See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/352155128

Deep Learning for Data-Driven Predictive Maintenance

Chapter in Intelligent Systems Reference Library · June 2021

DOI: 10.1007/978-3-030-75490-7_3

CITATIONS READS

6 1,872

3 authors:

Muhammad Sohaib Shiza Mushtaq

Zhejiang Normal University Lahore Garrison University
35 PUBLICATIONS 607 CITATIONS 4 PUBLICATIONS 74 CITATIONS

SEE PROFILE SEE PROFILE

Jia Uddin
Woosong University
188 PUBLICATIONS 1,368 CITATIONS

SEE PROFILE

All content following this page was uploaded by Jia Uddin on 14 February 2022.

The user has requested enhancement of the downloaded file.

 Deep Learning for Data-driven Predictive Maintenance

Muhammad Sohaiba, Shiza Mushtaqa, Jia Uddinb

a
Department Computer Science, Lahore Garrison University, Lahore, Pakistan
b
Technology Studies, Endicott College, Woosong University, Daejeon, Korea

Contents
Abstract ................................................................................................. 3

Introduction ........................................................................................... 4

2. Maintenance ...................................................................................... 4

2.1 Predictive maintenance ................................................................... 6

2.1.1. Data Collection ........................................................................ 8

2.1.2. Fault Detection and Diagnostic................................................ 8

2.1.3. Prognostic ................................................................................ 9

3. Sensors commonly used for Predicative maintenance ...................... 10

4. Sensor Data Analysis and Machine learning...................................... 12

4.1. Logistic Regression........................................................................ 13

4.1.1. Binary or Binominal Logistic Regression ................................ 13

4.1.2. Multinomial Logistic Regression ............................................ 13

4.1.3. Ordinal Logistic Regression .................................................... 13

4.2. k-Nearest neighbors ..................................................................... 13

4.3. Artificial Neural Network .............................................................. 15

2

4.4. Support Vector Machine ............................................................... 16

5. Limitations of Machine learning Algorithms used for Predictive

Maintenance ........................................................................................ 18

5. Deep Learning Methods ................................................................... 19

5.1. Deep Artificial Neural Network ..................................................... 19

5.2. Deep Convolutional Neural Network ............................................ 20

5.3. Deep Recurrent Neural Network .................................................. 22

................................................................. Error! Bookmark not defined.

5.4. Deep Auto-encoders ..................................................................... 23

5.5. Deep Belief Network ..................................................................... 25

6. Advantages of Deep Learning in Predictive maintenance .................. 26

7. Conclusion and Future Perspectives of Deep learning based Data-

Driven Predictive maintenance. ........................................................... 29

7.1. Conclusion .................................................................................... 29

7.2. Future Perspectives of Deep learning based Data-Driven Predictive

maintenance. ....................................................................................... 30

7.2. 1. Enhanced Generalization ...................................................... 30

7.2.2. Explainability.......................................................................... 30

7.2.3. Multimodal and Multi-Sensors Data Fusion .......................... 31

Glossary ............................................................................................... 31

Condition Monitoring: It is the constant monitoring of the parameters

from a machine which are associate with its health indication. Any
 3

significant fluctuation in the monitored values of the parameters

indicates the deviation of the machine from its normal state. .......... 31

Sensor: It is a device that takes a physical reading from the

environment and converts it into electrical data or signals. ............. 31

References ........................................................................................... 32

Abstract
With the advancement of technology, it is viable to obtain data from sen-
sors in real-time from industrial equipment, engines, heavy machines, and
manmade structures. The collected data in real-time can be further utilized
to perform maintenance of assets before these entities get entirely worn-
out which in technical terms known as predictive maintenance. The down-
time of dilapidated instruments can lead to loss of revenues and can be a
threat to the workers on the facilities. In time and precise prediction of
such failures using the data acquired through sensors can alleviate the
downtime of the equipment, hence, helps in cutting off the revenue losses
and ensure the safety of the workers. If enough historical data associated
with equipment is available, then artificial intelligence techniques like ma-
chine learning as well as deep learning may be utilized for identification of
equipment failures in parts or as a whole. Deep learning algorithms have
shown profound progress in the problem areas where practitioners and re-
searchers had been eluded for several decades. This chapter provides an
insight into the deep learning algorithms used for predictive maintenance.
It also provides an overview of industrial sensors and future research as-
pects of sensors using techniques of deep learning for predictive mainte-
nance.
4

Introduction
A sensor is a device that converts physical signals which are measured from
the environment into electrical signals [1]. It is possible to repeatedly meas-
ure and store the physical quantities from the environment. The stored
measurements can be utilized to study the behavior of these physical quan-
tities. Moreover, a realistic system can also be devised to analyze the meas-
urements taken from sensors in real-time to gain valuable insights and
make important decisions.

Certain internal and external physical quantities are affected and their
reading change during the operation of an asset due to the change in the
working environment. These physical quantities include but not limited to
internal oil temperature, and pressure, external temperature, and humid-
ity, etc. Continuous monitoring of these variables to identify the drift of
equipment from its normal condition and taking measures to evade them
is referred as predictive maintenance [2].

Predictive maintenance provides reduced downtime of an asset, enhanced

quality, limiting the revenue losses, reliability, and better safety of workers.
In addition, the provision of early warnings to avoid catastrophic outcomes
by recognizing uncommon behavior of equipment is the most important
goal of predictive maintenance.

In the past few decades, data-driven predictive maintenance has been ef-
ficacious due to the improvised data acquisition techniques, application of
different types of sensors, machine learning and deep learning, and availa-
bility of big data.

In summary, taking measurements through various types of sensors,

storage of these measurements, data pre-processing, and analysis of the
pre-processed data defines the complete predictive maintenance process
[3].

2. Maintenance
The failure of industrial equipment, heavy machines, engines, and struc-
tures leads to unwanted downtime, enough economic damages, and put
 5

the worker’s safety at risk. These issues can be mitigated if the mainte-
nance of such assets is performed on a timely basis. In general, mainte-
nance of an object can be defined as, servicing, functional checks, repairing
or replacing of required components, machinery, artificial infrastructures,
and supporting utilities in industrial, residential, business installations. In
other words, the maintenance process can be defined as maintaining the
equipment by troubleshooting problems either manually or through com-
puterized diagnostic tools [4].

In the context of manufacturing and processing industries, the process of

maintenance can be broadly categorized as reactive maintenance, preven-
tive maintenance, predictive maintenance alternatively called condition-
based maintenance, and proactive maintenance [5] [6].

In reactive maintenance, the ineffective or damaged components of an as-

set are repaired or replaced so that it may continue its function smoothly.
Therefore, no need to take preventive measures and problem is tackled
when it is reported. On the other hand, preventive maintenance is a sched-
ule for periodic inspection of equipment. The main goal of preventive
maintenance is to identify small problems as the equipment undergoes de-
terioration and addressing these issues before its complete collapse. Its
main advantage is that the equipment under observation does not breaks
down as the defected part is replaced with a new one on time. So, the
equipment makes from one planed service to another without any failure
due to fatigue, neglect, or wear. A few of the preventive maintenance ac-
tives include oil change, partial or complete work halts at specified inter-
vals, use of lubricants, minor adjustments, worn parts replacement, etc.
The objective of preventive maintenance is to take equipment from one
planned operation to another without any failures. In this, the mainte-
nance activities include partial or complete work halts after specified inter-
vals, oil change, lubrication use, minor adjustment, and so on. Further, the
workers can make a log of equipment wear and tear, so it is easy to replace
the defected part with a new one easily without failure of the overall sys-
tem.

Strategies of Predictive maintenance are adopted to identify the condition

of equipment in service with the aim to estimate when the maintenance of
equipment will be necessary. It helps in the reduction of maintenance costs
6

as the maintenance is just carried out when it is required. Thus, it is also

termed as condition-based maintenance because it is done when condition
of the item is degraded. The hallmark point of predictive maintenance is
that it allows scheduling of correcting activities and hence prevents sudden
machinery failures. Key features of predictive maintenance are to provide
correct information regarding lifetime of the equipment, enhanced plant
safety, few accidents with increased environment protection, and good
spare parts handling.

2.1 Predictive maintenance

In predictive maintenance condition and performance monitoring of equip-
ment during normal operation is performed to reduce the likelihood of fail-
ures. It is also known as condition-based maintenance because the equip-
ment is continuously monitored to timely detect and identify its anomalous
behavior. Predictive maintenance is an old field, but its history is not
properly documented. However, it is speculated that it is operational in the
industrial world since the 1990s.

Predictive maintenance aim is to predict a failure of equipment well before

its actual failure, which is then followed by preventive measures such as
regularly scheduled and corrective maintenance. The process of predictive
maintenance is given in Fig 1.
 7

Fig 1. Predictive maintenance process.

One of the key aspects of predictive maintenance is the continuous moni-

toring of equipment so that its optimal working can be ensured. Therefore,
it is appropriate to say that without condition monitoring predictive
maintenance is not feasible. Condition monitoring as three categories: pe-
riodic, online, and remote. In periodic condition monitoring behavior of
equipment under observation is examined after fixed intervals. The peri-
odic analysis of the data accumulated from the equipment can depict a cer-
tain trend that can define its health state. The benefit of periodic Condition
Monitoring is the regular inspection of the devices. If something is wrong,
it is detected and identified in time. Online condition monitoring is the uti-
lization of measurement hardware and software to continuously monitor
machines or production processes. Lastly, as the name suggests, remote
condition monitoring is the monitoring of equipment from a remote loca-
tion, with data transmitted for analysis.

Real life scenarios of predictive maintenance are:

• Recognizing that when fan get unbalanced.

8

• Recognizing when bearings demand lubrication.

• Alarming moment indication when oil replacement or when oil is
contaminated.
• Finding misalignment between two rotating pieces of equipment.

Further, predictive maintenance can be of two types, i.e., Data-driven

maintenance, and based on Model maintenance. Here the topic of discus-
sion would be data-driven maintenance of industrial installations. The pro-
cess of data driven-maintenance consists of four crucial stages, i.e., data
collection, fault detection and diagnostic, and prognosis. If properly imple-
mented and executed, predictive maintenance is the foundation of an effi-
cacious maintenance program.

2.1.1. Data Collection

To perform data-driven predictive maintenance, data associated with
health states of equipment under observation is vital. It can be obtained
through the data collection process in which information regarding target
variables is gathered and measured in an established system. The gathered
information can then be analyzed to answer questions related to the health
states of the equipment and evaluate outcomes. Data collection is an im-
portant constituent of research in various disciplines including but not lim-
ited to humanities, physical and social sciences, engineering, computer sci-
ence, and business. Further, data collection techniques may change
depending upon the discipline. Nevertheless, the data collection process
may vary but the significance of correct and honest data collection remains
the same. In terms of data-driven predictive maintenance, data is collected
via different types of sensors, i.e., temperature, accelerometer, acoustic,
oil, etc. the measurements recorded through the sensors are further ana-
lyzed through signal processing, machine learning and deep learning tech-
niques to infer momentous statistics about the conditions of the equip-
ment.

2.1.2. Fault Detection and Diagnostic

Fault detection is the identification of anomalous patterns in the recorded
data associated with the equipment under study. Fault detection does not
just involve breathing the threshold level but understanding about the
 9

dynamics of the environment in which the equipment is operational and

contextualizing the problem. It not only alerts the operators that there is a
spike in the readings of given measurements but also indicates that a com-
ponent is out of order.

Once the anomalies are being identified, the data can be further analyzed
to pinpoint the type and location of fault which is known as fault diagnosis.
Hence, the overall process of Fault Detection and Diagnostics (FDD) isolates
anomalies and identifies the types of problems in the performance of es-
sential equipment required in the manufacturing, and processing indus-
tries such as motors, boilers, chillers, motors, elevators, pumps, exhaust
fans, etc. The application of FDD is beyond the manufacturing and pro-
cessing industries. For instance, it can be applied to detect and identify
anomalies in the man mad structures including buildings and bridges or to
detect problems in the water, oil, and gas distribution networks, etc. More-
over, recent developments have empowered FDD systems to render anom-
alies into real-world faults and generate alerts to operators listing details
about the failure of a component and how to resolve the problem.

2.1.3. Prognostic
Prognostics is the prediction of remaining useful time of equipment. Its
purpose is to predict the time after that the complete system or one of its
components will not give its desired performance [1]. The loss of smooth
operation of a system is mostly assumed to the failure of the whole system
or its components. Therefore, such a system can no longer be used to
achieve the desired performance. The predicted time is known as the re-
maining useful life (RUL), which plays an important role for deciding allevi-
ating failure. Prognostics can also be defined as predicting the future per-
formance of equipment by evaluating the degree of deviance or
deterioration of a system from its baseline operating conditions [2]. The
concept of prognostics depends upon the exploration time to start progno-
sis (TSP), an examination of failure modes, detection of wear and aging of
the components in the incipient stage, and fault conditions. Prognostic of
a system can be helpful and effective if ample knowledge about the root
cause of failure in the system is known. Therefore, it is essential to gather
all the vital information on the possible failures in a product that includes
10

information about the site, mode, cause, and mechanism of failure. This
information is helpful to monitor and optimize the targeted parameters of
a system. Potential uses for prognostics is in predictive maintenance. The
field of studies that relates failure mechanisms to the lifecycle of a system
is denoted as prognostics and health management (PHM). Prognostics
techniques can be divided into data-driven, model-based, and hybrid ap-
proaches.

3. Sensors commonly used for Predicative mainte-

nance
There are different types of sensors that can be used for predictive
maintenance of industrial equipment. This section provides insight into dif-
ferent kinds of sensors used for predictive maintenance, for instance, tem-
perate, oil, accelerometers, acoustics emission sensors, and microphone.
The selection of sensors is application-specific in which some sensors can
be effective as compared to others. For instance, some sensors can detect
bearing faults at the inception, much earlier as compared to others. This
section covers different types of sensors commonly used in the predictive
maintenance of assets. It is imperative to first confirm the type of potential
fault in a system that may be encountered as predictive maintenance sys-
tems often utilizes a limited number of sensors. For instance, types of sen-
sors that can detect fault signatures from a rotary machine are given in Fig
2. Some of the commonly used sensors in predictive maintenance are listed
in Table 1 along with their specifications and types of faults they can detect.
 11

Fig 2. A machine life cycle vs types of sensor that can detect fault signals
[7][8].

Table 1. Types and specification of sensors used in predictive mainte-

nance [7][8].

Measurement Type Sensor Frequency Application

Integrity testing of metallic

structures, composite mate-
rials. Crack detection in hot
Acoustic Emissions Piezo Acoustic Emission 100-400 kHz reheat piping system. Plants
and woods drying process
monitoring, partial discharge
detection.

Corrosion detection in stor-

age tanks (flat bottom), oil
Acoustic Emissions Piezo Acoustic Emission 20-100 kHz
and water pipelines leakage
detection, concrete crack de-
tection, and partial discharge
12

detection in low noise sce-

nario.

Vibration Piezo accelerometer Up to 30 kHz Rotating machines.

Vibration MEMS accelerometer Up to 20 kHz Rotating machines.

Pressure leaks, rotating ma-

Sound Pressure Ultrasonic microphone Up to 20 kHz
chines.

Pressure leaks, rotating ma-

Sound Pressure Microphone Up to 100 kHz chines and gearbox fault di-
agnosis, pump cavitation.

Magnetometer, Hall, Rotor bar and end ring prob-

Magnetic Field
Search Coil lems.

Heat source identification,

change in load conditions, ir-
Temperature Infrared Thermography -
regular turn off, power sup-
ply issues.

Heat source identification,

change in load conditions, ir-
Temperature Thermocouple, RTD -
regular turn off, power sup-
ply issues.

Oil quality Particle Monitor - Debris detection.

4. Sensor Data Analysis and Machine learning

Machine learning is a subfield of artificial intelligence where algorithms
learn from input data without being explicitly programmed. These algo-
rithms can automatically learn input data which is then used for predicting
the out-of-sample data. This property makes a machine learning algorithm
useful for data-driven predictive maintenance of industrial equipment. This
section provides an overview of the popular machine learning algorithms
that have successfully been used in the predictive maintenance of indus-
trial equipment. Moreover, it also describes the pros-cons of each machine
learning algorithm in the context of predictive maintenance. Noteworthy,
machine learning algorithms that have been used for data-driven predic-
tive maintenance are listed as follow.
 13

4.1. Logistic Regression

Logistic regression is a type of supervised classification algorithm. It was
invented by David Cox in 1958. It is used for developing a regression model,
the model performs prediction analysis which describes the data and its
relationship between 1 dependent variable and 1 or many nominal, ordi-
nal, interval/ratio level independent variables. There are three types of lo-
gistic regressions: 1) binomial, 2) multinomial, and 3) ordinal.

4.1.1. Binary or Binominal Logistic Regression

It is a binary response variable. The targeted variables can only have two
possible types: 0 or 1, which can represent yes or no, scenarios.

4.1.2. Multinomial Logistic Regression

Multinomial logistic regression differs from binomial logistic regression
based on the output variable values. The output in multinomial logistic re-
gression may have greater than 2 possible discrete outputs that have no
quantitative significance.

4.1.3. Ordinal Logistic Regression

The third type of ordinal deals with ordered categories. Regression coeffi-
cients are not easy to interpret in logistic regression as compare to linear
regression. For example, a result card can be categorized in this form as
excellent, very good, good, or bad. Logistic Regression assumes that
the logit of the target variable is a linear function of the parameters. It out-
puts a variable between 0 and 1, and it is commonly used for binary classi-
fication. It is a discriminative model. It handles the linear solution.

4.2. k-Nearest neighbors

k-NN that is abbreviation of k-nearest neighbors is one of the simplest ma-
chine learning techniques used for classification purposes. In k-NN, each
data sample in the training set is represented by N-dimensional space rep-
resented by its “N” number of features [53]. Similarly, instances of the test
set are also represented in a similar feature space. The goal of k-NN is to
classify test set samples into appropriate classes based on neighborhood
14

criterion calculated by using some similarity metrics. The classes of k near-

est neighbors are determined and test samples is assigned with the class
having majority votes. The matrices used to calculate the similarity among
test samples and k nearest neighbors (training samples) include Euclidean
distance, Mahalanobis distance, Manhattan distance. Furthermore, cus-
tomized similarity metrics can also be designed which can effectively de-
termine k nearest neighbors in the presence of outliers. For instance, Eu-
clidian distance “DE” between nth test sample Tk and mth train sample Trm
can be calculated as:

1/ 2
éN 2ù
DE = êå (Tn - Trm ) ú , n = 1, 2, 3, …, N ,
ë n=1 û (1)
m = 1, 2, 3, …, M

where N is the number of features and M represents the total number

of training samples.

A schematic illustration of k-NN algorithms is presented in Fig 3, which ex-

plains the whole process graphically. Due to its simple nature and easiness
to implement the k-NN algorithm has been widely used in devising condi-
tion monitoring of equipment. To name a few it has been utilized in the
condition monitoring of storage tanks [9], bearing and gearbox fault diag-
nosis [10][11], shaft misalignment detection [12], structural health moni-
toring [13].
 15

Fig 3. A k-NN classifier [14].

4.3. Artificial Neural Network

Artificial Neural Network (ANN) is adopted by the biological neural network
that makes the human brain. The term “neural” in ANN is used to represent
neuron and ‘network’ denotes a structure like a graph. ANNs are also re-
ferred to as neural nets, artificial neural systems. These networks learn in-
formation from data and perform tasks without being explicitly pro-
grammed according to task-related rules.

A shallow ANN is composed of three layers as can be seen in Fig 4. The

very first layer is the input layer that receives the original data as input into
the system which is then further processed by following layers of artificial
neurons. The following layers are the hidden layer, where artificial neurons
get a set of defined inputs and generate an output via an activation func-
tion. The output layer, which is the last layer of ANN, gives the outputs for
the program. Some applications of ANN in data-driven predictive
16

maintenance can be seen in the form of condition monitoring [15], fault

diagnosis [16], and prognosis health management of equipment [17].

Fig 4. The basic architecture of a feed-forward artificial neural network

[18].

4.4. Support Vector Machine

Support-vector machine (SVM) or support-vector network, proposed by
Vapnik [19] is a supervised learning model that is used for regression and
classification. Statistic learning theory (SLT) is the basic building block of
SVM. The goal of SVM is to find hyperplane based on support vectors that
can effectively divide instances into different classes as given in Fig 5. A
hyperplane can be defined as a boundary that differentiates samples of dif-
ferent categories. This division can be performed in SVM either by finding
hyperplane with maximum margins from support vectors or an optimal one
that perform the categorization effectively.
 17

SVM can be used in both binary and multiclass classification problems.

In practice, multiclass problems are encountered frequently, therefore,
there are multiple variants of SVM available to solve this sort of problem.
A few of the variants of SVM includes one against one (OAO) approach, one
against all (OAA) approach, decision directed acyclic graph (DDAG), and hi-
erarchical structure for multiclass SVM [14]. SVMs have been considered
as the gold standard for data-driven predictive maintenance in industrial
applications. The concept of hyperplane makes SVMs different than other
machine learning classifiers and enhances its classification performance
as compared to other shallow networks. Moreover, the flexibility to use
different types of kernel functions with SVM enhances its performance
even on nonlinear and complex data. These kernel functions can also
solve problems that are encountered while working with high dimen-
sional data with features that are not easily separable [20]. It has been
frequently used for the condition monitoring and fault diagnosis of rotary
machines and gearbox [21] [22] [23] [24], centrifugal pumps [25] [26], dis-
tillation columns [27] [28], ship engine fault [29], air-conditioning system
[30].
18

Fig 5. A hyperplane separating two classes in support vector machine

[31].

5. Limitations of Machine learning Algorithms

used for Predictive Maintenance
Although machine learning algorithms have been extensively used in the
development of a predictive maintenance mechanism there are some lim-
itations associated with it. The reason behind the rendering of such algo-
rithms for predictive maintenance is to automatically detect and diagnose
any defect in equipment under observation. The detection of fault is also
crucial to implement an effective prognostic strategy for equipment. A
few of the limitations of machine learning in the context of predictive
maintenance are listed as follow [32][33][34] [35]:
1. Generalizability
The implementation mechanism of machine learning is domain
specific. It means each sort of application needs separate training
and fine-tuning of the algorithm.
2. Domain Related Knowledge
When using machine learning algorithms in predictive mainte-
nance tasks expert knowledge about the problem domain is re-
quired. A feature engineering step is mandatory in the machine
learning based fault detection, diagnostic, prognostic process.
Feature engineering is a difficult process and requires a lot of ex-
pertise to generate hand-crafted features that can structure the
dataset. And it can detect a growth in fault.

3. Learning ability, reliability and performance

As the network architecture is simple for machine learning algo-
rithms, therefore, such networks have limited learning capabil-
ity. In general, these networks are referred to as shallow net-
works. In practice, in data-driven predictive maintenance the
data used during the process is noisy, nonlinear, and complex.
Machine learning algorithms are unable to handle irregularities,
non-stationarity, non-linearity of the data, which is often the case
 19

if data is from industrial equipment. Therefore, shallow net-

works have limited ability to provide data abstraction in the form
of features that are used to predict faults. So, the overall perfor-
mance of machine learning algorithms deteriorates when used
with real time datasets for predictive maintenance.

Cross-domain Analysis
Poor performance in cross-domain applications. If the nature of
the application gets complex, satisfactory performance is not
guaranteed. Maintenance actions are performed according to the
failure prediction results.

5. Deep Learning Methods

Deep Learning is a subclass of machine learning algorithms which utilizes
ANNs having stacked layers each containing several processing units [36].
This section provides an overview of the popular deep learning algorithms
that have been used in predictive maintenance. Moreover, it highlights the
advantage of deep learning over machine learning. Few of the deep learn-
ing algorithms worthy to mention in terms of predicative maintenance are
described as follow.

5.1. Deep Artificial Neural Network

As the name suggests, a deep artificial neural network (DANN) is an archi-
tecture in which several layers each continuing multiple neurons. The sim-
plest form of DANN is feed-forward DANN (FF-DANN) in which several fully
connected layers of neurons are stacked together as presented in Fig 6.
These linkages are not all equal: every link can have a different quality or
importance. The loads on these networks feeds the information on a sys-
tem. Regularly the neural system units are also termed as hubs.

The first layer receives original inputs that travel through the whole net-
work layer by layer in the forward direction. In this way, FF-DANN automat-
ically learns approximation of the original input data which can be later
used as features for regression as well as classification tasks

A feedforward neural system is a naturally roused characterization calcula-

tion. It consists of several straightforward neuron-like individual units,
20

sorted out in layers. Each unit in a layer relates to all the units in the pre-
ceding layer.

Fig 6. The basic architecture of feed-forward deep artificial neural net-

work (FF-DANN) [18].

DANN has been used for the fault diagnosis [37], leakage detection [38] ,
and prognosis of industrial equipment [37]. Moreover, its application can
also be seen in structural health monitoring [39].

5.2. Deep Convolutional Neural Network

Deep Convolutional Neural Networks (DCNN) are stacked ANN that used
convolutional operation to extract information from the inputs. A back-
propagation (BP) algorithm is used in DCNN to reduce the value of the cost
role by adjusting weights and biases parameters of the network.

DCNN is a unique type of Neural Network that demonstrated exceptional

performance in several computer vision and image processing competi-
tions [40]. Many of the electrifying fields of operation of CNN involve image
 21

recognition and segmentation, object detection, video processing. Mainly

thanks to multiple phases of removal, which can automatically get repre-
sentative data, the learner-centered skill of DCNN. The accessibility of a
great deal of device information has increased data analysis in DCNNs.
There have been some exciting ideas for developments on CNNs, including
the use of various triggering and loss functions, optimization of parame-
ters, regularization, and architectural creativity. Nonetheless, technologi-
cal advances achieve a major increase in the representative capability of
the deep CNN. Significant attentiveness has been rewarded to the ideas of
using space and the channel details, architectural depth and width, and
multifaceted cognitive processes.

Likewise, there is also a concept of using a surface block as a supporting

structure. This research thus focuses on the underlying taxonomy of the
recently identified profound CNN architectures and thus divides recent de-
velopments into seven separate groups of CNN architectures. The usage of
these seven categories is centered on the space, depth, multistage, dis-
tance, charts, channel boosting, and attention. In contrast, general
knowledge of the CNN components is also provided, as well as current dif-
ficulties and CNN applications.

Due to the profound success in numerous fields, DCNN architectures have

widely been used in the field of predictive maintenance to detect, diagnose
faults in equipment, and perform prognostics.

Deep CNNs are traditional neural feedforward networks that use BP algo-
rithms for the adjustment of the network's parameters to optimize the cost
function. Fortunately, in four new aspects it is quite different from regular
BP networks: local receptive fields, shared weights, pooling, and different
layers combination. A basic deep neural network is given in Fig 7. Moreo-
ver, the selection of network architecture is domain and application-spe-
cific and can be modified accordingly. In the field of predictive mainte-
nance, researchers have utilized both one dimensional as well as two-
dimensional DCNNN architectures. It has been under study for predictive
maintenance in a number of domains including but not limited to bearing
[41] [35], high-velocity oxy-fuel machine [42], and buildings [43]
22

Fig 7. Convolutional Neural Network Architecture[44].

5.3. Deep Recurrent Neural Network

Deep recurrent neural networks (DRNN) have been successfully used for
various tasks like for learning word embeddings, for language modeling, for
online handwritten character recognition [45], and speech recognition
[46]. Like a traditional DNN, it is made up of multiple hidden layers. A DRNN
has recurrent (feedback) connections among hidden layers and works se-
quentially. Every concealed unit is linked to itself as well as to rest of the
nodes in that hidden layer. Thus, DRNN means multiple recurrent hidden
layers stacked to form a sequential network hierarchy.

DNN and DCNN are not much of a use for the processing of sequential
information., e.g., if the occurrence of two inputs depends on each other
and constitute to the next input in time and so on, in such a scenario it is
better to adopt the recurrent neural network (RNN) architecture. The RNN
network can be considered as the replication of the same recurrent unit
throughout the network. The main goal behind using RNN is to explore the
sequential nature of inputs and extract information for further usage. It
predicts the next occurrence based upon the analyzed sequences in any
application where data comes in sequential order. A 3-layered DRNN archi-
tecture is shown in Fig 8 which takes input “s” at the time, t−1, t, and t+1
and predicts the respective output P at each time instance. Similarly, ct is
the hidden state at time t, ct−1 is the hidden state at time t−1, and ct+1 is
the next hidden state at time t + 1. Here, all hidden states take the input
from the previous hidden layer and contract in the calculation of final out-
put Pt+1 at time t+1. Moreover, x, y, and z represent the weights for differ-
ent layers. The application of DRNN in predictive maintenance can be
 23

found for induction motors [47], aero-engine [48], reciprocating compres-

sor [49], etc.

Fig 8. Deep Recurrent Neural Network Architecture [50].

5.4. Deep Auto-encoders

An autoencoder is a neural network used to acquire informative codes
from the input data in an unaided way. There are multiple advantages as-
sociated with the usage of autoencoders which are listed as follows.
1. It can learn informative codes from raw data which can be used as
features in regression and classification tasks.
2. 2. It can be used for data dimensionality reduction.
3. There are variants of autoencoders which can help in the exploration
of incomplete inputs (noisy data with missing values).
4. These can be helpful in fine-tuning of DNNs.

The least complex type of an autoencoder is a feedforward, a non-inter-

mittent neural system like a multilayer perceptron (MLP) – having an input
24

layer, an intermediate or code (hidden) layer, and output. In this way, au-
toencoder comprises of an encoder, and a decoder layer as shown in Fig 9.

In each layer, there are multiple processing units (neurons). The number of
processing units may be the same in all layers or may vary depending upon
the nature of implementation. In practical scenarios, where dimensionality
reduction of the input data is also intended, the number of processing units
in the code layer is kept minor than the input layer. The hidden layer ex-
tracts prominent information from the input data which can be used as
features for the categorization task. If multiple autoencoders are stacked,
it creates a stacked autoencoder. The learning process of the autoencoder
is unsupervised, whereas, if used in the categorization task, an additional
layer is appended to the stacked autoencoders which is fine-tuned sepa-
rately in a supervised manner, to determine the classes for instance under
observation. Most often this supervised layer is of SoftMax classifier. Deep
auto-encoders have been extensively used for the condition monitoring as
well as fault diagnosis of rotary machine bearings [51], leakage detection
of storage tank[52], fault detection of elevator system [53], etc.
 25

Fig 9. Deep Autoencoder Architecture [54].

5.5. Deep Belief Network

Deep Belief Network (DBN) is the graphical representation model for the
inputs provided to the networks and is generative, i.e., it can generate all
the possible value for a given scenario. It is constituted on the principles of
machine learning, ANN, probability, and statistics. As in deep architecture,
DBN consists of multiple layers with values, having relation between the
layers but not among the values. Its main objective is to categorize the
data.

DBN is made up of unsupervised neural networks such as Restricted Boltz-

mann Machines (RBMs). Multiple RBMs are stacked to form a DBNN. An
illustration of a hidden layer constituted by RBM in a DBN is given in Fig 10.
In DBN each hidden layer of an RBM is an observable layer to the next one.
The hidden layers of DBN are conditionally independent. The invisible layer
of every sub-network is the visible layer of the next. The hidden or invisible
layers are not interconnected and are conditionally independent. The pre-
dictive maintenance researchers and experts have utilized DBN for the fault
detection and diagnosis of bearing [55], gearbox [56], complex chemical
processes [57], etc.
26

Fig 10. Restricted Boltzmann Machine [36].

6. Advantages of Deep Learning in Predictive

maintenance
The application of deep neural networks in various fields has achieved a
remarkable feat. It has made the tasks possible which previously perceived
to be impossible, such as the ability to handle big data as well as to charac-
terize it in a meaningful manner. Its role does not differ in the field of pre-
dictive maintenance. In recent years, a tendency has been evolved to apply
deep learning techniques for fault detection [58], fault diagnosis [59], and
its prognosis [60]. A few of the benefits that deep learning brings to the
field of predictive maintenance are listed below:

1. The main advantage of incorporating a deep learning algorithm is the

automated learning of structures from the fresh data. The nonlinear
transformations in hierarchical order make it easy to extrapolate in-
formation from coarse data without the need for feature extraction
and selection steps.
2. As the overhead of feature engineering and selection step is not man-
datory, relatively it is easy to develop condition monitoring, fault de-
tection and diagnosis, and prognosis strategy for predictive mainte-
nance.
3. It is the age of data and data is the new gold, therefore, in the field of
data-driven predictive maintenance, data plays an important role.
Without it, nothing is possible in data-driven predictive maintenance.
Through sensors, it is likely to recorded data regarding the health
states of equipment on regular intervals for longer duration constitut-
ing datasets with huge records. Deep learning algorithms are suitable
to tackle the challenge of big data analytics as compared to machine
learning.
4. Deep learning algorithms are more suitable for transfer learning. It
makes it feasible for the development of cross-domain data-driven
predictive maintenance solutions.
 27

5. An additional benefit that comes along with transfer learning is the

evasion of the training process that saves lots of time and computa-
tion power.
6. The generalization power of deep learning based predictive mainte-
nance strategies is greater as compared to machine learning ones.
7. Multi-task learning is possible in deep learning based predictive
maintenance solutions. It helps to create multiple threads for different
tasks instead of training a separate model for each one.
8. The larger number of layers and neurons in deep learning net-
work, it allows the notion of complex problems which provides
additional performance boost.
9. The most attractive part of applying deep learning in the domain
of predictive maintenance is the fact that these networks can au-
tomatically extract the correct feature from the data while elimi-
nating the need for manual feature engineering.
10. Use deep learning to predict failures when up to date, so that it
covers any new event or behaviour.

6. Deep Learning for Predictive maintenance

The goal of data-driven predictive maintenance is to automate the
fault identification and prediction procedure with the help of artifi-
cial intelligence techniques most with focus on machine learning al-
gorithms. However, in practice application of these techniques is not
an easy task due to multiple reasons. First and foremost is the lack of
availability of an efficient algorithm. Second, the data associated with
an object under examination for the predictive maintenance is the
composition of desired information coupled with noise. The addition
of this extra information makes the exploration of the data challeng-
ing due to its non-linear, non-stationary, and complex nature. As
stated in [50], the ability of traditional machine learning algorithms
is usually limited to process complex data in raw form. For this rea-
son, to develop an effective predictive maintenance system using con-
ventional machine learning techniques a data engineering step is in-
cluded in the predictive maintenance pipeline requires which
considerable domain expertise. Through this additional step raw
data can be transferred into carefully curated features that can be
used as data descriptors in abstract form for the maintenance
28

process. For the past few years there is a huge tendency towards us-
ing deep learning methods in data-driven predictive maintenance has
been observed. One of the main reasons behind this is the ability of
deep learning algorithms to learn from the raw data by their own. It
means that if a deep learning technique has been incorporated in the
predictive maintenance strategy, there is no need for in depth
knowledge of the problem domain for feature engineering. Further-
more, a few additional advantages of deep learning algorithms as
compare to traditional machine learning techniques in terms of pre-
dictive maintenance have been listed in Table 2 [41][61][62][63].

Deep learning Algorithm Superiority for Predictive

Maintenance

DANN Degradation mapping, and fail-

ures identification, when
enough history data could be
obtained, and the complexity of
target issue is relatively high.

DCNN It can be useful to use CNN as

a diagnostic tool for data
driven predictive maintenance
when dealing with two-dimen-
sional input data. The convolu-
tional filters can extract useful
local pattern from complex raw
data in a robust manner. Fur-
ther the implication of stacked
convolutional layer enhances
the fault diagnostic perfor-
mance.

DRNN Through different variations of

RNN deep sequential struc-
tures can be formed which pro-
vide the concept of memory
unites and is useful to exploit
 29

temporal data while perform-

ing predictive maintenance.
DSAE DSAE are semi-supervised neu-
ral networks which are used to
automatically mine fault signa-
tures from the raw data. These
networks are highly effective in
discriminant information anal-
ysis about the health states of
an object. Moreover, DSAE
can be used for the dimension-
ality reduction.
DBN DBN is an energy-based archi-
tecture that can learn hidden
information from the complex
inputs. It is beneficial in the
data-driven predictive mainte-
nance when the decision is not
solely based upon historical
data. Moreover, it is also useful
when the input data dimen-
sions are large.

7. Conclusion and Future Perspectives of Deep

learning based Data-Driven Predictive mainte-
nance.

7.1. Conclusion
With the advent of sensors, it became feasible to acquire data from indus-
trial equipment. The acquired data is further analyzed for insights. If the
analysis system is up to the mark that provides right perceptions without
any postponement about the industrial or motorized equipment, it can be
of an extra benefit to the expertise of the operators or engineers for taking
a precautionary action. Due to the current advancements in ML and DL
30

algorithms, accessibility of proper sensors and ubiquitous computing auto-

mated predictive maintenance is attainable.

It can be expected that the availability of data from industrial equipment

will further improvise by the incorporation of IoT using different sensors.
With the availability of humongous data, sophisticated big data analysis
techniques are required to devise reliable data-driven predictive mainte-
nance.

7.2. Future Perspectives of Deep learning based Data-Driven

Predictive maintenance.
There is a still scope available for the improvement of deep learning based
predictive maintenance. Some of the limitations that are faced by deep
learning algorithms in terms of predictive maintenance are listed pre-
sented in the following subsections

7.2. 1. Enhanced Generalization

Although advanced deep learning techniques such as fine-tune transfer
learning [64] and multitask learning [65] have brought a sense of generali-
zation into the data-driven predictive maintenance strategies, but still
these concepts must be explored in depth. Such concepts can be delved
into for the implementation of domain-independent data-driven predictive
maintenance.

7.2.2. Explainability
There is no doubt that the data handling and exploration strength of deep
learning is far greater than that of machine learning. Its implication in the
field of predictive maintenance has reduced a lot of overhead and compli-
cations that were faced by classical machine learning techniques. To name
a few it can cope up with big data easily and can learn salient information
from the inputs automatically without the requirement of domain-specific
feature engineering step. Nevertheless, instead of enhanced capability,
deep leering algorithms are like a black box. At present, there is no proper
explanation that how deep learning algorithms accurately approximate the
 31

complex, nonlinear, and nonstationary data in an abstract way. Further,

how the approximated codes which are also termed as features yield a bet-
ter predictive maintenance performance than that of its predecessors.
There is a need for explainable deep learning based predictive mainte-
nance strategies.

7.2.3. Multimodal and Multi-Sensors Data Fusion

The data fusion from multiple sensors and modalities is an interesting and
possible extension of the deep learning-based data-driven predictive
maintenance. Data fusion can provide in-depth details about the bearing
defects that can help in the enhancement of bearing fault diagnosis mod-
els. Data fusion from multiple sensors is also a practical consideration as in
practice for better performance multiple sensors are deployed on the con-
cerned component to collect the data.

Glossary

Condition Monitoring: It is the constant monitoring of the parameters

from a machine which are associate with its health indication. Any signifi-
cant fluctuation in the monitored values of the parameters indicates the de-
viation of the machine from its normal state.
Data-Driven: It is an adjective which means that a given activity is based
upon data.
Multimodal Data: Data taken through several modes of recordings
within one application.
Multi-Sensor: Data recorded through multiple sensors. Types of sensors
may be the same or may differ.

Sensor: It is a device that takes a physical reading from the environment

and converts it into electrical data or signals.
Support Vectors: The closest data point to the hyperplane are called
support vectors. These entities define the location and orientation of the
hyperplane.
32

References
[1] ELECTRONICS HUB, “What is a sensor?”
https://www.electronicshub.org/different-types-sensors (accessed
Aug. 15, 2020).
[2] J. Butler and C. Smalley, “An introduction to predictive
maintenance,” Pharm. Eng., 2017, doi: 10.1016/b978-0-7506-
7531-4.x5000-3.
[3] W. Zhang, D. Yang, and H. Wang, “Data-Driven Methods for
Predictive Maintenance of Industrial Equipment: A Survey,” IEEE
Syst. J., 2019, doi: 10.1109/JSYST.2019.2905565.
[4] R. K. Mobley, An Introduction to Predictive Maintenance (Second
Edition). 2002.
[5] R. Hemmerdinger, “Predictive maintenance strategy for building
operations: A better approach,” Schneider Electr., 2014.
[6] J. C. P. Cheng, W. Chen, K. Chen, and Q. Wang, “Data-driven
predictive maintenance planning framework for MEP components
based on BIM and IoT using machine learning algorithms,” Autom.
Constr., vol. 112, p. 103087, 2020.
[7] E. by AspenCore, “Choosing the most suitable predictive
maintenance sensor,” 2020.
https://www.embedded.com/choosing-the-most-suitable-
predictive-maintenance-sensor/ (accessed Aug. 15, 2020).
[8] V. Systeme, “Acoustic Emission Sensors,” 2019.
https://www.vallen.de/wp-content/uploads/2019/03/sov.pdf
(accessed Aug. 15, 2020).
[9] M. J. Hasan and J.-M. Kim, “Fault Detection of a Spherical Tank
Using a Genetic Algorithm-Based Hybrid Feature Pool and k-
Nearest Neighbor Algorithm,” Energies, vol. 12, no. 6, p. 991, 2019.
[10] Z. Li, X. Yan, C. Yuan, and L. Li, “Gear multi-faults diagnosis of a
rotating machinery based on independent component analysis and
fuzzy k-nearest neighbor,” 2010, doi:
10.4028/www.scientific.net/AMR.108-111.1033.
 33

[11] A. Sharma, R. Jigyasu, L. Mathew, and S. Chatterji, “Bearing Fault

Diagnosis Using Weighted K-Nearest Neighbor,” 2018, doi:
10.1109/ICOEI.2018.8553800.
[12] M. Gohari and A. M. Eydi, “Modelling of shaft unbalance:
Modelling a multi discs rotor using K-Nearest Neighbor and
Decision Tree Algorithms,” Meas. J. Int. Meas. Confed., 2020, doi:
10.1016/j.measurement.2019.107253.
[13] J. Vitola, F. Pozo, D. A. Tibaduiza, and M. Anaya, “A sensor data
fusion system based on k-nearest neighbor pattern classification
for structural health monitoring applications,” Sensors
(Switzerland), 2017, doi: 10.3390/s17020417.
[14] Y. Wei, Y. Li, M. Xu, and W. Huang, “A review of early fault
diagnosis approaches and their applications in rotating
machinery,” Entropy. 2019, doi: 10.3390/e21040409.
[15] D. V. S. S. S. Sarma and G. N. S. Kalyani, “ANN approach for
condition monitoring of power transformers using DGA,” 2004,
doi: 10.1109/tencon.2004.1414803.
[16] Z. Y. Zhang and K. S. Wang, “Wind turbine fault detection based on
SCADA data analysis using ANN,” Adv. Manuf., 2014, doi:
10.1007/s40436-014-0061-6.
[17] Z. Zhang, Y. Wang, and K. Wang, “Fault diagnosis and prognosis
using wavelet packet decomposition, Fourier transform and
artificial neural network,” J. Intell. Manuf., 2013, doi:
10.1007/s10845-012-0657-2.
[18] B. Yegnanarayana, Artificial neural networks. PHI Learning Pvt.
Ltd., 2009.
[19] C. Cortes and V. Vapnik, “Support-vector networks. 1995,” Mach.
Learn., vol. 20, no. 3, p. 273, 2017.
[20] P. F. Orrù, A. Zoccheddu, L. Sassu, C. Mattia, R. Cozza, and S.
Arena, “Machine learning approach using MLP and SVM
algorithms for the fault prediction of a centrifugal pump in the oil
and gas industry,” Sustain., 2020, doi: 10.3390/su12114776.
[21] X. Li, Y. Yang, H. Pan, J. Cheng, and J. Cheng, “A novel deep
34

stacking least squares support vector machine for rolling bearing

fault diagnosis,” Comput. Ind., 2019, doi:
10.1016/j.compind.2019.05.005.
[22] Y. Li, W. Zhang, Q. Xiong, D. Luo, G. Mei, and T. Zhang, “A rolling
bearing fault diagnosis strategy based on improved multiscale
permutation entropy and least squares SVM,” J. Mech. Sci.
Technol., 2017, doi: 10.1007/s12206-017-0514-5.
[23] F. Chen, B. Tang, and R. Chen, “A novel fault diagnosis model for
gearbox based on wavelet support vector machine with immune
genetic algorithm,” Meas. J. Int. Meas. Confed., 2013, doi:
10.1016/j.measurement.2012.06.009.
[24] M. Sohaib and J.-M. Kim, “Hierarchical radial basis function based
multiclass support vector machines and a hybrid feature pool for
bearings fault diagnosis,” 2018, doi: 10.1109/ICEE.2018.8566908.
[25] A. K. Panda, J. S. Rapur, and R. Tiwari, “Prediction of flow
blockages and impending cavitation in centrifugal pumps using
Support Vector Machine (SVM) algorithms based on vibration
measurements,” Meas. J. Int. Meas. Confed., 2018, doi:
10.1016/j.measurement.2018.07.092.
[26] D. J. Bordoloi and R. Tiwari, “Identification of suction flow
blockages and casing cavitations in centrifugal pumps by optimal
support vector machine techniques,” J. Brazilian Soc. Mech. Sci.
Eng., 2017, doi: 10.1007/s40430-017-0714-z.
[27] S. A. Taqvi, L. D. Tufa, H. Zabiri, A. S. Maulud, and F. Uddin,
“Multiple Fault Diagnosis in Distillation Column Using Multikernel
Support Vector Machine,” Ind. Eng. Chem. Res., 2018, doi:
10.1021/acs.iecr.8b03360.
[28] L. Liu and A. L. Liu, “Fault diagnosis of distillation column based on
improved genetic algorithm optimization-based support vector
machine,” Huadong Ligong Daxue Xuebao/Journal East China Univ.
Sci. Technol., 2011.
[29] C. Cai, H. Zong, and B. Zhang, “Ship diesel engine fault diagnosis
based on the SVM and association rule mining,” 2016, doi:
10.1109/CSCWD.2016.7566022.
 35

[30] K. Sun, G. Li, H. Chen, J. Liu, J. Li, and W. Hu, “A novel efficient
SVM-based fault diagnosis method for multi-split air conditioning
system’s refrigerant charge fault amount,” Appl. Therm. Eng.,
2016, doi: 10.1016/j.applthermaleng.2016.07.109.
[31] J. A. K. Suykens and J. Vandewalle, “Least squares support vector
machine classifiers,” Neural Process. Lett., vol. 9, no. 3, pp. 293–
300, 1999.
[32] Z. M. Çınar, A. Abdussalam Nuhu, Q. Zeeshan, O. Korhan, M.
Asmael, and B. Safaei, “Machine Learning in Predictive
Maintenance towards Sustainable Smart Manufacturing in
Industry 4.0,” Sustainability, vol. 12, no. 19, p. 8211, 2020.
[33] F. Lv, C. Wen, Z. Bao, and M. Liu, “Fault diagnosis based on deep
learning,” in 2016 American Control Conference (ACC), 2016, pp.
6851–6856, doi: 10.1109/ACC.2016.7526751.
[34] N. G. Lo, J.-M. Flaus, and O. Adrot, “Review of Machine Learning
Approaches In Fault Diagnosis applied to IoT Systems,” in 2019
International Conference on Control, Automation and Diagnosis
(ICCAD), 2019, pp. 1–6.
[35] M. J. Hasan, M. Sohaib, and J. M. Kim, “1D CNN-based transfer
learning model for bearing fault diagnosis under variable working
conditions,” 2019, doi: 10.1007/978-3-030-03302-6_2.
[36] Y. Bengio, I. Goodfellow, and A. Courville, Deep learning, vol. 1.
Citeseer, 2017.
[37] S. E. Pandarakone, M. Masuko, Y. Mizuno, and H. Nakamura,
“Deep neural network based bearing fault diagnosis of induction
motor using fast Fourier transform analysis,” in 2018 IEEE Energy
Conversion Congress and Exposition (ECCE), 2018, pp. 3214–3221.
[38] M. Sohaib and J.-M. Kim, “Data Driven Leakage Detection and
Classification of a Boiler Tube,” Appl. Sci., Jun. 2019, doi:
10.3390/app9122450.
[39] I. D. Khurjekar and J. B. Harley, “Uncertainty Aware Deep Neural
Network for Multistatic Localization with Application to Ultrasonic
Structural Health Monitoring,” arXiv Prepr. arXiv2007.06814, 2020.
36

[40] A. Khan, A. Sohail, U. Zahoora, and A. S. Qureshi, “A survey of the

recent architectures of deep convolutional neural networks,” Artif.
Intell. Rev., 2020, doi: 10.1007/s10462-020-09825-6.
[41] M. Sohaib and J.-M. Kim, “Fault diagnosis of rotary machine
bearings under inconsistent working conditions,” IEEE Trans.
Instrum. Meas., vol. 69, no. 6, pp. 3334–3347, 2019.
[42] K. Ibrahim and M. Masrom, “Predictive Maintenance of High-
Velocity Oxy-Fuel Machine Using Convolution Neural Network,”
Available SSRN 3660305, 2020.
[43] Ç. F. Özgenel and A. G. Sorguç, “Performance comparison of
pretrained convolutional neural networks on crack detection in
buildings,” in ISARC. Proceedings of the International Symposium
on Automation and Robotics in Construction, 2018, vol. 35, pp. 1–
8.
[44] Missinglink.ai, “The Complete Guide to Artificial Neural Networks:
Concepts and Models.” .
[45] H. Ren, W. Wang, and C. Liu, “Recognizing online handwritten
Chinese characters using RNNs with new computing
architectures,” Pattern Recognit., 2019, doi:
10.1016/j.patcog.2019.04.015.
[46] M. W. Y. Lam, X. Chen, S. Hu, J. Yu, X. Liu, and H. Meng, “Gaussian
Process Lstm Recurrent Neural Network Language Models for
Speech Recognition,” 2019, doi: 10.1109/ICASSP.2019.8683660.
[47] D. Xiao, Y. Huang, C. Qin, H. Shi, and Y. Li, “Fault diagnosis of
induction motors using recurrence quantification analysis and
LSTM with weighted BN,” Shock Vib., vol. 2019, 2019.
[48] M. Yuan, Y. Wu, and L. Lin, “Fault diagnosis and remaining useful
life estimation of aero engine using LSTM neural network,” in 2016
IEEE International Conference on Aircraft Utility Systems (AUS),
2016, pp. 135–140.
[49] D. Cabrera et al., “Bayesian approach and time series
dimensionality reduction to LSTM-based model-building for fault
diagnosis of a reciprocating compressor,” Neurocomputing, 2019.
 37

[50] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol.

521, no. 7553, pp. 436–444, 2015.
[51] M. Sohaib, C.-H. Kim, and J.-M. Kim, “A hybrid feature model and
deep-learning-based bearing fault diagnosis,” Sensors
(Switzerland), 2017, doi: 10.3390/s17122876.
[52] M. Sohaib, M. Islam, J. Kim, D.-C. Jeon, and J.-M. Kim, “Leakage
detection of a spherical water storage tank in a chemical industry
using acoustic emissions,” Appl. Sci., 2019, doi:
10.3390/app9010196.
[53] K. M. Mishra, T. R. Krogerus, and K. J. Huhtala, “Fault detection of
elevator systems using deep autoencoder feature extraction,” in
2019 13th International Conference on Research Challenges in
Information Science (RCIS), 2019, pp. 1–6.
[54] M. Sohaib and J.-M. Kim, “Reliable Fault Diagnosis of Rotary
Machine Bearings Using a Stacked Sparse Autoencoder-Based
Deep Neural Network,” Shock Vib., 2018, doi:
10.1155/2018/2919637.
[55] H. Shao, H. Jiang, H. Zhang, and T. Liang, “Electric locomotive
bearing fault diagnosis using a novel convolutional deep belief
network,” IEEE Trans. Ind. Electron., vol. 65, no. 3, pp. 2727–2736,
2017.
[56] Z. Chen, C. Li, and R.-V. Sánchez, “Multi-layer neural network with
deep belief network for gearbox fault diagnosis,” J.
Vibroengineering, vol. 17, no. 5, pp. 2379–2392, 2015.
[57] Z. Zhang and J. Zhao, “A deep belief network based fault diagnosis
model for complex chemical processes,” Comput. Chem. Eng., vol.
107, pp. 395–407, 2017.
[58] K. P. Lee, B. H. Wu, and S. L. Peng, “Deep-learning-based fault
detection and diagnosis of air-handling units,” Build. Environ.,
2019, doi: 10.1016/j.buildenv.2019.04.029.
[59] Y. Yu, D. Woradechjumroen, and D. Yu, “A review of fault
detection and diagnosis methodologies on air-handling units,”
Energy and Buildings. 2014, doi: 10.1016/j.enbuild.2014.06.042.
 38

[60] Y. Su, F. Tao, J. Jin, T. Wang, Q. Wang, and L. Wang, “Failure

Prognosis of Complex Equipment With Multistream Deep
Recurrent Neural Network,” J. Comput. Inf. Sci. Eng., 2020, doi:
10.1115/1.4045445.
[61] Z. Chen and W. Li, “Multisensor feature fusion for bearing fault
diagnosis using sparse autoencoder and deep belief network,” IEEE
Trans. Instrum. Meas., 2017, doi: 10.1109/TIM.2017.2669947.
[62] J. S. Rahhal and D. Abualnadi, “IOT Based Predictive Maintenance
Using LSTM RNN Estimator,” in 2020 International Conference on
Electrical, Communication, and Computer Engineering (ICECCE),
2020, pp. 1–5.
[63] S. Butte, A. R. Prashanth, and S. Patil, “Machine learning based
predictive maintenance strategy: a super learning approach with
deep neural networks,” in 2018 IEEE Workshop on
Microelectronics and Electron Devices (WMED), 2018, pp. 1–5.
[64] M. J. Hasan, M. M. M. Islam, and J.-M. Kim, “Acoustic spectral
imaging and transfer learning for reliable bearing fault diagnosis
under variable speed conditions,” Measurement, vol. 138, pp.
620–631, 2019.
[65] X. Cao, B. Chen, and N. Zeng, “A deep domain adaption model with
multi-task networks for planetary gearbox fault diagnosis,”
Neurocomputing, vol. 409, pp. 173–190, 2020.

View publication stats