Predictive maintenance using deep learning and

frequency harmonics analysis

Herman Maicol Gomez (  [email protected] )
Autonomous University of the West: Universidad Autonoma de Occidente https://orcid.org/0009-0006-
2167-2035
Juan Carlos Perafan-Villota
Universidad Autónoma de Occidente: Universidad Autonoma de Occidente

Research Article

Keywords: Predictive maintenance, Mechanical Vibration Analysis, Fast Fourier Transform, Finite
Elements Analysis, Condition Based Maintenance

Posted Date: August 31st, 2023

DOI: https://doi.org/10.21203/rs.3.rs-3267465/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License
 Predictive maintenance using deep learning and
frequency harmonics analysis
Gomez Herman M.1*† and Perafan-Villota Juan1*†
1 Automatics and Electronics Department, Universidad Autónoma de
Occidente, Cll 25 # 115-85
Km 2 Vı́a Cali - Jamundi, Cali, 760030, Colombia.

*Corresponding author(s). E-mail(s): herman mai.gomez [email protected];

[email protected];
† These authors contributed equally to this work.

Abstract
Reciprocating machines are critical assets in factories. Thus, predictive main-
tenance (PdM) plays a crucial role in avoiding unbudgeted costs. Firstly, a
preliminary diagnostic is carried out by qualified personnel who interpret data
from sensors based on the knowledge learned from maintenance manuals and
their expertise, which can turn into a spend and monotonous task. Currently,
machine learning and signals analysis hold great potential in predictive mainte-
nance, allowing assist experts to make appropriate decisions in advance. However,
there are many reasons why so far it is not something easy to do or that is
scalable. In this paper, we propose a low-cost architecture based on a convolu-
tional neural network feed by vibration data, which achieves high performance
in identifying the current asset health, getting robust features regardless of little
information available.

Keywords: Predictive maintenance, Mechanical Vibration Analysis, Fast Fourier

Transform, Finite Elements Analysis, Condition Based Maintenance.

1 Introduction
Modern industry has had many changes from the 20th century; one of these is the
way to keep in motion the machinery which has passed from only realizing repairs or
corrective maintenance to realize scheduling repairs and control testing or predictive

1
maintenance (PdM) [1]. The PdM relies on a paradigm shift in the mind of the main-
tenance team because it is not enough to know how the machine works, because it is
more important to know how it can fail in order to develop techniques that can predict
the failures and avoid the adverse effects that produce in the environment, the health,
production, and profitability of the company, thereby increasing the competitiveness
and the market position.
A monitoring system for reciprocating machinery is a challenging type of PdM since
its physical characteristics raise the difficulty in generating a reliable diagnosis based
on a single measurement. Another complex problem to be solved is the unbalanced
and limited amount of data available for every such status or condition machinery,
which are references needed in decision making.
A common type of reciprocating machinery is the 4-stroke internal combustion
engine, which can fail unexpectedly due to the difficulty of diagnosing the health of
its components. This type of machinery has many moving elements which are subject
to wear. However, the most critical points of its operation and performance are in the
power head, where all the components are related to the energy conversion perfor-
mance. Finally, it is important to stress that the complexity of the diagnosis further
increases on machinery with high levels of power.
Traditional diagnostic methods are vibration analysis [2], [3], [4] and oil analysis
[5], which can be very expensive and sometimes even impossible to carry out due to the
physical inaccessibility of some assets. Therefore, these tests are only made for critical
assets,leaving out many others assets that could present failures evolving faster. An
easy way to start monitoring these assets is using global vibration measurement at
specific points on an ongoing basis and following criteria such as those established in
the ISO 10816-7 standard (See Fig 1). This regulatory framework presents four health
states ranging from green(good) to red(dangerous) according to the vibration severity.
However, this kind of evaluation involves the human factor might have a significant
negative impact on such evaluation.

Fig. 1: Vibration Evaluation Standard Table for Reciprocating Machinery. According

to vibration levels in displacement, velocity, and acceleration for seven machine clas-
sifications based on size, construction, and assembly.

2
 On the other hand, in automatic diagnosis, we have as one reference the use of
convolutional neural networks (CNN) that are ideal for solving classification problems
efficiently [6]. In this paper, we evaluated the possibility of mixing traditional diagnosis
using vibration analysis and machine learning to realize a routine part of the process
at condition-based maintenance.
Having examined pertinent literature in Section II, which explores current meth-
ods for diagnosing asset conditions, we proceed to delineate our methodology and
contributions in Section III. We continue in Section IV, evaluating the effectiveness
of our approach in precisely executing the diagnostic process for assets. Finally, the
conclusion section underscores the principal findings of this study.

2 Related work
The landscape of machine fault diagnosis in the state of the art is not uniform, as the
literature classifies studies based on the specific machine types. Furthermore, there’s
a distinct categorization based on the class of machine learning techniques employed:
those relying on features and those utilizing deep learning. For instance, in [7] and
[8], it’s evident that the authors under review lean significantly towards traditional
machine learning methods, where feature extraction techniques constitute the core of
their proposals. However, these are tailored for different types of machines.
In [7], the authors found that up until 2018, the SVM method (45.7%) and ANN
(43.5%) were the most prevalent techniques in the existing literature for diagnosing
faults in reciprocating compressors. In contrast, deep learning (DL) had limited adop-
tion (4.4%). Meanwhile, in [8], a similar investigation was conducted, but this time
focusing on fault diagnosis in rotating machinery.
On a different note, more recent reviews [9],[10], [11] demonstrate an increasing
trend in the utilization of deep learning theories as fundamental tools for constructing
end-to-end diagnosis procedures. In [9], the authors argue two reasons for this increase:
the human dependence to feature extraction tasks which adversely affects the per-
formance of the diagnosis models, and the increasingly grown data due to the rapid
development of internet technologies and the Internet of things which is not appro-
priate for the use with traditional machine learning. Furthermore, they reviewed four
typical deep learning methods applied in machine fault diagnosis: stacked autoencoder
(Stacked AE), deep belief network (DBN), convolutional neural networks (CNN), and
residual net (ResNet). Authors concluded that DL-based methods could obtain high
diagnosis accuracy compared with traditional machine learning, subject to training
with sufficient labeled samples. But, the collected data for training are realistically
imbalanced because it is easier to collect healthy data than faulty data. Lastly, in [12]
and [13], the authors showed transfer learning as an alternative solution to overcome
such weaknesses.
To our knowledge, if we focus on the literature regarding diagnosis faults in recip-
rocating machinery, there are also specific resources for different types of reciprocating
machinery. There is little literature on diagnosis faults in 4-stroke combustion engines,
despite being one of the most commonly used machines in the industry. These engines

3
can fail unexpectedly due to the difficulty in diagnosing the health of their compo-
nents, which has many moving elements, where the most critical point of its operation
and performance is the power head, where the elements that comprise it perform the
energy conversion.
Costs associated with the traditional methods used for diagnosis are quite high,
which leads to a reduction or, in some cases, the complete elimination of PdM; this
leaves no option but corrective maintenance.
Considering that in the last few years, new diagnosis faults methods using deep
learning have been coming, we propose a low-cost system that enables pre-diagnosis
by utilizing audio signals related to the vibration of these machines. This approach
will lower the costs associated with commonly used yet expensive tests, such as oil
and vibration analysis.
In order to validate our study, we propose a case study involving two 4-stroke
internal combustion engines, where we focus on the elements of the power head and
their corresponding failure modes, particularly those related to piston cylinders and
bearings. Moreover, as a countermeasure to address the inherent dataset imbalance
common in such studies, we introduce a web application. This application facilitates
the execution of our automated diagnostic model, empowering individuals to not only
expand our dataset but also benefit from preliminary equipment pre-diagnosis.

3 Methodology

Fig. 2: Proposed architecture for the automated diagnostic process. The vibration
data is carefully pre-processed and subsequently transformed into spectrograms. These
spectrograms feed into our Convolutional Neural Network (CNN), enabling it to clas-
sify the motor’s condition accurately.

Figure 2 shows our architecture for classifying reciprocating machines health.

This architecture takes vibration data and converts it into spectrograms parame-
terized according to the operating frequencies of the motor elements.
The spectrograms feed our CNN classifier, which extracts the most important
features to classify them within the three category labels: Ok for engines in good
condition, Monitoring for engines with suspected defect and Caution for engines with

4
some kind of defect. The main reason why the spectrogram obtains better results in
the identification of robust characteristics is the fact that it can identify the behavior
of the frequency harmonics. This fact has already proven to solve successfully natural
language processing problems [14].
Importantly, proper training of the CNN classifier guarantees the successful func-
tioning of the architecture proposed. Below, we outline the creation of the dataset
process and the methodology used to reach an accurate labeling.

3.1 Dataset and data labeling

We obtain the dataset used to train our classifier from a series of vibration tests on
two 4-stroke, 97 cm3 engines under various conditions. The first engine contributes
approximately 90% of the data and has been used for 50,000 km. The second engine
provides around 10% of the data and has been used for nearly 300,000 km, showing
more noticeable wear. We performed two tests on the powerhead before maintenance
and seven after maintenance on the youngest engine. The data set covers a wide range
of scenarios and conditions to ensure robustness and generalization of the trained
model.
The next critical step was to label the dataset accurately. We combined manual and
automated labeling methods to ensure precision and reduce human errors. Appropriate
well-designed labeling is hugely important to get a better classification output. The
following is a description of the most significant processes used to achieve the correct
labeling of our dataset.
To label our dataset, we started using an experimental validation process as shown
in Fig. 3.

FEA Metrology

Vibration Theoretical STFT ISO 10816

Waveform
signal signal
FFT
Time MCI
synchronous Power FFT
Audio file Filters Spectrogram MCA
average Envelope
Natural
Time frequency Time Frequency
Domain estimation domain domain Diagnosis

Input Pre-processing Processing Analysis

Fig. 3: Initial architecture of the diagnostic process used for labeling, detailing the
methods and tools employed.

In the first block, we acquire a vibration signal that is the input to the prepro-
cessing block. Here we denoise a vibration signal by using a series of successive filters
parametrized according to the finite element analysis model for the engines (See Fig.
4).

5
Fig. 4: List of parts and simplified finite element analysis (FEA) model of the system
under evaluation, enabling the identification of the types of mechanical stresses that
affect the components.

In the processing block, we use several mathematic tools to analyze time, frequency,
and time-frequency to obtain spectrograms. Lastly, we use our strong knowledge base
about internal combustion engines, mechanical vibration analysis, and the norm ISO
10816 to interpret the plots estimating the vibration severity. Table 1 shows an example
of how we calculate velocities and their respective operating frequencies, which are
black highlighted.

Table 1: Gear shifting patterns - speed at shift points and corresponding

rpm / Hz
4-speed Manual Transmission Engine 1.
Gear z1 z2 z3 z4 z5 z6 z7 z8 z9 z10 z11 z12
R.P.M 1880.6 2279.1 716.3 1336.0 1840.8 2378.2
Hz 31.3 38.0 11.9 22.3 30.7 39.6
RPM Out 227.9 425.1 585.7 756.7
km/h 19.6 36.6 50.5 65.2

During vibration tests, we conducted measurements of pressure, sound, velocity,

and temperature. We also performed some oil analyses to ensure accurate labeling and
classification. The testing methodology used Zone D in the ISO 10816-7 chart (see
Fig. 1) to pre-categorize the collected signals based on the overall vibration value.
From Table 1, we get the forcing frequencies expected in the spectrum and make
the correlation with Table ISO 10816 to obtain a first estimate of the severity of the
engine vibration. We further employ these operational frequencies to establish suitable
spectral bands in the spectrogram, optimizing the recognition of robust features.
It is important to clarify that, despite the nature of the data that allows the
use of spectrogram representation, it alone is insufficient for labeling due to the
classification problem being dependent on the operational context and associated non-
deterministic phenomena that complicate straightforward diagnosis. To address this

6
issue, we designed an algorithm that identifies and quantify the number of harmon-
ics of the fundamental frequency in the signal. This algorithm formulates a series of
methods to align each input signal with its corresponding diagnosis. Table 2 provides
a summary of the additional calculations associated with quantifying the effects of
these frequencies.

Table 2: Bearing fault frequencies calculations.

Type of defect1 Equation2 Hertz
Shaft speed frequency h  i 50.000
Nb
Inner race defect frequency BP F I = 2
× S × (1 + Bd × cos(θ) 221.302
h Pd i
Nb Bd
Outer race defect frequency BP F O = 2 × S × (1 − Pd × cos(θ) 128.698
h  i
Cage defect frequency F T F = S2 × (1 − Bd × cos(θ) 18.835
Pd  2 
Ball spin frequency BSF = 2Bd × S × (1 − Bd
Pd
Pd
× cos(θ) 87.874
Rolling element defect frequency 175.747

1 The analysis involves calculating the bearing failure frequencies at the equipment’s ideal oper-
ating rotor speed of 3000 rpm (50 Hz).
2 Nb=Number of rolling elements Pd=Pitch diameter θ=Contact angle Bd=Ball diameter
S=Shaft speed

Using this data and the technical information related to the vibration sources, we
continue designing the algorithm required for automating the diagnosis. It’s crucial to
treat the machine as a black box so that the system can identify robust characteristics
in the signals that an analyst might overlook.
After conducting numerous tests, we identified the appropriate parameters for
spectrogram configuration, enabling the best recognition of robust characteristics. In
summary, with vibration data collected at the suitable sampling frequency and for
a relevant duration, the spectrogram focused on a specific frequency band provides
the graphical representation that feeds a convolutional neural network. This network
extracts the most important features to classify the signals into the three category
labels.

4 Results

The successful application of spectrograms in voice recognition systems demonstrated

its potential for identifying frequency harmonics’ behavior in signals. Harnessing this
potential, we embarked on the creation of an initial stage employing a CNN model
built from the ground up. For the model development, we utilize Tensorflow 2.4.1 and
adjust the spectrogram images to a size of 128x128 pixels. We categorize the training
images into three classes and then split the data into a 66% training set and a 33%
validation set.
However, the obtained results did not meet the desired level of stability, only achiev-
ing about 40% precision, showing a lower level of generalization, as evidenced in Table

7
3. In addition, the signal classification process revealed an imbalance of categories,
primarily caused by the scarcity of signals obtained from machines in poor real-life
conditions. Typically, most diagnoses fall into the Ok or Monitoring categories.
Consequently, the subsequent stages of the process involved iterative optimization
of various parameters to achieve the best possible performance.

Table 3: First stage: results obtained

with our CNN model from scratch
Parameters Precision recall f1-score
OK 0.67 0.50 0.57
Monitoring 0.50 0.38 0.43
Caution 0.43 0.75 0.55

accuracy 0.50
macro avg 0.50 0.54 0.52
weighted avg 0.52 0.50 0.49

As the achieved results fall short of the expected performance, we proactively

incorporate two additional concepts from our thorough review of the state-of-the-art:
transfer learning and operational context.
32 Neurons

16 Neurons Ok
LeakyRelu

LeakyRelu

LeakyRelu
8 Neurons
Conv1-1
Conv1-2

Conv5-1
Conv5-2
Conv5-3
Pooling

Pooling

Dense1-

Dense2-

Dense3-
... Monitoring

Caution

Convolutional Layers Dense Layers Output Layer

Fig. 5: Model Stage II, using as base a network VGG16 previously trained with the
ImageNet dataset

By employing transfer learning, we successfully address the overfitting issue. We

use the pre-trained VGG16 network with ImageNet as our base model and augment
it with three dense layers having 32, 16, and 8 neurons, respectively (See Fig. 5). All
these layers employ the LeakyReLU activation function with an alpha value of 0.3 for
the first two layers and 0.1 for the last layer. The obtained results, presented in Table
4, demonstrate a noticeable enhancement in the performance of our classifier.
During the training process, we observed that depending only on the Accuracy
metric did not adequately reveal the effects of changes in the input signal configura-
tion, particularly in pre-processing. Consequently, we determined crucial to include
additional performance metrics like Recall, F1-Score, and Precision.
On the other hand, we acknowledged that factors such as environmental conditions,
workload, and usage patterns exerted a pronounced influence on the performance and

8
 Table 4: Comparison of Results for Models stage I and II
Model Parameters Precision Recall F1-score
OK 0.67 0.50 0.57
Model I Monitoring 0.50 0.38 0.43
Caution 0.43 0.75 0.55
OK 0.67 ↑0.67 ↑0.67
Model II Monitoring ↑0.83 ↑0.83 ↑0.83
Caution ↑0.67 ↓0.67 ↑0.67

Average (Model I) Macro Avg 0.50 0.54 0.52

Weighted Avg 0.52 0.50 0.49
Average (Model II) Macro Avg ↑0.72 ↑0.72 ↑0.72
Weighted Avg ↑0.78 ↑0.78 ↑0.78

Accuracy (Model I) ... ... ... 0.50

Accuracy (Model II) ... ... ... ↑0.78

behavior of the machines. Considering that the accuracy of training results heavily
relies on the skilled classification of complex signals, a vibration analyst undertakes this
intricate duty by employing structured observation methods to heighten the certainty
of diagnostics. These systematic observational processes align with the concept of
operational context.
Given this awareness, we formulated a system to gather data concerning sup-
plementary variables. Within the domain of machinery diagnostics, these variables
constitute what is termed the operational context.
Figure 6 presents the system’s components, which are valuable for collecting infor-
mation from external databases. We added an Internet of Things (IoT) arrangement
to this framework, which collected data and enhanced the classification process.

Fig. 6: Acquisition and diagnosis scheme.

The operational context’s impact becomes evident in improved image classification

(see Table 5). Incorporating the operational context yielded enhanced accuracy within

9
the Caution category, even in light of the observed data imbalance. Moreover, this
incorporation circumvented the necessity of introducing additional signals.

Table 5: Comparison of Results for Models stage II and III

Model Parameters Precision Recall F1-score
OK 0.67 0.67 0.67
Model II Monitoring 0.83 0.83 0.83
Caution 0.67 0.67 0.67
OK ↑1.00 ↑1.00 ↑1.00
Model III Monitoring ↑1.00 ↓0.60 ↓0.75
Caution 0.67 ↑1.00 ↑0.80

Average (Model II) Macro Avg 0.72 0.72 0.72

Weighted Avg 0.52 0.50 0.49
Average (Model III) Macro Avg ↑0.89 ↑0.87 ↑0.85
Weighted Avg ↑0.91 ↑0.87 ↑0.86

Accuracy (Model II) ... ... ... 0.78

Accuracy (Model III) ... ... ... ↑0.87

The impact of the operational context is evident in the improved spectrogram

image classification (see Table 5). Remarkably, despite the data imbalance, we
improved the accuracy within the Caution category. In addition, this incorporation
avoided the need to introduce additional synthesized signals.
We can observe the operational context’s effects in the classification results for
the set of 25 test images captured with the engine running, which enhancing with the
quantification of the total harmonic component of the signal (THC), which ranges
between 82.5% and 90%, significantly improves the performance. It is noteworthy that
at certain rpm (Test 712) with the engine in second gear, the THC exceeds that of
first gear(Test 702), and in fourth gear, where the engine runs at a lower rpm (722),
the difference is notorious (see Fig.7).

10
 Fig. 7: Operational context and algorithm effect.

Furthermore, recognizing that a signal influenced by resonance can lead to an

inaccurate interpretation by our classifier, we adjusted our model (See Fig. 8).
32 Neurons

16 Neurons

Ok
LeakyRelu

LeakyRelu

LeakyRelu
8 Neurons
Resonance

Conv1-1
Conv1-2

Conv5-1
Conv5-2
Conv5-3
Pooling

Pooling
filtering

Dense1-

Dense2-

Dense3-

... Monitoring

Caution

Pre-processing Convolutional Layers Dense Layers Output Layer

Fig. 8: Model Stage III, showing how a filtering stage has been added in the pre-
processing block to exclude signals affected by the resonance phenomenon.

In doing so, we employed an algorithm that quantifies the harmonic activity of the
signals and excludes those affected by resonance phenomena (see Algorithm. 1).
The Table 5 showcases the robust capability of a CNN in identifying features,
enabling accurate categorization of vibration signals into three distinct classes.
Notably, the follow-up and normal categories achieved a precision rate of 100%,
while the precaution category exhibited precision exceeding 80% across 162 reference
tests. This performance is particularly commendable, considering the restricted signal
dataset and the inherent category imbalance.
Finally, we have successfully developed a fully functional prototype that utilizes
web-based tools to generate diagnostics, store and manage information, and present
results in real time through online platforms. This system incorporates state-of-the-art

11
Algorithm 1 harmonics
1: function harmonics(s, F s)
2: (spec, freq) ← calculate magnitude spectrum(s, F s)
3: (f 0, af ) ← find fundamental frequency(spec, freq)
4: vf ← find highest peaks(spec, freq, 40)
5: hc ← 0
6: for i ← 0 to 39 do
7: v ← find highest peak index(spec)
8: if spec[v] > spec[v − 1] and spec[v] > spec[v + 1] then
9: append(vf, freq[v])
10: hc ← hc + spec[v] ▷ Harmonic component
11: end if
12: end for
13: thc ← 100 × (hc − af )/hc
14: return thc
15: end function

techniques for validating data and authenticating users. Using the framework Flask,
we have deployed our application, naming it PREDIWEB, which seamlessly inte-
grates technologies like Tensorflow and Cuda to execute complex artificial intelligence
models. To ensure robust cybersecurity measures, we’ve established an authentication
gateway using the Flask SQLAlchemy V.2.4.4 module for database management and
Flask Login V.0.5.0 for secure access.
At its core, the prototype features a module known as PREDITEST. Its singular
input involves users uploading an image that encapsulates a meticulously processed
vibration signal aligned with specified standards. The corresponding output furnishes
a predictive analysis regarding the equipment’s health status, as inferred from the
loaded signal.
It’s important to clarify that image diagnostics occur on the backend or server
side, with the resulting insights conveyed to the front end. This frontend interface then
translates the data into an intuitive graphical representation, vividly depicted in Fig.
9.

5 Conclusions
Developing a Deep Learning model for predictive maintenance and condition moni-
toring of reciprocating machinery poses a significant challenge. This challenge arises
due to the intricate machinery, the data’s characteristics, and the inherent data
imbalance. Moreover, the data imbalance exacerbates due to the substantial disparity
in data volume between machines encountering failures and those functioning under
normal conditions. However, following our methodology, we can develop a robust
framework for a functional application. Is possible enhance the system, by fine-tuning
hyperparameters to achieve optimal performance.

12
 Fig. 9: Automated diagnosis carried out in WEB application.

Reducing uncertainty for generating reliable diagnoses involves leveraging perti-

nent process parameters. The utilization of IIOT technologies is highly advantageous
in this regard. While not individually sufficient, these technologies prove invaluable
when combined with enriched signals and other parameters. This synthesis produces
a range of relationships that unveil distinct behavioral patterns. Through iterative
refinement of this process is futher the training of an artificial intelligence model.
Eventually, this model becomes adept at identifying abnormal operational conditions
or issues linked to disruptions like resonance and bending modes of external elements,
which are atypical for the combustion system within an internal combustion engine.

Is necessary a multidisciplinary approach to develop full automatic diagnostics

because process engineers provide domain expertise, and data engineers enable the
neural network to leverage that expertise and make accurate diagnoses, even with
limited information. This collaborative effort is crucial for building a robust and
effective diagnostic system for reciprocating machinery.

References
[1] Aguilar-Otero, J.R., Torres-Arcique, R., Magaña-Jiménez, D.: Análisis de modos
de falla, efectos y criticidad (amfec) para la planeación del mantenimiento emple-
ando criterios de riesgo y confiabilidad. Tecnologı́a, Ciencia, Educación 25(1),
15–26 (2010)

[2] White, G.: Introduction to machine vibration. USA, Azima DLI, 22–29 (2010)

[3] Hernández-Montero, F.E., Gutiérrez-Garcı́a, M.: Analysis of encircling to the

prosecution of vibrations for the diagnosis of. Ingenierı́a Mecánica 13(1), 31–40
(2010)

13
 [4] Matienzo, J.R.: Diagnotic by vibrations in motor of internal combustion. Inge-
nierı́a Mecánica 1(1), 11–26 (1998)

[5] Carpio, L.B., Granizo, V.G.: The use of using oil analysis for early detection of
faults in diesel internal combustion engines. Ciencia Unemi 8(15), 84–95 (2015)

[6] Chatfield, K., Simonyan, K., Zisserman, A.: Efficient on-the-fly category retrieval
using convnets and gpus. In: Computer Vision–ACCV 2014: 12th Asian Confer-
ence on Computer Vision, Singapore, Singapore, November 1-5, 2014, Revised
Selected Papers, Part I 12, pp. 129–145 (2015). Springer

[7] Lv, Q., Yu, X., Ma, H., Ye, J., Wu, W., Wang, X.: Applications of machine
learning to reciprocating compressor fault diagnosis: A review. Processes 9(6)
(2021) https://doi.org/10.3390/pr9060909

[8] Liu, R., Yang, B., Zio, E., Chen, X.: Artificial intelligence for fault diagnosis of
rotating machinery: A review. Mechanical Systems and Signal Processing 108,
33–47 (2018) https://doi.org/10.1016/j.ymssp.2018.02.016

[9] Lei, Y., Yang, B., Jiang, X., Jia, F., Li, N., Nandi, A.K.: Applications of machine
learning to machine fault diagnosis: A review and roadmap. Mechanical Systems
and Signal Processing 138, 106587 (2020) https://doi.org/10.1016/j.ymssp.2019.
106587

[10] Hoang, D.-T., Kang, H.-J.: A survey on deep learning based bearing fault diag-
nosis. Neurocomputing 335, 327–335 (2019) https://doi.org/10.1016/j.neucom.
2018.06.078

[11] Zhao, J., Yang, S., Li, Q., Liu, Y.-q., Gu, X.Q., Liu, W.: A new bearing fault
diagnosis method based on signal-to-image mapping and convolutional neural
network. Measurement 176, 109088 (2021)

[12] Liu, Q., Huang, C.: A fault diagnosis method based on transfer convolutional
neural networks. IEEE Access 7, 171423–171430 (2019) https://doi.org/10.1109/
ACCESS.2019.2956052

[13] Xu, G., Liu, M., Jiang, Z., Shen, W., Huang, C.: Online fault diagnosis method
based on transfer convolutional neural networks. IEEE Transactions on Instru-
mentation and Measurement 69(2), 509–520 (2020) https://doi.org/10.1109/
TIM.2019.2902003

[14] Pfister, T., Simonyan, K., Charles, J., Zisserman, A.: Deep convolutional neural
networks for efficient pose estimation in gesture videos. In: Computer Vision–
ACCV 2014: 12th Asian Conference on Computer Vision, Singapore, Singapore,
November 1-5, 2014, Revised Selected Papers, Part I 12, pp. 538–552 (2015).
Springer

14