 9. Percentile distribution of the considered bearing fault types.

Fig. 10. The frequency spectrum of (a) healthy gear, (b) tooth wear damage, (c) cracked/broken tooth, and (d) gear misalignment.

the same faults having different severity values are treated as labeled unique fault classes. The mean overall accuracy of the proposed
model is 96.8%. Regarding fault-specific results, the accuracy values differ between 91.4% and 98.9%. Although precision and recall
metrics are not given, they can be evaluated through a confusion matrix. Chemseddine et al. [42] used a neural network approach for
gear FD of a gearbox. The gears in the gearbox contain CRC, CT, MT, and WT faults. The performance of their model is measured by

13
O. Das et al. Heliyon 9 (2023) e17584

Table 2
Intelligent fault diagnosis and prognosis studies related to gear faults.
Studies Year Procedures Fault Types

FD FDE FI FP CT BT MT PT CRC WT Comp./Oth.

[91] 2014 X X X X
(c)
[92] 2015 X X X X X Chafing, Face Wear
(c)
[93] 2015 x X X X X Chaffing
[81] 2015 X X X Face Wear, Misallignment
[24] 2016 X X X X X Chafing
[17] 2016 X X X
[95] 2017 X X X
(c)
[96] 2017 X Unknown
[43] 2017 X X X X X Chafing, Weak Root
[82] 2017 x X X
[74] 2018 X X X
(c)
[97] 2018 X X X X Chafing, Misalignment
[98] 2018 X X X
[42] 2018 X X X X X
[21] 2018 X X X Eccentricity
[99] 2018 X X X Surface Fatigue
[131] 2018 X X X
[24] 2019 X X X
(c)
[130] 2019 X X X X X
[40] 2019 X X X X BT + CT,Gear + Looseness, Gear + Looseness + Bearing, Gear +
Bearing
[129] 2020 X X X X X
[113] 2020 X X
[8] 2020 X Unknown
[83] 2020 X X X X Chafing
[9] 2020 X X X X Eccentricity, Abrasion
[107] 2021 X X X
(c)
[13] 2021 X Unknown
[27] 2021 X X X X X
[51] 2021 X X X X
[128] 2021 X X X X Gnash
[85] 2021 X X
[101] 2022 X X
[75] 2022 X X
[84] 2022 X X Spalling

Fig. 11. Percentile distribution of the considered gear fault types.

14
O. Das et al. Heliyon 9 (2023) e17584

(caption on next page)

15
O. Das et al. Heliyon 9 (2023) e17584

Fig. 12. The frequency spectrum of (a) unbalance fault (force, dynamic, couple), (b) unbalance fault (overhung rotor), (c) rub-impact fault, (d)
horizontal misalignment, (e) angular misalignment, (e) eccentric rotor, and (f) bent shaft.

considering only the overall model accuracy metric regarding the five sequences that they performed. They obtained an overall ac­
curacy of 99.36% in classifying gear faults. Azamfar et al. [9] proposed a deep learning-based technique for gear FD. For this purpose,
they used the data obtained from a test simulator where eccentricity (ECC), PT, MT, CT, and abrasion (ABR) faults exist. The model
performance is reflected by the accuracy and confusion matrix. The study also comprises FP procedures regarding ABR fault. Similar to
various studies, a classification-based approach is pursued by considering different ABR severity values as unique fault classes. Shi
et al. [84] considered the sun-planet gear mechanism for FD of gear faults. They used a deep learning-based approach for gears having
CRC and spalling (SP) faults. They performed a complete FD that includes fault detection, fault type identification, and fault location.
In addition, they also considered finding the fault direction. To assess the model, they considered both overall model accuracy and
confusion matrix metrics. Hence, fault-specific model performance can be evaluated. On the other hand, the fault-specific results are
not summarized, yet they can be evaluated from the confusion matrices.
As mentioned in bearing faults, the studies related to the FP of gear faults pursued a classification-based approach [91,107,127].
Although classification-based FP may give an idea about the performance of the proposed approach, it may misdirect one as the
developed model receives an input that indicates a severity value close to the mean value of two severity classes introduced to the
model. Regarding FD procedure, various studies generally focused on overall model accuracy [8,21,24,27,40,42,43,97,129], and a few
considered confusion matrix, precision, and recall [85,101,107]. The gear faults considered in related studies are diverse since
numerous faulty gear cases are examined. On the other hand, the literature lacks compound gear faults or compound faults including
gear faults and other rotating machine faults.

5.3. Rotor faults

The rotor is used for producing the rotational motion that drives or is driven by a machine. It is surrounded by a stator, which is a
stationary part that provides electric current to rotate the rotor. Rotor faults occur due to several reasons including manufacturing and

Table 3
Intelligent fault diagnosis and prognosis studies related to rotor faults.
Studies Year Procedures Fault Types

FD FDE FI FP MA UB RI BR Comp./Oth.

[136] 2013 X X(c) X

[25] 2013 X ECC.
[6] 2013 X X
[30] 2013 X X(c) X
[71] 2014 X BWR
[97] 2014 X X(c) CR
[76] 2015 X X X X
[138] 2015 X X X X CR
[94] 2016 X X
[137] 2016 X X(c) X
[7] 2016 X X X
[87] 2016 X X X BWR
[77] 2017 X X X
[140] 2017 X X(c) X
[96] 2017 X X
[135] 2018 X X X X
[26] 2018 X X
[78] 2018 X X
[100] 2018 X X X
[49] 2019 X DICB
[133] 2019 X X X X OW
[134] 2019 X X
[22] 2019 X DR
[40] 2019 X X X Rotor + Gear Faults
[79] 2019 X X
[114] 2020 X X Rotor + Bearing Faults
[115] 2020 X DR
[12] 2021 X X(c) X
[132] 2021 X X X
[34] 2021 X X
[33] 2022 X X
[141] 2022 X X(r) RWF
[80] 2022 X X
[11] 2023 X X X Rotor + Bearing Faults
[139] 2023 X X(c) CR

16
O. Das et al. Heliyon 9 (2023) e17584

maintenance issues, exposure to environmental factors, excessive loading, faulty installation, aging, and operating conditions. The
types of rotor faults are broken rotor bars, unbalanced rotor, misaligned rotor bars, rub-impact faults, corrosion, and electrical faults.
Each kind of fault has a unique impact on the rotating machine data that includes vibration, temperature, pressure, current, etc. The
illustration of the impact of the rotor-related faults on the frequency spectrum is shown in Fig. 12 [102].
Table 3 presents the studies related to the intelligent rotor fault diagnosis and prognosis of rotating machines. The abbreviations of
the rotor fault types in Table 3 are; MA: Misalignment, UB: Unbalance, RI: Rub-Impact, BR: Broken Rotor, BWR: Bowed Rotor, ECC.:
Eccentricity, DICB: Damaged Impeller and Cover Board, OW: Oil Whirl, DR: Damaged Rotor, RWF: Rotor Winding Fault, Comp.: Compound
Faults, Oth.: Other Faults, (c): Classification. Fig. 13 shows the percentile distribution of the bearing fault types investigated within the
scope of the studies given in Table 3.
It is seen from Table 3 that compared with the gear and bearing faults, there are more studies in percentile for rotors that focused on
the FP procedures. Besides, although numerous studies comprise FD of rotor faults, there are a considerable amount of FDE studies. The
FDE studies are based on fault detection and do not aim to identify the type of faults. Since there is only a healthy and faulty condition
of the component, the importance of the considered defects is only related to the level of distinguishability between faulty and healthy
states. Regarding FDE procedures, RI faults are considered more than the other rotor faults. In Table 3, some studies are referred to as
FD studies that have a single rotor fault. This is because these studies are focused on not only rotor faults but also other components and
assessed their proposed approaches by considering all kinds of faults. Some studies related to the IFDP of rotors are presented as
follows. Lu et al. [79] employed a machine learning-based method for the FD of rotor faults. They considered UB, MA, and RI faults that
existed in a rotating machinery setup. Although they used accuracy as the sole performance metric, they also present fault-specific
results regarding the assessment of their proposed approach. They obtained a perfect score for all kinds of conditions and fault
types of the rotor. Zgarni et al. [78] performed an intelligent FDE procedure regarding the BR fault of the rotor. To evaluate the
performance of their proposed method, they considered ROC curves, F1 score, and false faction classification (FFC) score. Obtaining an
F1 score close to 1 indicates a precise and sensitive model. FFC score is another metric that indicates the average false predictions made
by the model. Hence, a lower FFC means more accurate results. According to their experimental results, the highest metrics are ob­
tained as 1.0 for F1 and 0.0 for FFC. Since intelligent FDE procedures are binary classification problems, a confusion matrix may not be
essentially required. Xiang et al. [139] adopted a deep transfer learning-based approach for rotor FDE and FP where cracked rotor fault
(CF) having two severity values is considered. Although it is an FDE problem, the confusion matrix is considered a performance metric
in addition to overall accuracy. The inclusion of the confusion matrix is because of the FP procedure where two crack severities are
identified in a classification-based approach. According to their experimental results, they obtained an average model accuracy of
92.67% and perfectly distinguish one crack from another.
As seen in Table 3, there are plenty of studies related to the FP of rotor faults. However, similar to bearing and gear faults, almost all
of those are classification-based approaches. Hence, regression-based strategies for FP procedures [141] constitute an essential gap in
the literature regarding rotor faults. The performance metrics of the IFDP studies related to rotor faults mostly include only overall
model accuracy, which may not be sufficient to understand the fault-specific performance of a proposed technique. As seen in Fig. 13, a
significant variety of rotor faults are covered. On the other hand, these rotor faults are mostly examined regarding single-fault oc­
currences where it is assumed that a single fault exists in the system. Although there are studies that investigated compound fault
conditions [11,40,114], they considered a rotor fault that occurred at the same time as a fault of another component such as a bearing
or gear.

Fig. 13. Percentile distribution of the considered rotor fault types.

17
O. Das et al. Heliyon 9 (2023) e17584

5.4. Other faults

Rotating machines comprise various other faults that occur on the bearing, gear, or rotor including belt-pulley faults, fan or turbine
faults, looseness, shaft faults, etc. As seen in Fig. 5, these faults constitute the minority of the studies related to IFDP. Table 4 presents
the IFDP studies of the different faults that may occur in rotating machines.
According to Table 4, the majority of the studies are related to the FD procedure of shaft and stator faults. The faults of the remained
components such as the belt-pulley, turbine blade, or fasteners of the rotating machine are rarely investigated. There are also a few
studies related to the looseness faults that may occur in different parts of the rotating machines [40,132,135]. Similar to the studies
related to bearing, gear, or rotor faults only a few studies take confusion matrix, precision, recall, or any other metrics in FD procedures
of rotating machines. Most of the studies assessed their intelligent model based on the overall accuracy of fault detection and fault
identification in rotating machines [15,21,25,40,46,88]. In addition, the majority of the studies are focused on single fault occurrences
rather than compound faults [40,86].

6. AI-based approaches in fault diagnosis and prognostics of rotating machines

6.1. Data sources

In the IFDP of rotating machines, various kinds of data sources and data acquisition methods are used to constitute an intelligent
model for FD and FP. Each fault has a unique impact on the machine monitoring data in different ways. Hence, it is essential to choose a
suitable data source and data acquisition technique for the faulty conditions of a rotating machine.
The most common data sources are as follows. Visual Data comprises image data obtained through high-resolution cameras, optical
devices, or images derived from a signal. Visual data may include thermal images [32,114], x-ray images [143,144], acoustic signals
[11,23,24], or vibration data [75,85,106] or their processed form including spectrogram [145,146], waveforms [147], etc that usually
reflects the characteristics of the healthy and faulty behavior of a component of a machine. Acoustic Data includes the sound or noise
that is produced by a rotating machine that differs when operating in a healthy and faulty condition. Acoustic data is gathered via a
microphone, acoustic emission sensors, and accelerometer [11,23,24] to conduct an effective FD and FP procedure for rotating ma­
chines. Vibration/Kinematic Data is constituted by the dynamic behavior of a rotating machine such as vibration, acceleration, or
velocity. Vibration/Kinematic Data can be gathered by accelerometers, tachometers, and gyroscopes [11,30,74,84,91,96] to deter­
mine the condition of the rotating machinery. There are also other data sources and data acquisition methods for IFDP of rotating
machines including the measurement of feeding currents by transducers, magnetic field by magnetometers, EMG, and pressure/depth
values by depth sensors [9,27,47,97,141]. The selection of the data source and data acquisition method depends on cost, effectiveness,
real-time adaptability, environmental conditions, robustness, sensitivity, installation, and maintenance complexity.
As seen in Fig. 14, vibration data is generally used for the IFDP of the bearing, gear, rotor, and other components. Visual/image data
are also considered for mostly FD of rotating machine components. However, these data are broadly used for bearings and rotors. In
addition, visual/image data are frequently used for FD procedures whereas a few studies used them for FDE, FI, and FP. Acoustic
signals are also used, yet such studies are lesser than the other data sources regarding the FD of rotating machines. Apart from those
data sources, some studies used current signals, EMG, SCADA data, temperature, torque, pressure, and velocity for the IFDP of rotating
machine components.

Table 4
Intelligent fault diagnosis and prognosis studies related to faults of other components in rotating machine.
Studies Year Procedures Component Fault Types

FD FDE FI FP

[22] 2013 X Stator Stator Winding

[67] 2014 X Shaft Bowed
[85] 2014 X X Belt-pulley Belt Looseness, Belt Damage,
Pulley Damage
[68] 2015 X Shaft Bowed Shaft
[7] 2016 X X Stator Shorted Turns
[83] 2016 X Stator Shorted Turns
[82] 2017 X Shaft Misalignment, Imbalance,
Misalignment + Imbalance
[78] 2017 X Shaft Shaft Crack
[18] 2018 X Shaft Imbalance, Damaged Shaft
[131] 2018 X Bearing Block Looseness
[43] 2019 X Shaft Misalignment
[37] 2019 X Shaft Looseness, Looseness + Gear Fault
[128] 2021 X Couplings Looseness
[15] 2022 X X Turbine Blade Blade Fault
[25] 2023 X Turbine Blade, Fastener Blade damage, Bolt Damage

18
O. Das et al. Heliyon 9 (2023) e17584

Fig. 14. The considered type of data source distribution for IFDP of rotating machines.

Table 5 shows the studies and the data sources used for FD, FDE, FI, and FP procedures of rotating machine components. The
abbreviations given in the table are; ES: Electric Signal, TQ: Torque, PR: Pressue, TM: Tachometer. Some of those are briefly presented as
follows.
Kang et al. [71] used two-dimensional gray-level images for the FD of the bearing, shaft, and rotors of rotating machines. For this
purpose, they employed wavelet transform (WT) considering Shannon wavelets for multiresolution analysis of the vibration signals.
They considered angular misalignment, broken rotor bar fault, rotor unbalance fault, bearing outer race fault, bowed shaft, and bowed
rotor faults to assess their proposed approach. Li et al. [24] considered acoustic emission and vibration signals with several ML al­
gorithms for the FD of bearing and gear faults. They used a benchmark dataset where gear wore tooth, chaffing tooth, pitting tooth,
chipped tooth, root crack tooth, bearing inner race fault, bearing outer race fault, bearing ball fault, and bearing house eccentricity
exist. Zolfaghari et al. [140] performed the FDE procedure of rotors in a rotating machine regarding broken rotor bar fault cases. They
used the current signal acquired from the stator to represent the healthy and faulty components. Chen et al. [27] used vibration, torque,
and pressure data for FD of a planetary gearbox existed in a wind turbine drivetrain diagnosis simulator. They considered five con­
ditions of the planetary gearbox including healthy gear, surface worn fault, missing tooth, chipped tooth, and cracked tooth. Zhao et al.
[29] conducted an FP procedure to predict the RUL of a bearing component of the direct-drive wind turbine. For this purpose, they
considered a digital-twin-based method where environmental parameters and operational parameters are given as input to constitute a
virtual output in terms of vibration signals. Afterwards, they used actual vibration data obtained from the main bearing of a wind
turbine. Both vibration data are then processed to build an intelligent FP model.

6.2. Feature extraction

IFDP models require the specific features of the given data to constitute meaningful relationships to give accurate results. There are
various feature extraction methods used in the IFDP of rotating machines. Time-domain analysis is used to express the behavior of the
signals acquired from the machine in the time domain. Some time-domain features are peak values, root means square values, and crest
factors. Frequency-domain analysis transforms the acquired raw data into the frequency domain by considering specific techniques such
as Fast Fourier Transform, Wavelet Transform, Empirical Mode Decomposition, and Singular Value Decomposition. Some frequency-
domain features are spectral energy, bandwidth, dominant frequency, etc. Statistical feature extraction comprises the evaluation of the
statistical metrics including mean, median, nth percentile, kurtosis, skewness, and standard deviation from the raw data. Signal pro­
cessing methods process the given data by filtering, normalizing, standardizing, or smoothing to find suitable features. Apart from those
manual techniques, machine learning algorithms are also used to find the patterns in the raw data and find the relevant features
automatically.
The acquired data sources are subjected to manual or automatic feature extraction procedures for IFDP procedures. The most
feature extraction techniques regarding FD, FDE, FI, or FP procedures of rotating machine components are statistical feature extraction
(STFE) [10,27,29,101], Fast Fourier Transform (FFT) [8,33,34,109], Wavelet Transform (WWT) [20,73,78,134], and EMD [21,41,52].
There are also some methodologies derived from those techniques including Discrete Fourier Transform (DFT) [95], Short Time
Fourier Transform (STFT) [36], Wavelet Packet Transform (WPT) [37,83], Continuous Wavelet Transform (CWT) [36,38], Discrete
Wavelet Transform (DWT) [76,122], and Ensemble Empirical Mode Decomposition (EEMD) [112,121]. There are also other methods
such as Singular Value Decomposition (SVD) [42,82,121], Similarity-Based Modelling (SBM) [100], and Hilbert Transform [35,42]
that are all considered manual feature extraction techniques. On the other hand, developed ML models such as Convolutional Neural
Networks (CNN) [106,107], Generative Adversarial Networks (GAN) [108], Long-Short Term Memory (LSTM) [83,107,130], and
Autoencoders (AE) [87,105] are used for automatic feature extraction procedures. Considering the popular feature extraction

19
O. Das et al. Heliyon 9 (2023) e17584

Table 5
Intelligent fault diagnosis and prognosis studies and considered data source types.
Studies Year Component Procedures Data source

FD FDE FI FP Visual Acoustic Vibration Others

[136] 2013 Rotor X X X

[16] 2013 Bearing X X
[71] 2014 Bearing, Shaft, Rotor X X X
[91] 2014 Bearing, Gear, Rotor X X(c) X
[89] 2014 Belt-pulley X X X
[138] 2015 Rotor X X X
[72] 2015 Bearing, Shaft X X
[24] 2016 Bearing, Gear X X X
[137] 2016 Rotor X X(c) X
[7] 2016 Bearing, Stator, Rotor X X ES
[41] 2017 Bearing X X(c) X
[90] 2017 Machine Faults X X
[86] 2017 Shaft, Rotor X X
[140] 2017 Rotor X X(c) ES
[135] 2018 Rotor X X
[98] 2018 Bearing, Gear X X
[21] 2018 Bearing, Gear, Shaft X X
[100] 2018 Bearing, Rotor X X
[11] 2019 Bearing X X
[23] 2019 Bearing X X
[127] 2019 Bearing, Gear X X(c) X
[133] 2019 Rotor X X
[134] 2019 Rotor X ES
[22] 2019 Bearing, Rotor X X
[46] 2019 Shaft X X
[40] 2019 Bearing, Gear, Rotor X X
[109] 2020 Bearing X X X
[111] 2020 Bearing X X(c) X
[115] 2020 Bearing, Rotor X X
[8] 2020 Gear X X
[79] 2019 Rotor X X X
[142] 2020 Journal Bearings X X
[107] 2021 Bearing, Gear X X(c) X
[13] 2021 Bearing, Gear X SCADA
[27] 2021 Gear X X TQ, PR
[12] 2021 Rotor X X(c) X
[132] 2021 Bearing, Rotor X X
[85] 2021 Gear X X
[104] 2022 Bearing X X X
[101] 2022 Bearing, Gear X X
[75] 2022 Bearing, Gear X X
[84] 2022 Gear X X TM
[80] 2022 Rotor X X
[15] 2022 Bearing, Turbine blades X X X
[141] 2022 Rotor X X ES
[11] 2023 Bearing, Rotor X X X TM
[139] 2023 Rotor X X(c) X
[125] 2023 Bearing X X(c) X
[29] 2023 Bearing X X X SCADA

strategies used for IFDP of rotating machines, it is seen from Fig. 15 that the statistical features of raw signals or processed signals are
the most preferred method followed by WT-based techniques, deep learning-based (DL-based) methods, FFT-based techniques, and
EMD-based approaches, respectively.
Table 6 show the studies related to IFDP of rotating machines considering the employed feature extraction and data types.

6.3. Fusion techniques

There are several fusion methods in machine learning models to develop effective intelligent techniques. Fusion techniques are
referred to as combining data or related information from different sources at various levels. By using fusion techniques, it is possible to
obtain ML models that give more accurate and robust decisions.
In general, the fusion procedures are constituted from three levels. Data fusion is related to combining data from different sources to
improve the representability of an intelligent model. For instance, data fusion may be effective when combining data collected from
the same data sources which are placed in different locations of machinery [148,149]. A fault’s signature may appear differently in
those sources. However, that difference may result in a positive impact on the performance of the intelligent model. Based on this

20
O. Das et al. Heliyon 9 (2023) e17584

Fig. 15. Feature extraction techniques used for IFDP of rotating machines.

example, multi-modal sensor fusion or multi-location sensor fusion methods are effective approaches in data fusion. Multi-modal sensor
fusion [148] comprises combining data acquired from distinct sensors (e.g., accelerometer + microphone, tachometer + thermal
cameral + acoustic emission sensor, etc.) that measure different modalities or any other different physical data. By doing so, the
limitations of each sensor may be mitigated. Multi-location sensor fusion [149] is the integration of the data collected from the same or
different sensors that are placed in different locations. It is mainly considered to improve the accuracy and reliability of an intelligent
model. Multi-location sensor fusion may require additional techniques including Kalman Filters, Bayesian Networks, or different
regression methods. Feature fusion denotes the integration of the feature sets extracted from the same or different sensors [150]. For
instance, a bearing fault may be found by using the feature sets obtained from a thermal camera and accelerometer. Decision fusion is
related to giving a final decision by combining the decisions made by different intelligent methods based on different data sources and
feature sets [151].
Fusion techniques are useful to solve complex problems, especially where the error tolerance is required to be very small. By using
fusion techniques, improved accuracy, robustness, and reliability may be provided.
Regarding the IFDP of rotating machines, there are only a few studies that investigated the effectiveness of the data fusion tech­
niques regarding multi-modal sensor fusion, feature fusion, and decision fusion [9,11,24,27,29]. Most IFDP approaches seek the use of
a single data source obtained from a single sensor or multi-location sensor fusion approach where data is acquired from the same type
of sensors placed in different locations on the components of the rotating machine.

6.4. Machine learning and data mining

ML can be categorized as supervised learning, unsupervised learning, and semi-supervised learning. ML models are used in a vast
range of areas including image recognition [1,2], speech recognition [4], recommendation systems [152], autonomous driving systems
[153,154], etc. Supervised learning is based on learning the relationships between the input and output. The performance of the model
is assessed by testing it with unseen data. The inputs may be categorical, numeric, ordered values, statistical values, time-series values,
or matrices. The outputs may be categorical data (classification) or continuous data (regression) [155–157]. In other words, a su­
pervised learning-based ML algorithm can identify the fault type of a faulty machine (classification) or aim to predict a continuously
changing fault severity value (regression). Supervised learning-based algorithms may be linear (Logistic Regression, Linear
Discriminant Analysis) or non-linear (Naive Bayes, Support Vector Machines, Non-Linear Logistic Regression). In multi-dimensional
heterogeneous datasets, there are mostly non-linear relationships among the data. It is necessary to use suitable models that can
express these relationships effectively. In semi-supervised learning, labeled and unlabeled data are used together to train the ML
model. In this kind of learning strategy, the majority of the training data is unlabeled. Semi-supervised learning is generally used in
problems where lots of data exist and labeling the data is difficult or costly [155–157]. Unsupervised learning is another learning
strategy where labeling is based on extracting patterns and relationships among the unlabeled input data [155–157]. Some unsu­
pervised learning techniques are Association Rule Mining, Clustering, and Anomaly Detection. In an ARM task, the co-occurrences of
the inputs are calculated. On the other hand, clustering is about aggregation based on the patterns between the data. Anomaly
detection clusters the data by finding and discarding the outliers. As seen in Fig. 16, the majority of the studies considered supervised
learning approaches followed by unsupervised learning and semi-supervised learning.
In the field of rotating machines, various ML algorithms are used or improved to propose an effective IFDP method. Table 7 shows
the studies related to the IFDP of rotating machines regarding the considered components and ML algorithms. It is seen from Table 7
that the use of popular single learners such as Support Vector Machines (SVM) [24,41,87,88,117], K-Nearest Neighbors (KNN) [44,90,
96,115], and Decision Tree (DT) [16,17,74] are widely used for IFDP of bearing [16,41,78,120], gear [8,27,83,130], rotor [26,134,
137], shaft [46,71], and stator [7,87]. On the other hand, ensemble learning approaches such as Random Forest (RF) [81,94,97],
AdaBoost (AB) [8,11,113], Gradient Boosting Decision Tree (GBDT) [14], Light Gradient Boosting Machine (LGBM) [12], eXtreme
Gradient Boosting (XGB) [14], and CatBoost (CB) [52] are rarely used when compared with single learners, deep learning approaches,

21
O. Das et al. Heliyon 9 (2023) e17584

Table 6
Intelligent fault diagnosis and prognosis studies and feature extraction methods.
Studies Year Component Data Types Feature extraction

STFE FFT WWT EMD Other

[16] 2013 Bearing Acoustic X

[6] 2013 Bearing, Rotor Voltage, Current X X X
[71] 2014 Bearing, Shaft, Rotor Visual, Vibration X X
[92] 2015 Bearing, Gear Vibration X X
[93] 2015 Bearing, Gear Vibration X X
[122] 2015 Bearing Vibration X X
[123] 2015 Bearing Vibration X X
[72] 2015 Bearing, Shaft Vibration X
[24] 2016 Bearing, Gear Acoustic, Vibration X
[94] 2016 Bearing, Rotor Vibration X X
[31] 2016 Bearing Vibration X X
[87] 2016 Bearing, Stator, Rotor Vibration X
[41] 2017 Bearing Vibration X X X
[95] 2017 Bearing, Gear Vibration X
[119] 2017 Bearing Vibration X X X
[121] 2017 Bearing Vibration X X
[73] 2018 Bearing Vibration X X X
[97] 2018 Bearing, Gear EMG X X
[21] 2018 Bearing, Gear, Shaft Vibration X
[99] 2018 Bearing, Gear Vibration X X
[48] 2018 Bearing Vibration X
[116] 2019 Bearing Visual X
[23] 2019 Bearing Acoustic X
[126] 2019 Bearing Visual X
[44] 2019 Bearing Vibration X
[40] 2019 Bearing, Gear, Rotor Vibration X X X
[108] 2020 Bearing Vibration X
[50] 2020 Bearing Vibration X
[109] 2020 Bearing Visual X X X
[47] 2020 Bearing Visual, Current
[110] 2020 Bearing Vibration X
[111] 2020 Bearing Vibration X X
[112] 2020 Bearing Vibration X X
[115] 2020 Bearing, Rotor Visual X
[37] 2021 Bearing Vibration X
[14] 2021 Bearing Vibration X X
[32] 2021 Bearing Visual X
[35] 2021 Bearing Vibration X X
[107] 2021 Bearing, Gear Vibration X
[13] 2021 Bearing, Gear SCADA X
[132] 2021 Bearing, Rotor Vibration X
[34] 2021 Bearing, Rotor Visual X
[52] 2022 Bearing Vibration X
[104] 2022 Bearing Visual, Vibration X X
[10] 2022 Bearing Vibration X X X
[46] 2022 Bearing Vibration X
[101] 2022 Bearing, Gear Vibration X X
[75] 2022 Bearing, Gear Vibration X
[105] 2022 Bearing Vibration X
[15] 2022 Bearing, Turbine blades Vibration X
[11] 2023 Bearing, Rotor Acoustic, Vibration, Tachometer X X
[125] 2023 Bearing Visual X
[29] 2023 Bearing Visual, Vibration, SCADA X X X

and neural networks. These approaches are considered for IFDP of bearing [97,100,116], gear [43,81,97], and rotor [26,100]. Neural
networks such as Artificial Neural Networks (ANN) [7,30,39,122,140] and Multilayer Perceptrons (MLP) [8,25,47,95] are extensively
used for IFDP of bearing [39,74,122], gear [8,27,83] rotor [25,86,134,140], stator [25], shafts [46,86], and other components [89] of
rotating machines. Deep learning algorithms are increasingly considered the IFDP of rotating machines. Especially the use and
improvement of Convolutional Neural Networks (CNN) [29,92,119,125,126,135,139] have drawn the attention of many researchers
as seen in Table 7. Deep learning approaches such as Deep Belief Neural Networks (DNN) [15,114,127], Deep Neural Networks (DNN)
[14,15,87,114,120], Long-Short Term Memory (LSTM) [29,84,101,107,132], Generative Adversarial Network (GAN) [108],
Autoencoder (AE) [34,105,111,114], Domain Adaptive Neural Networks (DANN) [45,75,125,139] are considered for bearing [73,75,
87,119,120,125], gear [43,75], rotor [87,133,135,139], stator [87], shaft [46], and other components [15,28] regarding FD, FDE, FI,
and FP procedures. As seen from Fig. 17, SVM and CNN-based ML algorithms are the most used techniques for the IFDP of rotating
machines. On the other hand, ensemble learning methods are the least considered algorithms in this field.

22
O. Das et al. Heliyon 9 (2023) e17584

Fig. 16. The learning approaches considered for IFDP of rotating machines.

6.5. Types of rotating machines and real-world settings

There are numerous types of rotating machines developed for a specific purpose used in the industry. Each type has different
challenges in fault analysis. For instance, the operational conditions of a centrifugal pump may significantly alter depending on the
type of fluid, fluid pressure, temperature, and variable system needs. All of these parameters make challenging the identification of
faulty or even healthy pumps. Another example may be given for gas and steam turbines. These machines operate at very high
temperatures which makes collecting and processing data by a mounted sensor compelling. Regarding those machines, sensor
placement, effective data collection, and effective data processing may be considered key difficulties. As one of the most common
rotating machine types, wind turbines bring specific challenges in intelligent fault analysis. The collected data is highly affected by
external noise, aerodynamic forces, and environmental conditions. For instance, vibration and acoustic data are remarkably affected
by the excessive noise generated by the turbine itself and the aerodynamic forces developed by the wind. In such machines, multi-
modal fault diagnosis may be more beneficial to reduce such drawbacks. Similar challenges are observed for each rotating machine
and therefore, it is essential to understand the machine-specific difficulties when collecting and processing data and constituting an
IFDP model.
As presented in the former sections, numerous studies are conducted in the field of fault analysis of rotating machines. A majority of
those studies considered benchmark datasets collected from hand-crafted laboratory test beds having a motor that drives a shaft
connected to a set of gears and/or supported with bearings. Such test beds are also denoted as machine fault simulators. Some other
built an experimental setup to represent the rotating machines or directly consider the real-world rotating machines, including wind
turbines [13,99,125,130,158–160], rotor test rigs of aircraft [8], compressors [161–163], gas turbines [164,165], steam turbines [166,
167], pumps [168–170], train bogie [171] and other rotating machine-included systems such as unmanned underwater or aerial
vehicles [8,172] for machine learning-based fault analysis. Some recent works regarding the fault diagnosis of the rotating machines
mentioned above are presented as follows.
Han and Li proposed an out-of-distribution assisted trustworthy machinery fault diagnosis considering bearing faults of a wind
turbine and gear fault of a single-stage planetary gearbox. They integrated deep ensemble neural networks to constitute an ensemble
fault diagnosis structure. The ensemble was established by five individual deep base learners including a three-convolutional layered
CNN, four-convolutional layered CNN, five-convolutional layered CNN, a gradient-skipping ResNet architecture, and an interception
block built by GoogleNet. They compared the results of the deep ensemble learning with each individual deep base learner. They
concluded that the developed deep ensemble learning method was superior to others and gave promising results in handling the
problem of unseen faults [158].
Zhang et al. employed a convolutional deep belief network-based fault diagnosis approach for reciprocating compressors. They
developed an auto-denoising network in which the signal is denoised and the most representative features are extracted. To evaluate
the probability of fault occurrence, they adopted multiple Gaussian process classifiers. Following the combination of those properties
considering an optimized weight, the fault type is identified. They concluded that their proposed approach effectively diagnosed the
reciprocating compressor faults with accuracy rates of up to 91.89% [161]. Hasan et al. presented a grayscale-converted scalog­
ram-based adaptive deep convolutional neural network (ADCNN) for fault diagnosis of centrifugal pumps. For this purpose, they
collected the vibration signals and decompose them by using Continuous Wavelet Transform to obtain scalograms. Afterwards, they
converted the scalograms into grayscale images to be fed as input to ADCNN. They compared their proposed approach with different
fault diagnosis models such as CNN, FFT + CNN, and Statistical Feature Extraction + kNN. They concluded that the developed
approach outperformed these models by an improvement in accuracy of up to 15.6% [168]. Manikandan and Duraivelu developed a
deep CNN-based approach for fault diagnosis of the impeller and mechanical seal in industrial mono-block centrifugal pumps. For this
purpose, they converted the collected vibration signals into 2D images. They trained and test the deep CNN classifier using such image
data. They concluded that their proposed approach identified the faults with an accuracy value of 99.07% [170]. Ding et al. proposed a
multiscale lightweight network with adaptive pruning for fault diagnosis of train bogie bearings. They constituted weight-sharing

23
 Table 7
Intelligent fault diagnosis and prognosis studies and used machine learning algorithms.
Studies Year Component Algorithm

O. Das et al.
CNN DBN DNN LSTM GAN ANN MLP CB XGB LGBM GBDT RF AB DT SVM KNN Others

[16] 2013 Bearing X

[22] 2013 Stator, Rotor X X
[120] 2014 Bearing X X
[88] 2015 Bearing, Gear X X
[118] 2015 Bearing X X X
[21] 2016 Bearing, Gear X X X X
[90] 2016 Bearing, Rotor X X X
[17] 2016 Gear X X
[83] 2016 Bearing, Stator, Rotor X X X X
[91] 2017 Bearing, Gear X X X X
[92] 2017 Bearing, Gear X X X
[40] 2017 Gear X X X X
[82] 2017 Shaft, Rotor X X X X
[116] 2017 Bearing X X X
[136] 2017 Rotor X X
[69] 2018 Bearing X X
[70] 2018 Bearing, Gear X X
[131] 2018 Rotor X X X
[93] 2018 Bearing, Gear X X
[45] 2018 Bearing X X X
[41] 2019 Bearing X X X X
[123] 2019 Bearing, Gear X X
[46] 2019 Bearing, Rotor X X X
[126] 2019 Gear X X X X
[130] 2019 Rotor X X X
24

[43] 2019 Shaft X X X

[104] 2020 Bearing X X X
[44] 2020 Bearing X X X X
[106] 2020 Bearing X X
[125] 2020 Bearing, Gear X X X X
[109] 2020 Bearing, Gear X X X X X
[110] 2020 Bearing, Rotor X X X X
[111] 2020 Bearing, Rotor X X X X
[8] 2020 Gear X X X X X X X X
[9] 2020 Gear X X X X X
[14] 2021 Bearing X X X X X X
[29] 2021 Bearing X X
[32] 2021 Bearing X X X X
[103] 2021 Bearing, Gear X X
[13] 2021 Bearing, Gear X X X X
[24] 2021 Gear X X X X X
[124] 2021 Gear X X
[12] 2021 Rotor X X X X X
[128] 2021 Bearing, Rotor X X
[49] 2022 Bearing X X

Heliyon 9 (2023) e17584

[10] 2022 Bearing X X X X
[42] 2022 Bearing X X
[97] 2022 Bearing, Gear X X
[80] 2022 Gear X X
[99] 2022 Gear X X X X
[76] 2022 Rotor X
[11] 2023 Bearing, Rotor X X X
[135] 2023 Rotor X X
[121] 2023 Bearing X X
[26] 2023 Bearing X X X
O. Das et al. Heliyon 9 (2023) e17584

Fig. 17. The usage distribution of the ML algorithms in the field of IFDP of rotating machines.

multiscale convolutions to obtain the multi-time scale features of vibration signals. Afterwards, they structured the inverse separable
convolution blocks to obtain highly representative features. Besides, they implemented an adaptive pruning technique to eliminate
redundant network structures during training. They concluded that the developed technique is superior to other recent methods in
fault diagnosis of lightweight bearings working under variable conditions [171].
Some of those studies indicated that their proposed approach is based on real-time since the data that they used are acquired from
one or multiple experimental setups that are given above. It is controversial to consider such approaches as a real-time-based IFDP
method since gathering the data from a rotating machine in real-time does not precisely mean that the constituted model is going to
operate in real time with the same performance. Besides, in real-time, the IFDP model may completely fail in distinguishing the healthy
machine from the damaged one or identifying the correct fault type and/or fault severity due to improper algorithm and/or parameter
selection or the differences in the environmental/operational conditions of the same machine used to train the model. Although the
hand-crafted fault simulators or rotating machines that are developed in laboratories may mimic real-world operating machines, they
still remain as developed or tested in a laboratory environment. All such laboratory equipment are denoted as idealized since they do
not completely represent real-world settings such as environmental parameters (i.e., temperature, humidity, pressure, and ambient
noise) and operation conditions (i.e., motor speed, load, voltage, disruptive or coupled vibrations). In addition, the fault types existing
in those laboratory setups are artificial and therefore, do not depict the natural forming of the faults. This may cause deceptive or
ineffective outcomes since it is essential to detect, localize, and identify the fault in its early-stage to plan or execute the maintenance
procedure.
Some studies in the literature comprise the real-time assessment of their proposed approach by developing an interface where a
simulation including data acquisition, data processing, and fault analysis procedures are conducted sequentially regarding the same
machine in real-time after training the model [22,38,172–174].
Lee et al. used CNN for fault diagnosis of induction motors in real time. They considered rotor and bearing faults (punctured rotor,
worn bearing) for the fault detection procedure. They employed a simple pipeline where the data is collected via a vibration sensor and
sent to the build model from a data acquisition device. They constituted an interface in LabVIEW software in which all the steps of the
fault diagnosis can be monitored easily. They achieved an overall accuracy of 98.66% regarding the identification of the normal state
and all faulty conditions of the machine [22]. Guo et al. proposed a multi-task CNN with information fusion for bearing fault diagnosis.
They included an online data processing step in their approach where the data is given to multi-task CNN as input for fault diagnosis.
They tested their proposed approach considering two benchmark datasets, namely Case Western Reserve University Data (CWRU) and
Machinery Fault Simulator – Rotor Dynamics Simulator Bearing Data (MFS-RDS) where bearing faults exist. They achieved an average
accuracy of 95.04% regarding fault identification and localization in real-time [38]. Bagci Das and Birant used the combination of
genetic algorithm and ensemble learning (GASEL) for fault detection of autonomous underwater vehicles (AUV). The vehicle includes
propeller damages with two severities, load-increase fault, and sensor faults. They assessed the performance of their proposed
approach by employing Flask environment in Phyton and Postman API to receive the raw data and send it to the fault diagnosis
module. They achieved an average accuracy of 99.33% including the prediction of the healthy state and all faulty conditions of the
vehicle [172]. Zhao et al. combined transfer learning and embedded convolutional LSTM with digital twin technology for bearing life
prediction in wind turbines. For this purpose, they generated the digital twin of the vibrational behavior of the main bearing of the
wind turbine in the ANSYS environment. They tested their proposed approach regarding two benchmark datasets, namely IEEE
Prognostics Health Management (PHM) Data Challenge and XJTU-SY. Afterwards, they conducted a second experiment on a

25
O. Das et al. Heliyon 9 (2023) e17584

real-world 1.5 MW wind turbine in Inner Mongolia. Regarding the real-world test results, they predicted the life of the bearing several
months earlier than its actual lifespan [173]. Syafrudin et al. proposed a Density-Based Spatial Clustering (DBSCAN) with Noise-based
outlier detection and RF for real-time fault detection in an automotive assembly line process. For this purpose, they utilized Apache
Kafka, Apache Storm, and Mongo DB to handle the real-time procedures. They achieve a perfect score regarding fault detection when
using DBSCAN with RF [174].
Such studies given above provide promising outcomes regarding the performance of their proposed IFDP methods when they are
operated in real-time conditions. On the other hand, they still do not satisfy the real-world settings thoroughly since the IFDP ap­
proaches developed in those studies are constituted in a conservative environment that mostly neglects the environmental and
operational variable parameters.
It is seen from the literature survey that the implementation of IFDP models in real-world settings is rarely explored. The devel­
opment of the IFDP models that satisfies the uncertainties due to environmental and operational parameters is ultimately required to
constitute robustness and reliance in the field of intelligent fault analysis of rotating machines. This may require time since it takes time
for fault development in normal operation. Such time consumption may be reduced by improper use of the machine. However, doing
so may result in the complete failure of the machine and may even cause undesired costly consequences. Regarding real-world ap­
plications, a semi-supervised or unsupervised technique may be more beneficial since gathering the faulty data to train an IFDP model
is challenging as there are numerous fault types that show up in certain circumstances and requires time as mentioned above. However,
using semi-supervised or unsupervised methods may cause a struggle in the identification of fault type since a semi-supervised or
unsupervised IFDP method may mispredict the fault type due to the similar or distinctive characteristics that show up in the vitals of a
machine. Hence, a blend of supervised and semi-supervised/unsupervised methods may give more accurate and reliable results in real-
world applications. Such a combination may be denoted as another important feature direction. Finally, it is favorable to use an IFDP
model that effectively operates in different machines in the real world. For instance, it is preferred to constitute an IFDP model that
successfully performs fault analysis in gearboxes both used for wind turbines and for industrial reduction gearboxes. For such purposes,
numerous studies are focused on transfer learning-based approaches [75,103,118,139,165] to achieve an IFDP model as generalizable
as possible. Although generalizability is an important aspect of IFDP procedures, it is more critical to focus on a specific rotating
machine in real-world settings since it is possible to develop distinct IFDP models for each machine in a factory that are in commu­
nication with each other in a harmony with the concept of Industry 4.0. On the other hand, from the perspective of Industry 4.0, it is a
must to use a reliable and robust IFDP approach that gives accurate, meaningful, and explainable results in real-time.

7. Summary and recent challenges

This review encompasses the literature related to IFDP of rotating machines regarding various aspects including machine com­
ponents, fault types, fault existence, fault analysis strategy, data sources, data fusion, feature extraction, and machine learning
technique. It is seen that the most examined component is the bearing followed by the gear, rotor, shaft, stator, and other parts (RQ1).
Regarding the fault types, bearing faults are investigated the most but are limited to inner race, outer race, cage, and ball faults.
Following the bearing faults, gear and rotor faults are commonly considered. For gear faults, crack, wear, chipped tooth, broken tooth,
missing tooth, and pitting faults are widely remarked. For rotor faults, broken rotor, misalignment, unbalance, and rub-impact faults
are extensively considered. Stator, shaft, and other faults are rarely taken into account and the IFDP approaches regarding those parts
are scarce (RQ2).
To constitute an IFDP approach various data sources including, vibration, acoustic, thermal, pressure, torque, current, and voltage
data are considered in the form of 1-D signal or visual data. Among these data sources, vibration is the most utilized one followed by
acoustic data (RQ3).
It is possible to improve the performance and reliability of an IFDP model by employing data fusion techniques. The data fusion
procedure is conducted at various levels in intelligent fault analysis of rotating machines. It is seen that the fusion process is usually
performed at the data level, including multi-modal sensor fusion and multi-location sensor fusion in the literature related to the IFDP of
rotating machines. Apart from those techniques, feature level fusion and decision level fusion are also considered in the literature
(RQ4).
Regarding the feature extraction techniques, hand-crafted techniques such as statistical features, FFT-based features (STFT and
DFT), WWT-based features (DWT, CWT, and WPT), and EMD-based features (EEMD, CEMD, CEEMDAN) are considered based on the
data source. In numerous studies, raw signals are directly used or processed to obtain visual or signal data to be used especially for deep
learning-based techniques. These techniques may include feature extraction layers that obtain the characteristic properties of a given
data automatically (RQ5).
The machine learning approaches used for the IFDP of rotating machines comprise supervised, unsupervised, and semi-supervised
learning. In general, supervised learning approaches are considered for FD, FDE, FI, and FP of rotating machines. All fault analysis
strategies are considered mostly as classification problems. On the other hand, it is more meaningful to pursue a regression-based
approach for FP procedures since fault severity is a continuous parameter. Besides, FP is widely conducted to help or assess the
RUL of a component. Hence, adopting a classification-based approach regarding FP only provides the capability of a model in catching
fault severity values at a certain level and therefore, does not present the trend of component condition that helps to find its RUL (RQ6).
A variety of machine learning techniques are used or improved to develop effective IFDP approaches regarding rotating machines.
In the last decade, it is seen that Support Vector Machines (SVM) and Convolutional Neural Networks (CNN) are the most used machine
learning techniques in the field of IFDP of rotating machines. In addition, Neural Network-based approaches and K-Neareast Neighbors
are commonly used in this field (RQ7).

26
O. Das et al. Heliyon 9 (2023) e17584

The developed IFDP approaches are assessed mostly by overall model accuracy regarding the FD procedures. On the other hand,
based on the considered machine learning approach (classification, regression, or clustering), precision, recall, confusion matrix, F1
score, AUC, ROC, mean squared error, root mean squared error, mean absolute error, and time consumption are considered (RQ8).
Introducing artificial intelligence to fault analysis of a machine has a significant impact on the maintenance strategy. Since the IFDP
procedures involve the condition monitoring and assessment of the RUL of a component or a machine the best strategies can be
predictive maintenance (PdM) and condition-based maintenance (CBM) (RQ9).
Although fault analysis is a helpful task, it has some challenges that should be overcome. The main challenges are given below
(RQ10).

• Data collection: A very common challenge is the collection of relevant and accurate fault data because of its strong impact on the
performance of machine learning methods. The rotating machines have numerous operational indicators including vibration,
speed, temperature, pressure, current and voltage data that may be useful for intelligent fault analysis. On the other hand, some
IFDP techniques such as deep learning-based approaches usually require higher amounts of data than other ML algorithms
regardless of the data type. To satisfy such a limitation, it is needed to acquire such an amount of data for both healthy and faulty
machines. It is challenging to gather data from a faulty machine in real-world settings since a machine operates mostly in healthy
conditions. The level of such a challenge increased even higher when a data-hunger ML method is adopted for fault analysis.
• Big data: With the recent development of smart manufacturing and the Internet of Things, the amount of collected fault-related data
has grown in an exponential manner and has high-dimensional characteristics. Besides, each machine condition poses uniqueness
in the collected data that increase the complexity of the problem. It is challenging to overcome such an issue with fewer data since
IFDP methods have limited capabilities to conduct the learning process grounded on real-world physics. Besides, they do not
comprehend the ties between cause and effect and therefore, do not come up with a conclusion based on such a relationship. Hence,
the requirement for big data becomes inevitable, especially for more complex problems such as observing multiple kinds of faults at
the same time. Analyzing large fault data is also a challenging task since it requires the use of powerful big data analytical tech­
niques (i.e., feature dimensionality reduction methods), specialized algorithms and methodologies, and a vast amount of storage
and computing resources.
• Imbalanced data: The fault-related data collected from rotating machinery in a real-world industrial system is highly imbalanced
since there are few machine failures relative to the number of normal conditions. The data imbalance may result in fallacious results
as it is possible that a developed IFDP method may predict all instances as healthy by discarding all faulty cases. This situation may
result in high accuracy, but poor precision and recall. Hence, the data imbalance constitutes another challenge to achieving
effective standard machine-learning solutions for fault analysis.
• Data pre-processing: Data pre-processing is a significant challenge in the field of IFDP of rotating machines since each kind of data
has a different impact on the model performance. It is required to obtain the most representative features to feed the ML model to
achieve successful results. However, the selection of the signal or image processing method, tuning its parameters, and extracting
the right features requires tedious work and expertise. For instance, regarding 1D time series signals, using Fast Fourier Transform
or Wavelet Transform generally constitute differences in model performance. Similarly, parameter tuning in signals processing (e.
g., sliding window time, overlap ratio, decomposition level, wavelet type, etc.) affected the model performance considerably as
they unearth or hide the fault-specific features that existed in the data.
• Dynamic system: Due to the fast-changing and dynamic Industry 4.0 environment, the machine learning system should be capable
to learn and flexible under changing conditions. The machine learning model should be continuously updated by the newly
collected data. In addition, it is challenging for an IFDP model under variable physical conditions including operational and
environmental parameters such as temperature, humidity, pressure, ambient noise, coupled vibrations, etc. An effective IFDP
model or models have to recognize the changes in such parameters and make predictions accordingly. In general, IFDP models are
used to solve a particular problem under confined conditions. Therefore, their model performance would be significantly affected
by changes in physical conditions. Such a limitation constitutes an extreme challenge regarding the development and employment
of IFDP models in real-world settings.
• Algorithm selection: Another significant difficulty is the selection of the machine learning technique. Each machine learning study
is unique in its context since the data availability affects the algorithm’s performance. Thus, the best machine learning method that
is capable to solve the target fault analysis problem is not clear. Selecting the appropriate learning strategy stands as another
challenge. Although IFDP techniques have a high capability in distinguishing data with different labels, it may not be useful to
employ supervised learning techniques in cases where an excessive amount of diversity exists in data. In such conditions, it is more
convenient to pursue a fault detection strategy rather than fault identification. Supervised algorithms require labeled data for
learning. On the other hand, it is difficult to constitute a dataset including all kinds of faults to train the supervised model due to the
possibility to obtain faulty data and the issues related to big data as mentioned above. In addition, the suitability of the chosen IFDP
method may change for different fault cases. It is expected that the researchers will be aware of potential ML algorithms suitable for
a given problem/data and will compare their performances based on various metrics.
• Optimal parameter selection: The performance of each algorithm on fault detection depends on the parameter settings. The optimal
parameter settings can be determined based on a lot of experimental work, which is computationally expensive. It is challenging to
assess the optimal parameters since each fault type has a unique characteristic impact on the data which may require a more
complex or more simple algorithm. Hence, it may be necessary to conduct a fault-specific parameter tuning of the selected IFDP
approach.

27
O. Das et al. Heliyon 9 (2023) e17584

• Evaluation: The assessment of the machine learning model in terms of understanding and interpreting the results obtained in fault
analysis is one of the most challenging processes. The question arises as to how to measure the performance and assess the validity
of the results. It is also necessary to prevent the underfitting and overfitting of the models. Hence, it is needed to provide a balanced
and satisfying amount of data for training since an IFDP approach may be prone to memorize the characteristics of the data instead
of learning may result in false predictions in the test set (overfitting) or may not learn anything due to insufficient amount of data
(underfitting). Another challenge is testing a proposed IFDP approach with unseen data that have different characteristics from the
training data in one way or another. In addition, a proposed method is desired to be tested under real-time and real-world con­
ditions. Although it is possible to test a model in real-time by using distinctive data, assessing the model in real-world settings is
extremely challenging and time-consuming since the fault growth procedure may take a long time and the consequences of a fault
may result in catastrophic failure of the machinery if the fault is not detected by the model. Another issue regarding the evaluation
of a proposed IFDP approach is dealing with unbalanced data. In the field of fault diagnosis of rotating machinery or any kind of
machine, the amount of healthy machine data is abundant since the machine operates normally in most cases. On the other hand,
collecting faulty data is challenging and even if it is collected, it will not be as plenty as that of the healthy data because the machine
eventually has to stop since operating a machine in a faulty condition poses a significant danger. Unbalanced data may result in
fallacious results such that a developed IFDP model may be successful in predicting the healthy condition while it may completely
fail to assess the faulty machine. Even in such a scenario, the overall accuracy of the IFDP model may surpass 90% due to the
majority of the healthy cases included in the training and testing data. In such cases, it is extremely significant to measure the model
performance by additional metrics such as precision, recall, confusion matrix, actual-predicted diagrams/values, mean absolute
error rates, etc. This may constitute an additional challenge since it is necessary to be aware of such situations. In addition, it may be
compelling to select suitable performance metrics for the model evaluation.
• Interdisciplinary collaboration: Machine learning for rotating machinery is a multi-disciplinary research area. Hence, expertise
from disciplines such as mechanical and computer sciences may be required. Cooperation among the corresponding experts is likely
needed so as to accomplish machine learning implementations.
• Real-time adaptation: A fundamental challenge is developing an effective intelligent model that is adaptable for real-time appli­
cations because machinery components are significantly affected by environmental factors such as temperature, pressure, hu­
midity, pollutants, noise, etc. In addition, machines usually do not operate as they are assumed to do in controlled real or virtual
environments such as laboratories or simulations. Hence, it is essential to develop a model that is resilient against such factors. On
the other hand, it is also challenging to assess the performance of the IFDP model in real-world settings due to time consumption
and potential failures that may result in unnecessary costs.
• Domain adaptation: In general, intelligent techniques are trained and tested considering a single benchmark or commercial dataset.
Although pursuing such an approach is usually fruitful, it may still give fallacious results regarding the performance of the
developed model when it is tested with unseen data acquired from the same kind of component from a different machine.
Therefore, it is necessary to constitute an intelligent method that is able to give accurate predictions regarding the same domain
(component) of different machines.
• Fault types and compound faults: Another significant challenge is to determine the limits in terms of fault type or fault occurrences
when developing an IFDP model. It is necessary to consider all kinds of faults that are likely to occur during an operation. In several
cases, a fault development may cause to arise another fault in the same or different component, which may adversely impact the
model performance. Hence, it is necessary to pursue a deterministic approach and consider compound faults when constituting an
IFDP approach for rotating machines.

8. Research directions

Based on the studies in the literature and recent challenges, the following research directions may be considered in further studies
(RQ11).

1. Real-world dataset: The majority of the dataset is fabricated in laboratory conditions where environmental or other disruptive
factors are discarded. Most of the studies related to the IFDP of rotating machines used such benchmark datasets to assess their
proposed approach. On the other hand, it is necessary to consider these factors because they somehow affect the indicators of a
machine. Hence, future studies may comprise or propose datasets in real-world settings for intelligent fault analysis of rotating
machines although collecting such data may take a significant amount of time. By doing so, it may be possible to constitute IFDP
models that are able to handle the changes in environmental and operational parameters.
2. Component and fault-specific models: Although it is possible to constitute a generalizable model regarding the fault analysis of
multiple components, it may be more beneficial to develop an IFDP model for each type of component and tune its parameters
considering the type of fault. By doing so, the complexity of the problem would be decreased since only the faults of a specific
component are considered instead of taking numerous kinds of faults belonging to various components. Hence, further studies may
investigate the effectiveness of a component and fault-specific IFDP approach.
3. Multiple or Compound faults: Rotating machines may include multiple faults occurring in the same or different components. There
are scarce studies in the literature that measures the effectiveness of the IFDP approaches regarding the compound fault condition
in rotating machines. Future studies may measure the effectiveness of IFDP approaches considering compound faults as such sit­
uations may be observed in real-world settings.

28
O. Das et al. Heliyon 9 (2023) e17584

4. Complete Fault Analysis: Most studies proposed IFDP approaches by considering fault detection, fault identification, fault locali­
zation, or fault prognosis procedures. Developing effective intelligent techniques that are capable to conduct a complete fault
analysis where all procedures are covered will significantly contribute to the related community. For real-world settings, it is also
suggested to combine different ML methods in a fault analysis pipeline where each algorithm is responsible to handle a specific task.
For instance, one semi-supervised model may be developed for fault detection while one supervised model may in charge to
distinguish the environmental anomalies that affect to data used for fault detection. By combining ML approaches in such a way, it
may be possible to avoid deceiving results and constitute a more robust IFDP method.
5. Ensemble Methods: The majority of IFDP techniques developed for rotating machines considered deep learning (DL) approaches
due to the positive aspects regarding automated feature extraction and high accuracy. On the other hand, these approaches may
require expensive powerful devices, which may not be suitable for embedded systems. In addition, DL methods are generally time-
consuming, which is not practical for real-world applications. Hence, the ensemble learning techniques are needed to be examined
in more depth since they may not demand a high computational load and, at the same time, are accurate and robust.
6. Semi-supervised and Unsupervised Methods: The majority of the studies take supervised approaches into account where labeled
instances are used to train the IFDP model. On the other hand, using supervised methods may not be practical in real-world settings
since it is extremely challenging to constitute a dataset in which all possible faults that may occur in a component are included. This
would take so much time, and effort and may result in undesirable consequences including the complete failure of a machine.
Practically, semi-supervised and unsupervised approaches are more convenient for fault detection of rotating machines. In addi­
tion, having numerous and a wide variety of labeled data may not be handled by a proposed approach for fault diagnosis. In such
cases, employing a semi-supervised or unsupervised approach may be more fruitful for fault detection instead of trying to identify
and localize the fault.
7. Human-in-the-Loop: As a recent popular approach, human-in-the-loop aims to combine supervised learning with human feedback
active learning where an intelligent approach makes more accurate decisions as it receives correct commands or inputs from a
human. Future studies related to the IFDP of rotating machines would significantly benefit from such an approach, especially in
fault diagnosis procedures where the fault is identified and localized. By employing human-in-the-loop it is possible to constitute
more reliable IFDP approaches with a smaller size of initial data since the method will reinforce its learning by receiving a
command that includes the details of the condition of the machine.
8. Data Augmentation: The acquisition of the healthy data of a rotating machine component is easy because the machine generally
operates in its healthy condition. On the other hand, gathering faulty data from an industrial machine is a challenging process since
the machine may completely fail or even be destroyed during the data collection procedure. This raises another challenge, espe­
cially in the fault detection procedure regarding the amount of healthy and faulty data. Although there are numerous studies that
examined data augmentation to obtain higher amounts of data, it is still worth investigating as handling and the requirement of the
data may become compelling, especially in more complex rotating machines.
9. Model Assessment: Interpreting a performance of a developed intelligent model based on only accuracy may give fallacious results,
especially for unbalanced data. An IFDP model may give high accuracy and still make a false prediction. For instance, consider a
dataset with 1000 instances, which is split by 900 in the favor of “healthy condition” and 100 for “faulty condition”. A prediction
accuracy of 90% may be erroneous since all accurately predicted instances may cover the “healthy state” (900 of 1000) whereas the
entire “faulty state” instances are mispredicted as healthy. Hence, it is necessary to include other metrics such as precision and
recall. Regarding fault-specific tasks, the confusion matrix is another required metric to assess the model performance in a fault-
specific domain. For regression-based approaches, using solely coefficient of determination (R2), root mean squared error, or mean
absolute error may present misleading conclusions regarding the model performance. In addition to such metrics, it is critical to
explore the actual and predicted values and interpret how close they are. Hence, the future classification or regression-based IFDP
approaches may be assessed in more detail and comprehensively by including additional metrics. In addition to performance
metrics, it is strongly recommended to evaluate an IFDP model with unseen data and/or different benchmark or real-world
datasets.

9. Conclusions

This study presents a comprehensive review of IFDP procedures for rotating machinery by minding all the existing challenges. The
proposed IFDP approaches are widely discussed considering the fault analysis strategies, deemed data sources, data types, data fusion
techniques, machine learning techniques within the frame of the fault type, and compound faults that occur in components such as
bearings, gear, rotor, stator, shaft, and other parts. Besides, the existing challenges in developing, evaluating, and using IFDP ap­
proaches are presented. Finally, future directions are presented considering the issues and gaps existing in the literature to the
community to propose effective IFDP approaches for rotating machines.

Author contribution statement

All authors listed have significantly contributed to the development and the writing of this article.

Data availability statement

No data was used for the research described in the article.

29
O. Das et al. Heliyon 9 (2023) e17584

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.

References

[1] I. Sutskever Krizhevsky, G.E. Hinton, ImageNet classification with deep convolutional neural networks, Commun. ACM 60 (2012) 84–90, https://doi.org/
10.1145/3065386.
[2] K. Zhang, W. Zuo, Y. Chen, D. Meng, L. Zhang, Beyond a Gaussian denoiser: residual learning of deep CNN for image denoising, IEEE Trans. Image Process. 26
(2017) 3142–3155, https://doi.org/10.1109/tip.2017.2662206.
[3] V. Mnih, K. Kavukcuoglu, D. Silver, A.A. Rusu, J. Veness, M.G. Bellemare, A. Graves, M. Riedmiller, A.K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie,
A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, D. Hassabis, Human-level control through deep reinforcement learning, Nature 518 (2015)
529–533, https://doi.org/10.1038/nature14236.
[4] G. Hinton, L. Deng, D. Yu, G. Dahl, A. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen, T. Sainath, B. Kingsbury, Deep neural networks for acoustic
modeling in speech recognition: the shared views of four research groups, IEEE Signal Process. Mag. 29 (2012) 82–97, https://doi.org/10.1109/
msp.2012.2205597.
[5] J. Goodfellow, M. Pouget-Abadie, B. Mirza, D. Xu, Farley Warde, S. Ozair, A. Courville, Y. Bengio, Generative adversarial networks, Commun. ACM 63 (2020)
139–144, https://doi.org/10.1145/3422622.
[6] G. Clerc Soualhi, H. Razik, Detection and diagnosis of faults in induction motor using an improved artificial ant clustering technique, IEEE Trans. Ind. Electron.
60 (2013) 4053–4062, https://doi.org/10.1109/tie.2012.2230598.
[7] R.A. Patel, B.R. Bhalja, Condition monitoring and Fault Diagnosis of induction motor using support vector machine, Elec. Power Compon. Syst. 44 (2016)
683–692, https://doi.org/10.1080/15325008.2015.1131762.
[8] P.B. Mallikarjuna, M. Sreenatha, S. Manjunath, N.C. Kundur, Aircraft gearbox Fault Diagnosis system: an approach based on deep learning techniques, J. Intell.
Syst. 30 (2020) 258–272, https://doi.org/10.1515/jisys-2019-0237.
[9] M. Azamfar, J. Singh, I. Bravo-Imaz, J. Lee, Multisensor data fusion for gearbox fault diagnosis using 2-D convolutional neural network and motor current
signature analysis, Mech. Syst. Signal Process. 144 (2020), 106861, https://doi.org/10.1016/j.ymssp.2020.106861.
[10] K.N. Ravikumar, S.S. Aralikatti, H. Kumar, G.N. Kumar, K.V. Gangadharan, Fault diagnosis of antifriction bearing in internal combustion engine gearbox using
data mining techniques, International Journal of System Assurance Engineering and Management (2022), https://doi.org/10.1007/s13198-021-01407-1.
[11] O. Das, D. Bagci Das, Smart machine fault diagnostics based on fault specified discrete wavelet transform, J. Braz. Soc. Mech. Sci. Eng. 45 (2023), https://doi.
org/10.1007/s40430-022-03975-0.
[12] S. Liu, Z. Ji, Y. Wang, Z. Zhang, Z. Xu, C. Kan, K. Jin, Multi-feature fusion for fault diagnosis of rotating machinery based on convolutional neural network,
Comput. Commun. 173 (2021) 160–169, https://doi.org/10.1016/j.comcom.2021.04.016.
[13] H.S. Dhiman, D. Deb, S.M. Muyeen, I. Kamwa, Wind turbine gearbox anomaly detection based on adaptive threshold and twin support vector machines, IEEE
Trans. Energy Convers. (2021) 1, https://doi.org/10.1109/tec.2021.3075897, 1.
[14] C. Xiang, Z. Ren, P. Shi, H. Zhao, Data-driven Fault Diagnosis for rolling bearing based on DIT-FFT and XGBoost, Complexity 2021 (2021) 1–13, https://doi.
org/10.1155/2021/4941966.
[15] S.R. Saufi, M.F. Isham, Z.A. Ahmad, M.D.A. Hasan, Machinery fault diagnosis based on a modified hybrid deep sparse autoencoder using a raw vibration time-
series signal, J. Ambient Intell. Hum. Comput. (2022), https://doi.org/10.1007/s12652-022-04436-1.
[16] M. Amarnath, V. Sugumaran, H. Kumar, Exploiting sound signals for fault diagnosis of bearings using decision tree, Measurement 46 (2013) 1250–1256,
https://doi.org/10.1016/j.measurement.2012.11.011.
[17] A. Krishnakumari, M. Elayaperumal, C. Saravanan, Arvindan, Fault diagnostics of spur gear using decision tree and fuzzy classifier, Int. J. Adv. Des. Manuf.
Technol. 89 (2016) 3487–3494, https://doi.org/10.1007/s00170-016-9307-8.
[18] G. Dinardo, L. Fabbiano, G. Vacca, A smart and intuitive machine condition monitoring in the Industry 4.0 scenario, Measurement 126 (2018) 1–12, https://
doi.org/10.1016/j.measurement.2018.05.041.
[19] U. Kumar, D. Galar, Maintenance in the era of industry 4.0: issues and challenges, in: Quality, IT and Business Operations, Springer Singapore, 2017,
pp. 231–250, https://doi.org/10.1007/978-981-10-5577-5_19.
[20] 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, Comput. Ind. 123 (2020), 103298, https://doi.org/10.1016/j.compind.2020.103298.
[21] J. Chen, D. Zhou, C. Lyu, C. Lu, An integrated method based on CEEMD-SampEn and the correlation analysis algorithm for the fault diagnosis of a gearbox
under different working conditions, Mech. Syst. Signal Process. 113 (2018) 102–111, https://doi.org/10.1016/j.ymssp.2017.08.010.
[22] J.-H. Lee, J.-H. Pack, I.-S. Lee, Fault diagnosis of induction motor using convolutional neural network, Appl. Sci. 9 (2019) 2950, https://doi.org/10.3390/
app9152950.
[23] M.M.M. Islam, J.-M. Kim, Automated bearing fault diagnosis scheme using 2D representation of wavelet packet transform and deep convolutional neural
network, Comput. Ind. 106 (2019) 142–153, https://doi.org/10.1016/j.compind.2019.01.008.
[24] C. Li, R.-V. Sanchez, G. Zurita, M. Cerrada, D. Cabrera, R.E. Vásquez, Gearbox fault diagnosis based on deep random forest fusion of acoustic and vibratory
signals, Mech. Syst. Signal Process. 76–77 (2016) 283–293, https://doi.org/10.1016/j.ymssp.2016.02.007.
[25] M. Seera, C.P. Lim, D. Ishak, H. Singh, Application of the fuzzy min–max neural network to fault detection and diagnosis of induction motors, Neural Comput.
Appl. 23 (2013) 191–200, https://doi.org/10.1007/s00521-012-1310-x.
[26] D. Martin-Diaz, Morinigo-Sotelo, O. Duque-Perez, R.A. Osornio-Rios, R.J. Romero-Troncoso, Hybrid algorithmic approach oriented to incipient rotor fault
diagnosis on induction motors, ISA (Instrum. Soc. Am.) Trans. 80 (2018) 427–438, https://doi.org/10.1016/j.isatra.2018.07.033.
[27] S. Chen, R. Yang, M. Zhong, Graph-based semi-supervised random forest for rotating machinery gearbox fault diagnosis, Control Eng. Pract. 117 (2021),
104952, https://doi.org/10.1016/j.conengprac.2021.104952.
[28] J. Liu, G. Yang, X. Li, S. Hao, Y. Guan, Y. Li, A deep generative model based on CNN-CVAE for wind turbine condition monitoring, Meas. Sci. Technol. 34
(2023), 035902, https://doi.org/10.1088/1361-6501/aca496.
[29] W. Zhao, C. Zhang, J. Wang, O.G. Peyrano, F. Gu, S. Wang, D. Lv, Research on main bearing life prediction of direct-drive wind turbine based on digital twin
technology, Meas. Sci. Technol. 34 (2023), 025013, https://doi.org/10.1088/1361-6501/ac99f4.
[30] Z. Zhang, Y. Wang, K. Wang, Intelligent fault diagnosis and prognosis approach for rotating machinery integrating wavelet transform, principal component
analysis, and artificial neural networks, Int. J. Adv. Des. Manuf. Technol. 68 (2013) 763–773, https://doi.org/10.1007/s00170-013-4797-0.
[31] D. Dou, S. Zhou, Comparison of four direct classification methods for intelligent fault diagnosis of rotating machinery, Appl. Soft Comput. 46 (2016) 459–468,
https://doi.org/10.1016/j.asoc.2016.05.015.
[32] T. Choudhary, S. Mian, Fatima, Convolutional neural network based bearing fault diagnosis of rotating machine using thermal images, Measurement 176
(2021), 109196, https://doi.org/10.1016/j.measurement.2021.109196.
[33] S. Gundewar, P. Kane, A. Andhare, Detection of broken rotor bar fault in an induction motor using convolution neural network, Journal of Advanced
Mechanical Design, Systems, and Manufacturing 16 (2022), https://doi.org/10.1299/jamdsm.2022jamdsm0020. JAMDSM0020–JAMDSM0020.
[34] X. Wu, Y. Zhang, C. Cheng, Z. Peng, A hybrid classification autoencoder for semi-supervised fault diagnosis in rotating machinery, Mech. Syst. Signal Process.
149 (2021), 107327, https://doi.org/10.1016/j.ymssp.2020.107327.

30
O. Das et al. Heliyon 9 (2023) e17584

[35] S. Shen, H. Lu, M. Sadoughi, C. Hu, V. Nemani, A. Thelen, K. Webster, M. Darr, J. Sidon, S. Kenny, A Physics-Informed Deep Learning Approach for Bearing
Fault Detection, Engineering Applications of Artificial Intelligence, vol. 103, 2021, https://doi.org/10.1016/j.engappai.2021.104295.
[36] M. Hussain, D. Kumar Soother, I. Hussain Kalwar, T. Din Memon, Z. Ahmed Memon, K. Nisar, B. Shankar Chowdhry, Stator winding Fault Detection and
classification in three-phase induction motor, Intelligent Automation & Soft Computing 29 (2021) 869–883, https://doi.org/10.32604/iasc.2021.017790.
[37] L. Wan, K. Gong, G. Zhang, X. Yuan, C. Li, X. Deng, An efficient rolling bearing Fault Diagnosis method based on spark and improved random forest algorithm,
IEEE Access 9 (2021) 37866–37882, https://doi.org/10.1109/access.2021.3063929.
[38] S. Guo, B. Zhang, T. Yang, D. Lyu, W. Gao, Multitask convolutional neural network with information fusion for bearing Fault Diagnosis and localization, IEEE
Trans. Ind. Electron. 67 (2020) 8005–8015, https://doi.org/10.1109/tie.2019.2942548.
[39] S. Dong, D. Sun, B. Tang, Z. Gao, W. Yu, M. Xia, A Fault Diagnosis method for rotating machinery based on PCA and morlet kernel SVM, Math. Probl Eng.
(2014) 1–8, https://doi.org/10.1155/2014/293878, 2014.
[40] X.-B. Wang, Z.-X. Yang, P.K. Wong, C. Deng, Novel paralleled extreme learning machine networks for fault diagnosis of wind turbine drivetrain, Memetic
Computing 11 (2019) 127–142, https://doi.org/10.1007/s12293-018-0277-2.
[41] M. Sohaib, C.-H. Kim, J.-M. Kim, A hybrid feature model and deep-learning-based bearing Fault Diagnosis, Sensors 17 (2017), https://doi.org/10.3390/
s17122876.
[42] R. Chemseddine, M. Boualem, B. Djamel, F. Semchedine, Gear fault feature extraction and classification of singular value decomposition based on Hilbert
empirical wavelet transform, Journal of Vibroengineering 20 (2018) 1603–1618, https://doi.org/10.21595/jve.2017.18917.
[43] L. Jing, M. Zhao, P. Li, X. Xu, A convolutional neural network based feature learning and fault diagnosis method for the condition monitoring of gearbox,
Measurement 111 (2017) 1–10, https://doi.org/10.1016/j.measurement.2017.07.017.
[44] W. Gong, H. Chen, Z. Zhang, M. Zhang, R. Wang, C. Guan, Q. Wang, A novel deep learning method for Intelligent Fault Diagnosis of rotating machinery based
on improved CNN-SVM and multichannel data fusion, Sensors 19 (2019) 1693, https://doi.org/10.3390/s19071693.
[45] M. Kim, J.U. Ko, J. Lee, B.D. Youn, J.H. Jung, K.H. Sun, A Domain Adaptation with Semantic Clustering (DASC) method for fault diagnosis of rotating
machinery, ISA (Instrum. Soc. Am.) Trans. 120 (2022) 372–382, https://doi.org/10.1016/j.isatra.2021.03.002.
[46] X. Wu, Z. Penga, J. Renc, C. Chenga, W. Zhanga, D. Wang, Rub-impact Fault Diagnosis of rotating machinery based on 1-D convolutional neural networks, IEEE
Sensor. J. (2019), https://doi.org/10.1109/jsen.2019.2944157.
[47] D.T. Hoang, H.J. Kang, A motor current signal-based bearing Fault Diagnosis using deep learning and information fusion, IEEE Trans. Instrum. Meas. 69 (2020)
3325–3333, https://doi.org/10.1109/tim.2019.2933119.
[48] N. Zhang, L. Wu, J. Yang, Y. Guan, Naive Bayes bearing Fault Diagnosis based on enhanced independence of data, Sensors 18 (2018) 463, https://doi.org/
10.3390/s18020463.
[49] Y. Yang, H. Zheng, Y. Li, M. Xu, Y. Chen, A fault diagnosis scheme for rotating machinery using hierarchical symbolic analysis and convolutional neural
network, ISA (Instrum. Soc. Am.) Trans. 91 (2019) 235–252, https://doi.org/10.1016/j.isatra.2019.01.018.
[50] M. Qiao, S. Yan, X. Tang, C. Xu, Deep convolutional and LSTM recurrent neural networks for rolling bearing Fault Diagnosis under strong noises and variable
loads, IEEE Access 8 (2020) 66257–66269, https://doi.org/10.1109/access.2020.2985617.
[51] D. Zhang, X. Ren, H. Zuo, Compound Fault diagnosis for gearbox based using of euclidean matrix sample entropy and one-dimensional convolutional neural
network, Shock Vib. 2021 (2021) 1–26, https://doi.org/10.1155/2021/6669006.
[52] M.Y. Miaomiao Yang, W.L. Miaomiao Yang, W.Z. Weizhi Liu, M.W. Wenxuan Zhang, X.F. Mei Wang, Bearing vibration signal Fault Diagnosis based on LSTM-
cascade CatBoost, J. Internet Technol. 23 (2022) 1155–1161, https://doi.org/10.53106/160792642022092305024.
[53] P. Gangsar, R. Tiwari, Signal based condition monitoring techniques for fault detection and diagnosis of induction motors: a state-of-the-art review, Mech. Syst.
Signal Process. 144 (2020), 106908, https://doi.org/10.1016/j.ymssp.2020.106908.
[54] D. Neupane, J. Seok, Bearing Fault detection and diagnosis using case western Reserve university dataset with deep learning approaches: a review, IEEE Access
8 (2020) 93155–93178, https://doi.org/10.1109/access.2020.2990528.
[55] S.R. Saufi, Z.A.B. Ahmad, M.S. Leong, M.H. Lim, Challenges and opportunities of deep learning models for Machinery Fault Detection and diagnosis: a review,
IEEE Access 7 (2019) 122644–122662, https://doi.org/10.1109/ACCESS.2019.2938227.
[56] Y. Zhao, T. Li, X. Zhang, C. Zhang, Artificial intelligence-based fault detection and diagnosis methods for building energy systems: advantages, challenges and
the future, Renew. Sustain. Energy Rev. 109 (2019) 85–101, https://doi.org/10.1016/j.rser.2019.04.021.
[57] Z. Zhu, Y. Lei, G. Qi, Y. Chai, N. Mazur, Y. An, X. Huang, A review of the application of deep learning in intelligent fault diagnosis of rotating machinery,
Measurement 206 (2023), 112346, https://doi.org/10.1016/j.measurement.2022.112346.
[58] Y. Lei, B. Yang, X. Jiang, F. Jia, N. Li, A.K. Nandi, Applications of machine learning to machine fault diagnosis: a review and roadmap, Mech. Syst. Signal
Process. 138 (2020), 106587, https://doi.org/10.1016/j.ymssp.2019.106587.
[59] S. Lu, Q. He, J. Wang, A review of stochastic resonance in rotating machine fault detection, Mech. Syst. Signal Process. 116 (2019) 230–260, https://doi.org/
10.1016/j.ymssp.2018.06.032.
[60] Z. Gao, C. Cecati, S.X. Ding, A survey of Fault Diagnosis and fault-tolerant techniques—Part I: fault Diagnosis with model-based and signal-based approaches,
IEEE Trans. Ind. Electron. 62 (2015) 3757–3767, https://doi.org/10.1109/tie.2015.2417501.
[61] A.G. Nath, S.S. Udmale, S.K. Singh, Role of artificial intelligence in rotor fault diagnosis: a comprehensive review, Artif. Intell. Rev. 54 (2021), https://doi.org/
10.1007/s10462-020-09910-w.
[62] Y. Wang, J. Xiang, R. Markert, M. Liang, Spectral Kurtosis for Fault Detection, Diagnosis and Prognostics of Rotating Machines: A Review with Applications,
Mechanical Systems and Signal Processing, vols. 66–67, 2016, pp. 679–698, https://doi.org/10.1016/j.ymssp.2015.04.039.
[63] C. Stenström, P. Norrbin, A. Parida, U. Kumar, Preventive and corrective maintenance – cost comparison and cost–benefit analysis, Structure and Infrastructure
Engineering 12 (2015) 603–617, https://doi.org/10.1080/15732479.2015.1032983.
[64] S. Selcuk, Predictive maintenance, its implementation and latest trends, Proc. IME B J. Eng. Manufact. 231 (2016) 1670–1679, https://doi.org/10.1177/
0954405415601640.
[65] R. Ahmad, S. Kamaruddin, An overview of time-based and condition-based maintenance in industrial application, Comput. Ind. Eng. 63 (2012) 135–149,
https://doi.org/10.1016/j.cie.2012.02.002.
[66] K. Fischer, F. Besnard, L. Bertling, Reliability-centered maintenance for wind turbines based on statistical analysis and practical experience, IEEE Trans. Energy
Convers. 27 (2012) 184–195, https://doi.org/10.1109/tec.2011.2176129.
[67] F.O. Olorunniwo, A. Izuchukwu, Scheduling imperfect preventive and overhaul maintenance, Int. J. Qual. Reliab. Manag. 8 (1991), https://doi.org/10.1108/
02656719110141123.
[68] P. Gackowiec, General overview of maintenance strategies – concepts and approaches, Multidisciplinary Aspects of Production Engineering 2 (2019) 126–139,
https://doi.org/10.2478/mape-2019-0013.
[69] S.M. Lee, D. Lee, Y.S. Kim, The quality management ecosystem for predictive maintenance in the Industry 4.0 era, Int. J. Qual. Innovat. 5 (2019), https://doi.
org/10.1186/s40887-019-0029-5.
[70] Z. Li, Y. Wang, K.-S. Wang, Intelligent predictive maintenance for fault diagnosis and prognosis in machine centers: industry 4.0 scenario, Advances in
Manufacturing 5 (2017) 377–387, https://doi.org/10.1007/s40436-017-0203-8.
[71] M. Kang, J.-M. Kim, Reliable Fault diagnosis of multiple induction motor defects using a 2-D representation of Shannon wavelets, IEEE Trans. Magn. 50 (2014)
1–13, https://doi.org/10.1109/tmag.2014.2316474.
[72] J. Xiong, Q. Zhang, Z. Peng, G. Sun, W. Xu, Q. Wang, A diagnosis method for rotation machinery faults based on dimensionless indexes combined K-nearest
neighbor algorithm, Math. Probl Eng. (2015), https://doi.org/10.1155/2015/563954 (2015) 1–9.
[73] H. Pan, X. He, S. Tang, F. Meng, An improved bearing Fault Diagnosis method using one-dimensional CNN and LSTM, Strojniški Vestnik - Journal of
Mechanical Engineering (2018), https://doi.org/10.5545/sv-jme.2018.5249.

31
O. Das et al. Heliyon 9 (2023) e17584

[74] K. Vernekar, H. Kumar, G. KV, Engine gearbox fault diagnosis using machine learning approach, J. Qual. Mainten. Eng. 24 (2018) 345–357, https://doi.org/
10.1108/jqme-11-2015-0058.
[75] H. Cao, H. Shao, B. Liu, B. Cai, J. Cheng, Clustering-Guided novel unsupervised domain adversarial network for partial transfer Fault Diagnosis of rotating
machinery, IEEE Sensor. J. 22 (2022) 14387–14396, https://doi.org/10.1109/jsen.2022.3182727.
[76] N. Lu, Z. Xiao, O.P. Malik, Feature extraction using adaptive multiwavelets and synthetic detection index for rotor fault diagnosis of rotating machinery, Mech.
Syst. Signal Process. 52–53 (2015) 393–415, https://doi.org/10.1016/j.ymssp.2014.07.024.
[77] S. Altaf, M.W. Soomro, M.S. Mehmood, Fault Diagnosis and Detection in Industrial Motor Network Environment Using Knowledge-Level Modelling Technique,
Modelling and Simulation in Engineering, 2017, pp. 1–10, https://doi.org/10.1155/2017/1292190, 2017.
[78] S. Zgarni, H. Keskes, A. Braham, Nested SVDD in DAG SVM for induction motor condition monitoring, Eng. Appl. Artif. Intell. 71 (2018) 210–215, https://doi.
org/10.1016/j.engappai.2018.02.019.
[79] B. Wu, S. Feng, G. Sun, L. Xu, C. Ai, Fine-grained fault recognition method for shaft orbit of rotary machine based on convolutional neural network, Journal of
Vibroengineering 21 (2019) 2106–2120, https://doi.org/10.21595/jve.2019.20359.
[80] V. Biot-Monterde, A. Navarro-Navarro, I. Zamudio-Ramirez, J.A. Antonino-Daviu, R.A. Osornio-Rios, Automatic classification of rotor faults in soft-started
induction motors, based on persistence spectrum and convolutional neural network applied to stray-flux signals, Sensors 23 (2022) 316, https://doi.org/
10.3390/s23010316.
[81] D. Cabrera, F. Sancho, R.-V. Sánchez, G. Zurita, M. Cerrada, C. Li, R.E. Vásquez, Fault diagnosis of spur gearbox based on random forest and wavelet packet
decomposition, Front. Mech. Eng. 10 (2015) 277–286, https://doi.org/10.1007/s11465-015-0348-8.
[82] B. Merainani, C. Rahmoune, D. Benazzouz, B. Ould-Bouamama, A novel gearbox fault feature extraction and classification using Hilbert empirical wavelet
transform, singular value decomposition, and SOM neural network, J. Vib. Control 24 (2017) 2512–2531, https://doi.org/10.1177/1077546316688991.
[83] A. Yin, Y. Yan, Z. Zhang, C. Li, R.-V. Sánchez, Fault diagnosis of wind turbine gearbox based on the optimized LSTM neural network with cosine loss, Sensors
20 (2020) 2339, https://doi.org/10.3390/s20082339.
[84] J. Shi, D. Peng, Z. Peng, Z. Zhang, K. Goebel, D. Wu, Planetary gearbox fault diagnosis using bidirectional-convolutional LSTM networks, Mech. Syst. Signal
Process. 162 (2022), https://doi.org/10.1016/j.ymssp.2021.107996.
[85] G. Krishna Durbhaka, B. Selvaraj, M. Mittal, T. Saba, A. Rehman, L. Mohan Goyal, Swarm-LSTM: condition monitoring of gearbox Fault Diagnosis based on
hybrid LSTM deep neural network optimized by swarm intelligence algorithms, Comput. Mater. Continua (CMC) 66 (2021) 2041–2059, https://doi.org/
10.32604/cmc.2020.013131.
[86] M.A. Hassan, M.R. Habib, R.A. Abul Seoud, A.M. Bayoumi, Wavelet-based multiresolution bispectral analysis for detection and classification of helicopter
drive-shaft problems, J. Dyn. Syst. Meas. Control 140 (2017), https://doi.org/10.1115/1.4038243.
[87] W. Sun, S. Shao, R. Zhao, R. Yan, X. Zhang, X. Chen, A sparse auto-encoder-based deep neural network approach for induction motor faults classification,
Measurement 89 (2016) 171–178, https://doi.org/10.1016/j.measurement.2016.04.007.
[88] X. Jin, J. Feng, S. Du, G. Li, Y. Zhao, Rotor fault classification technique and precision analysis with kernel principal component analysis and multi-support
vector machines, Journal of Vibroengineering 16 (2014).
[89] H. Sun, K. Li, P. Chen, H. Wang, X. Ping, Y. Cao, A sequential fuzzy diagnosis method for rotating machinery using ant colony optimization and possibility
theory, J. Mech. Sci. Technol. 28 (2014) 1189–1201, https://doi.org/10.1007/s12206-014-0112-8.
[90] R. Chen, S. Chen, L. Yang, J. Wang, X. Xu, T. Luo, Looseness diagnosis method for connecting bolt of fan foundation based on sensitive mixed-domain features
of excitation-response and manifold learning, Neurocomputing 219 (2017) 376–388, https://doi.org/10.1016/j.neucom.2016.09.041.
[91] L. Jiang, H. Yin, X. Li, S. Tang, Fault diagnosis of rotating machinery based on multisensor information fusion using SVM and time-domain features, Shock Vib.
2014 (2014) 1–8, https://doi.org/10.1155/2014/418178.
[92] Z. Chen, C. Li, R.-V. Sanchez, Gearbox Fault Identification and Classification with Convolutional Neural Networks, Shock and Vibration. 2015, 2015, pp. 1–10,
https://doi.org/10.1155/2015/390134.
[93] C. Li, R.-V. Sanchez, G. Zurita, M. Cerrada, D. Cabrera, R.E. Vásquez, Multimodal deep support vector classification with homologous features and its
application to gearbox fault diagnosis, Neurocomputing 168 (2015) 119–127, https://doi.org/10.1016/j.neucom.2015.06.008.
[94] O. Janssens, V. Slavkovikj, B. Vervisch, K. Stockman, M. Loccufier, S. Verstockt, R. Van de Walle, S. Van Hoecke, Convolutional neural network based Fault
Detection for rotating machinery, J. Sound Vib. 377 (2016) 331–345, https://doi.org/10.1016/j.jsv.2016.05.027.
[95] X. Zhang, D. Jiang, T. Han, N. Wang, W. Yang, Y. Yang, Rotating Machinery Fault Diagnosis for imbalanced data based on fast clustering algorithm and support
vector machine, J. Sens. (2017) 1–15, https://doi.org/10.1155/2017/8092691, 2017.
[96] M. Xia, T. Li, L. Xu, L. Liu, C.W. de Silva, Fault diagnosis for rotating machinery using multiple sensors and convolutional neural networks, IEEE/ASME
Transactions on Mechatronics. 23 (2017) 101–110, https://doi.org/10.1109/TMECH.2017.2728371.
[97] R.-V. Sánchez, P. Lucero, R.E. Vásquez, M. Cerrada, J.-C. Macancela, D. Cabrera, Feature ranking for multi-fault diagnosis of rotating machinery by using
random forest and KNN, J. Intell. Fuzzy Syst. 34 (2018) 3463–3473, https://doi.org/10.3233/jifs-169526.
[98] Z. An, S. Li, J. Wang, W. Qian, Q. Wu, An Intelligent Fault Diagnosis approach considering the elimination of the weight matrix multi-correlation, Appl. Sci. 8
(2018) 906, https://doi.org/10.3390/app8060906.
[99] C. Huitao, J. Shuangxi, W. Xianhui, W. Zhiyang, Fault diagnosis of wind turbine gearbox based on wavelet neural network, J. Low Freq. Noise Vib. Act. Control
37 (2018) 977–986, https://doi.org/10.1177/1461348418795376.
[100] M.A. Marins, F.M.L. Ribeiro, S.L. Netto, E.A.B. da Silva, Improved similarity-based modeling for the classification of rotating-machine failures, J. Franklin Inst.
355 (2018) 1913–1930, https://doi.org/10.1016/j.jfranklin.2017.07.038.
[101] K. Vos, Z. Peng, C. Jenkins, M.R. Shahriar, P. Borghesani, W. Wang, Vibration-based anomaly detection using LSTM/SVM approaches, Mech. Syst. Signal
Process. 169 (2022), https://doi.org/10.1016/j.ymssp.2021.108752.
[102] F.K. Geitner, H.P. Bloch, Vibration Analysis, Machinery Failure Analysis and Troubleshooting, 2012, pp. 391–478, https://doi.org/10.1016/b978-0-12-
386045-3.00005-2.
[103] Z. Tang, L. Bo, X. Liu, D. Wei, A semi-supervised transferable LSTM with feature evaluation for fault diagnosis of rotating machinery, Appl. Intell. 52 (2022)
1703–1717, https://doi.org/10.1007/s10489-021-02504-1.
[104] F. He, Q. Ye, A bearing Fault Diagnosis method based on wavelet packet transform and convolutional neural network optimized by simulated annealing
algorithm, Sensors 22 (2022) 1410, https://doi.org/10.3390/s22041410.
[105] B. Yang, H. Sun, A zero-shot learning fault diagnosis method of rolling bearing based on extended semantic information under unknown conditions, J. Braz.
Soc. Mech. Sci. Eng. 45 (2022), https://doi.org/10.1007/s40430-022-03965-2.
[106] K. Yu, T.R. Lin, H. Ma, X. Li, X. Li, A multi-stage semi-supervised learning approach for intelligent fault diagnosis of rolling bearing using data augmentation
and metric learning, Mech. Syst. Signal Process. 146 (2021), 107043, https://doi.org/10.1016/j.ymssp.2020.107043.
[107] K.N. Ravikumar, A. Yadav, H. Kumar, K.V. Gangadharan, A.V. Narasimhadhan, Gearbox fault diagnosis based on Multi-Scale deep residual learning and
stacked LSTM model, Measurement 186 (2021), 110099, https://doi.org/10.1016/j.measurement.2021.110099.
[108] H. Yin, Z. Li, J. Zuo, H. Liu, K. Yang, F. Li, Wasserstein generative adversarial network and convolutional neural network (WG-CNN) for bearing Fault
Diagnosis, Math. Probl Eng. (2020) 1–16, https://doi.org/10.1155/2020/2604191, 2020.
[109] T. Han, Z. Tian, Z. Yin, A.C.C. Tan, Bearing fault identification based on convolutional neural network by different input modes, J. Braz. Soc. Mech. Sci. Eng. 42
(2020), https://doi.org/10.1007/s40430-020-02561-6.
[110] X. Li, X. Li, H. Ma, Deep representation clustering-based fault diagnosis method with unsupervised data applied to rotating machinery, Mech. Syst. Signal
Process. 143 (2020), 106825, https://doi.org/10.1016/j.ymssp.2020.106825.
[111] X. Zhao, M. Jia, A novel unsupervised deep learning network for intelligent fault diagnosis of rotating machinery, Struct. Health Monit. 19 (2020) 1745–1763,
https://doi.org/10.1177/1475921719897317.

32
O. Das et al. Heliyon 9 (2023) e17584

[112] S. Sun, K. Przystupa, M. Wei, H. Yu, Z. Ye, O. Kochan, Fast bearing fault diagnosis of rolling element using Lévy Moth-Flame optimization algorithm and Naive
Bayes, Eksploatacja I Niezawodnosc - Maintenance and Reliability 22 (2020) 730–740, https://doi.org/10.17531/ein.2020.4.17.
[113] Z. Jiang, Q. Han, X. Xu, Fault diagnosis of planetary gearbox based on motor current signal analysis, Shock Vib. (2020) 1–13, https://doi.org/10.1155/2020/
8854776, 2020.
[114] Y. Li, X. Du, F. Wan, X. Wang, H. Yu, Rotating machinery fault diagnosis based on convolutional neural network and infrared thermal imaging, Chin. J.
Aeronaut. 33 (2020) 427–438, https://doi.org/10.1016/j.cja.2019.08.014.
[115] P. Kumar, A. Shankar Hati, Convolutional neural network with batch normalisation for fault detection in squirrel cage induction motor, IET Electr. Power Appl.
15 (2020) 39–50, https://doi.org/10.1049/elp2.12005.
[116] G. Xu, M. Liu, Z. Jiang, D. Söffker, W. Shen, Bearing Fault diagnosis method based on deep convolutional neural network and random forest ensemble learning,
Sensors 19 (2019) 1088, https://doi.org/10.3390/s19051088.
[117] B. Wang, X. Hu, W. Wang, D. Sun, Fault diagnosis using improved pattern spectrum and fruit fly optimization algorithm–support vector machine, Adv. Mech.
Eng. 10 (2018), https://doi.org/10.1177/1687814018810935.
[118] R. Zhang, H. Tao, L. Wu, Y. Guan, Transfer learning with neural networks for bearing Fault Diagnosis in changing working conditions, IEEE Access 5 (2017)
14347–14357, https://doi.org/10.1109/access.2017.2720965.
[119] Y. Xie, T. Zhang, Fault diagnosis for rotating machinery based on convolutional neural network and empirical mode decomposition, Shock Vib. (2017) 1–12,
https://doi.org/10.1155/2017/3084197, 2017.
[120] M. Xia, T. Li, L. Liu, L. Xu, C.W. Silva, Intelligent fault diagnosis approach with unsupervised feature learning by stacked denoising autoencoder, IET Science,
Measurement & Technology 11 (2017) 687–695, https://doi.org/10.1049/iet-smt.2016.0423.
[121] K. Yu, T.R. Lin, J.W. Tan, A bearing fault diagnosis technique based on singular values of EEMD spatial condition matrix and Gath-Geva clustering, Appl.
Acoust. 121 (2017) 33–45, https://doi.org/10.1016/j.apacoust.2017.01.023.
[122] C. Gnanaprakasam, K. Chitra, S-transform and ANFIS for detecting and classifying the vibration signals of induction motor, J. Intell. Fuzzy Syst. 29 (2015)
2073–2085, https://doi.org/10.3233/ifs-151684.
[123] J. Li, S. Li, X. Chen, L. Wang, The hybrid KICA-GDA-LSSVM method research on rolling bearing fault feature extraction and classification, Shock Vib. (2015)
1–9, https://doi.org/10.1155/2015/512163, 2015.
[124] J. Zarei, M.A. Tajeddini, H.R. Karimi, Vibration analysis for bearing fault detection and classification using an intelligent filter, Mechatronics 24 (2014)
151–157, https://doi.org/10.1016/j.mechatronics.2014.01.003.
[125] P. Liang, B. Wang, G. Jiang, N. Li, L. Zhang, Unsupervised fault diagnosis of wind turbine bearing via a deep residual deformable convolution network based on
subdomain adaptation under time-varying speeds, Eng. Appl. Artif. Intell. 118 (2023), 105656, https://doi.org/10.1016/j.engappai.2022.105656.
[126] Z. Zhu, G. Peng, Y. Chen, H. Gao, A convolutional neural network based on a capsule network with strong generalization for bearing fault diagnosis,
Neurocomputing 323 (2019) 62–75, https://doi.org/10.1016/j.neucom.2018.09.050.
[127] Y. Li, L. Zou, L. Jiang, X. Zhou, Fault diagnosis of rotating machinery based on combination of deep belief network and one-dimensional convolutional neural
network, IEEE Access 7 (2019) 165710–165723, https://doi.org/10.1109/access.2019.2953490.
[128] Z. Wang, H. Huang, Y. Wang, Fault diagnosis of planetary gearbox using multi-criteria feature selection and heterogeneous ensemble learning classification,
Measurement 173 (2021), https://doi.org/10.1016/j.measurement.2020.108654.
[129] Y. Li, K. Wang, Modified Convolutional Neural Network with Global Average Pooling for Intelligent Fault Diagnosis of Industrial Gearbox, Eksploatacja I
Niezawodnosc - Maintenance and Reliability, vol. 22, 2020, https://doi.org/10.17531/ein.2020.1.8.
[130] L. Cao, Z. Qian, H. Zareipour, Z. Huang, F. Zhang, Fault diagnosis of wind turbine gearbox based on deep Bi-directional long short-term memory under time-
varying non-stationary operating conditions, IEEE Access 7 (2019) 155219–155228, https://doi.org/10.1109/access.2019.2947501.
[131] L.-M. Wang, Y.-M. Shao, Crack Fault classification for planetary gearbox based on feature selection technique and K-means clustering method, Chin. J. Mech.
Eng. 31 (2018), https://doi.org/10.1186/s10033-018-0202-0.
[132] A.G. Nath, A. Sharma, S.S. Udmale, S.K. Singh, An early classification approach for improving structural rotor Fault Diagnosis, IEEE Trans. Instrum. Meas. 70
(2021) 1–13, https://doi.org/10.1109/tim.2020.3043959.
[133] X. Zhu, D. Hou, P. Zhou, Z. Han, Y. Yuan, W. Zhou, Q. Yin, Rotor fault diagnosis using a convolutional neural network with symmetrized dot pattern images,
Measurement 138 (2019) 526–535, https://doi.org/10.1016/j.measurement.2019.02.022.
[134] T. Ince, Real-time broken rotor bar fault detection and classification by shallow 1D convolutional neural networks, Electrical Engineering 101 (2019) 599–608,
https://doi.org/10.1007/s00202-019-00808-7.
[135] S. Guo, T. Yang, W. Gao, C. Zhang, A novel Fault Diagnosis method for rotating machinery based on a convolutional neural network, Sensors 18 (2018) 1429,
https://doi.org/10.3390/s18051429.
[136] D. Zhao, J. Yan, Ant colony clustering analysis based intelligent fault diagnosis method and its application to rotating machinery, Pattern Anal. Appl. 16 (2013)
19–29, https://doi.org/10.1007/s10044-012-0289-3.
[137] L. Duan, M. Xie, T. Bai, J. Wang, A new support vector data description method for machinery fault diagnosis with unbalanced datasets, Expert Syst. Appl. 64
(2016) 239–246, https://doi.org/10.1016/j.eswa.2016.07.039.
[138] Z. Wang, J. Fan, Fault early recognition and health monitoring on aeroengine rotor system, J. Aero. Eng. 28 (2015), 04014065, https://doi.org/10.1061/(asce)
as.1943-5525.0000386.
[139] L. Xiang, X. Zhang, Y. Zhang, A. Hu, H. Bing, A novel method for rotor fault diagnosis based on deep transfer learning with simulated samples, Measurement
207 (2023), 112350, https://doi.org/10.1016/j.measurement.2022.112350.
[140] S. Zolfaghari, S. Noor, M. Rezazadeh Mehrjou, M. Marhaban, N. Mariun, Broken rotor bar Fault Detection and classification using wavelet packet signature
analysis based on fourier transform and multi-layer perceptron neural network, Appl. Sci. 8 (2018) 25, https://doi.org/10.3390/app8010025.
[141] C.E. Gonzalez Guillen, A.M. De Porras Cosano, P. Tian, J. Colmenares Diaz, A. Zarzo, C.A. Platero, Synchronous machines field winding turn-to-turn fault
severity estimation through machine learning regression algorithms, IEEE Trans. Energy Convers. 37 (2022), https://doi.org/10.1109/tec.2022.3159772, 1–1.
[142] D.S. Alves, G.B. Daniel, H.F. de Castro, T.H. Machado, K.L. Cavalca, O. Gecgel, J.P. Dias, S. Ekwaro-Osire, Uncertainty quantification in deep convolutional
neural network diagnostics of journal bearings with ovalization fault, Mech. Mach. Theor. 149 (2020), 103835, https://doi.org/10.1016/j.
mechmachtheory.2020.103835.
[143] J. Ren, R. Ren, M. Green, X. Huang, Defect detection from X-ray images using A three-stage deep learning algorithm, in: 2019 IEEE Canadian Conference of
Electrical and Computer Engineering (CCECE), 2019, https://doi.org/10.1109/ccece.2019.8861944.
[144] L. Jiang, Y. Wang, Z. Tang, Y. Miao, S. Chen, Casting defect detection in X-ray images using convolutional neural networks and attention-guided data
augmentation, Measurement 170 (2021), 108736, https://doi.org/10.1016/j.measurement.2020.108736.
[145] J. Wang, R.X. Gao, R. Yan, Multi-scale enveloping order spectrogram for rotating machine health diagnosis, Mech. Syst. Signal Process. 46 (2014) 28–44,
https://doi.org/10.1016/j.ymssp.2013.06.001.
[146] M.T. Pham, J.-M. Kim, C.H. Kim, Accurate bearing Fault Diagnosis under variable shaft speed using convolutional neural networks and vibration spectrogram,
Appl. Sci. 10 (2020) 6385, https://doi.org/10.3390/app10186385.
[147] X. Jiang, J. Wang, J. Shi, C. Shen, W. Huang, Z. Zhu, A coarse-to-fine decomposing strategy of VMD for extraction of weak repetitive transients in fault
diagnosis of rotating machines, Mech. Syst. Signal Process. 116 (2019) 668–692, https://doi.org/10.1016/j.ymssp.2018.07.014.
[148] R.-T. Wu, M.R. Jahanshahi, Data fusion approaches for structural health monitoring and system identification: past, present, and future, Struct. Health Monit.
19 (2020), 147592171879876, https://doi.org/10.1177/1475921718798769.
[149] O. Kullu, E. Cinar, A deep-learning-based multi-modal sensor fusion approach for detection of equipment faults, Machines 10 (2022) 1105, https://doi.org/
10.3390/machines10111105.
[150] Z. Chen, W. Li, Multisensor feature fusion for bearing Fault Diagnosis using sparse autoencoder and deep belief network, IEEE Trans. Instrum. Meas. 66 (2017)
1693–1702, https://doi.org/10.1109/tim.2017.2669947.

33
O. Das et al. Heliyon 9 (2023) e17584

[151] Q. Chao, H. Gao, J. Tao, Y. Wang, J. Zhou, C. Liu, Adaptive decision-level fusion strategy for the fault diagnosis of axial piston pumps using multiple channels
of vibration signals, Sci. China Technol. Sci. 65 (2022) 470–480, https://doi.org/10.1007/s11431-021-1904-7.
[152] I. Portugal, P. Alencar, D. Cowan, The use of machine learning algorithms in recommender systems: a systematic review, Expert Syst. Appl. 97 (2018) 205–227,
https://doi.org/10.1016/j.eswa.2017.12.020.
[153] A. Sallab, M. Abdou, E. Perot, S. Yogamani, Deep reinforcement learning framework for autonomous driving, Electron. Imag. (2017) 70–76, https://doi.org/
10.2352/issn.2470-1173.2017.19.avm-023, 2017.
[154] H. Fujiyoshi, T. Hirakawa, T. Yamashita, Deep learning-based image recognition for autonomous driving, IATSS Res. 43 (2019) 244–252, https://doi.org/
10.1016/j.iatssr.2019.11.008.
[155] Ethem Alpaydin, Introduction to Machine Learning, fourth ed., MIT Press, 2020.
[156] I.H. Witten, E. Al, Data Mining : Practical Machine Learning Tools and Techniques, Morgan Kaufmann, Amsterdam, 2017.
[157] J. Han, M. Kamber, P. Computer, Data Mining : Concepts and Techniques, Elsevier/Morgan Kaufmann, Amsterdam ; Boston, 2012.
[158] T. Han, Y.-F. Li, Out-of-distribution detection-assisted trustworthy machinery fault diagnosis approach with uncertainty-aware deep ensembles, Reliab. Eng.
Syst. Saf. 226 (2022), 108648, https://doi.org/10.1016/j.ress.2022.108648.
[159] Z. Liang, L. Zhang, X. Wang, A novel intelligent method for Fault Diagnosis of steam turbines based on T-SNE and XGBoost, Algorithms 16 (2023) 98, https://
doi.org/10.3390/a16020098.
[160] A. Amin, A. Bibo, M. Panyam, P. Tallapragada, Vibration Based Fault Diagnostics in a Wind Turbine Planetary Gearbox Using Machine Learning, vol. 47, Wind
Engineering, 2022, 0309524X2211239, https://doi.org/10.1177/0309524x221123968.
[161] Y. Zhang, J. Ji, B. Ma, Fault diagnosis of reciprocating compressor using a novel ensemble empirical mode decomposition-convolutional deep belief network,
Measurement 156 (2020), 107619, https://doi.org/10.1016/j.measurement.2020.107619.
[162] Y. Zhang, L. Zhang, Intelligent fault detection of reciprocating compressor using a novel discrete state space, Mech. Syst. Signal Process. 169 (2022), 108583,
https://doi.org/10.1016/j.ymssp.2021.108583.
[163] D. Cabrera, A. Guamán, S. Zhang, M. Cerrada, R.-V. Sánchez, J. Cevallos, J. Long, C. Li, Bayesian approach and time series dimensionality reduction to LSTM-
based model-building for fault diagnosis of a reciprocating compressor, Neurocomputing 380 (2020) 51–66, https://doi.org/10.1016/j.neucom.2019.11.006.
[164] W. Molla Salilew, Z. Ambri Abdul Karim, T. Alemu Lemma, Investigation of fault detection and isolation accuracy of different Machine learning techniques
with different data processing methods for gas turbine, Alex. Eng. J. 61 (2022) 12635–12651, https://doi.org/10.1016/j.aej.2022.06.026.
[165] X. Yang, M. Bai, J. Liu, J. Liu, D. Yu, Gas path fault diagnosis for gas turbine group based on deep transfer learning, Measurement 181 (2021), 109631, https://
doi.org/10.1016/j.measurement.2021.109631.
[166] K. Yao, S. Fan, Y. Wang, J. Wan, D. Yang, Y. Cao, Anomaly detection of steam turbine with hierarchical pre-warning strategy, IET Generation, Transm. Distrib.
16 (2022) 2357–2369, https://doi.org/10.1049/gtd2.12452.
[167] A. Dhini, I. Surjandari, B. Kusumoputro, A. Kusiak, Extreme learning machine – radial basis function (ELM-RBF) networks for diagnosing faults in a steam
turbine, Journal of Industrial and Production Engineering 39 (2021) 572–580, https://doi.org/10.1080/21681015.2021.1887948.
[168] M.J. Hasan, A. Rai, Z. Ahmad, J.-M. Kim, A Fault Diagnosis framework for centrifugal pumps by scalogram-based imaging and deep learning, IEEE Access 9
(2021) 58052–58066, https://doi.org/10.1109/ACCESS.2021.3072854.
[169] R. Tiwari, D.J. Bordoloi, A. Dewangan, Blockage and cavitation detection in centrifugal pumps from dynamic pressure signal using deep learning algorithm,
Measurement 173 (2021), 108676, https://doi.org/10.1016/j.measurement.2020.108676.
[170] A. Ding, Y. Qin, B. Wang, L. Jia, X. Cheng, Lightweight multiscale convolutional networks with adaptive pruning for Intelligent Fault Diagnosis of train bogie
bearings in edge computing scenarios, IEEE Trans. Instrum. Meas. 72 (2023), https://doi.org/10.1109/tim.2022.3231325, 1–13.
[171] S. Manikandan, K. Duraivelu, Vibration-based Fault Diagnosis of broken impeller and mechanical seal failure in industrial mono-block centrifugal pumps using
deep convolutional neural network, Journal of Vibration Engineering & Technologies 11 (2022), https://doi.org/10.1007/s42417-022-00566-0.
[172] D. Bagci Das, D. Birant, GASEL: genetic algorithm-supported ensemble learning for fault detection in autonomous underwater vehicles, Ocean Engineering 272
(2023), 113844, https://doi.org/10.1016/j.oceaneng.2023.113844.
[173] M. Syafrudin, G. Alfian, N. Fitriyani, J. Rhee, Performance analysis of IoT-based sensor, big data processing, and machine learning model for real-time
monitoring system in automotive manufacturing, Sensors 18 (2018) 2946, https://doi.org/10.3390/s18092946.
[174] W. Zhao, C. Zhang, J. Wang, O.G. Peyrano, F. Gu, S. Wang, D. Lv, Research on main bearing life prediction of direct-drive wind turbine based on digital twin
technology, Meas. Sci. Technol. 34 (2022), 025013, https://doi.org/10.1088/1361-6501/ac99f4.

34