Version of Record: https://www.sciencedirect.

com/science/article/pii/S0263224121011805
Manuscript_3a79a46d5aadfacdc074d3162253ee8e

Recent Advances and Trends of Predictive Maintenance from

Data-driven Machine Prognostics Perspective

Yuxin Wen1, Md. Fashiar Rahman2, Honglun Xu3 and Tzu-Liang (Bill) Tseng2∗

Abstract
In the Engineering discipline, prognostics play an essential role in improving system safety,
reliability and enabling predictive maintenance decision-making. Due to the adoption of emerging
sensing techniques and big data analytics tools, data-driven prognostic approaches are gaining
popularity. This paper aims to deliver an extensive review of recent advances and trends of data-
driven machine prognostics, with a focus on their applications in practice. The primary purpose of
this review is to categorize existing literature and report the latest research progress and directions
to support researchers and practitioners in acquiring a clear comprehension of the subject area.
This paper first summarizes fundamental methodologies on data-driven approaches for predictive
maintenance. Then, the article further conducts a comprehensive investigation on the different
fields of applications of machine prognostics. Finally, a discussion on the challenges,
opportunities, and future trends of predictive maintenance is presented to conclude this paper.
Keywords: Machine prognostics, predictive maintenance, condition-based maintenance, machine
learning, prognostics and health management, remaining useful life

1. Introduction
In the past decades, the science of prognostics and health management (PHM) of complex
engineering systems attracts the research community and industrial practitioners [1]. The primary
motivation of PHM is to improve system safety, increase machine reliability and availability, and
reduce maintenance costs. Rapid development in digital technologies, such as 3D printing, robots,
artificial intelligence (AI), digital automation, cloud computing, the Internet of Things (IoT), and

∗
Corresponding author, Email: btseng@utep.edu
1
Dale E. and Sarah Ann Fowler School of Engineering, Chapman University, CA 92866, USA.
2
Department of Industrial, Manufacturing and Systems Engineering, The University of Texas at El Paso, TX 79968,
USA
3
Computational Science Program, The University of Texas at El Paso, TX 79968, USA

© 2021 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
more, has accelerated the Fourth Industrial Revolution, which is often known as Industry 4.0 [2].
In the context of Industry 4.0, to satisfy the escalating demand for functionality and quality,
systems have become much more interconnected and complex than any other time before.
Therefore, diligent monitoring of the system operations and then proposing appropriate
maintenance strategies have become indispensable since an unexpected failure can result in
catastrophic consequences. A proper maintenance strategy is critical to minimize unplanned
downtime and ensure that facilities operate at the highest efficiency.
Corrective and preventive maintenances are the two widely used maintenance strategies since
the early 90s [3]. Corrective maintenance, also called failure-driven maintenance, is carried out
only after the occurrence of a malfunction or breakdown of equipment. However, it frequently
results in unpredictable performance in the industry, i.e., high production cost, extensive repair
time, not to mention the cost and penalties associated with machine breakdown. Preventive
maintenance is carried out regularly. Most maintenance decisions are made by experts based on
their experience with equipment fabricants, historic breakdowns or failure data. However, it is
difficult to make a proper maintenance schedule in advance. Due to the increasing requirement of
reliability, availability, maintainability, and safety of systems, preventive maintenance is
becoming less effective and obsolete. Recently, predictive maintenance, also known as condition-
based maintenance [4], which uses predictive tools to determine when maintenance actions are
necessary, has become prevalent in the industry due to the capability of reducing maintenance cost,
unexpected downtime, and while extending the life span of equipment [5]. Predictive maintenance
employs non-intrusive testing techniques, such as thermodynamics, acoustics, vibration analysis,
infrared analysis, etc., to monitor and evaluate equipment performance trends. The core procedures
for implementing predictive maintenance include data collection, fault detection and diagnostics,
and prognostics, which are later used to guide maintenance decisions such as maintenance
scheduling or resource optimization. A comprehensive review of each component can be found in
[6-8]. Of these procedures, diagnostics and prognostics are two critical aspects of predictive
maintenance. Diagnostics refer to identifying the presence of operational faults and determining
the root cause and effect to the functional equipment. In contrast, prognostics deal with predicting
the future state or remaining useful life (RUL) based on the current and historical conditions.
Accurately predicting “life-span” is the key to the success of predictive maintenance. It involves
analytical computations of historical or real-time data streamed from applications, sensors,

2
devices, etc. In general, prognostics measures the extent of deviation and degradation of any
machine or system from the normal operating behavior to predicts the RUL and future
performance. However, the task of prognostics is not trivial as predicting future performance
depends on the analysis of failure modes, early signals of wear and aging, and the nature of faults.
It also requires sound knowledge of the failure mechanisms which have a certain amount of
physical randomness. Moreover, prognostics identifies the potential system parameters that are
likely to cause the degradations, leading to eventual failures, which involves considerable
uncertainty and complicates the prediction. Therefore, prognostics is much more challenging than
diagnostics and requires effective and efficient predictive models to monitor the machine health
conditions.
In general, prognostic models can be classified into two groups, physical-based and data-driven
models [9]. The physical-based models capture the failure mechanisms or physical phenomena to
build a mathematical representation of the degradation process. It always requires a thorough
understanding of the sophisticated degradation mechanisms, making it infeasible or ineffective in
practical applications due to the system complexity or unclear degrading mechanism [10, 11]. On
the other hand, data-driven prognostics usually deploy data mining techniques to identify the
pattern and anomalies within the raw signals/data to detect any changes in system states. Due to
the promising applications and data availability, data-driven models are becoming attractive in
recent years. Data-driven models can be further classified into three subcategories, statistical-based
models, conventional machine learning based models, and deep learning based models. In the first
subcategory, the general path and stochastic process models are usually designed to track the
trajectory of the degradation in a probabilistic manner. Conventional machine learning approaches,
including random forest (RF), artificial neural network (ANN), support vector machine (SVM),
etc., are commonly designed to extract features for machine RUL prediction. As data increases in
dimensionality and volume, deep learning with automatic feature learning demonstrates
outstanding performance in reliability estimation using degradation data.
Several efforts have been performed to review the topic of degradation modeling and machine
prognostics in the recent decade. Si et al. provided a systematic review on data-driven models for
RUL prediction. They classified the data-driven models into two broad types of models according
to the criterion that if the models rely on directly observed state information or not [12]. Ye and
Xie [13] classified existing degradation models into general path models, stochastic process

3
models, and others with a focus on stochastic models. Zhang et al. [14] provided a review on
degradation model-based RUL estimation approaches with an emphasis on the heterogeneity in
the systems. All of the above works only focused on statistical-based models. It has been observed
a proliferation of data-driven algorithms to help with prognostics in the latest few years. For
example, Wang et al. [15] provided an in-depth review on health indicator construction for
vibration-based bearing and gear. Khan and Yairi [16] presented a systematic review of artificial
intelligence based system health management and recent trends of deep learning in the reliability
field. Recently, [17] [18] surveyed contemporary advancements of deep learning and its
applications to machine health monitoring. All of them only survey AI techniques for system
health management. In another recent work, Kordestani et al. [19] and Guo et al. [20] reviewed
and summarized the emerging prognostic modeling methods, which can be classified into data-
driven, physics-based/model-based, and hybrid approaches. Baur et al. [21] presented a review on
diagnostics and prognostics approaches from knowledge-based, model-based, statistical-based,
and data-driven concepts. This work only focuses on applying machine tools considering the feed
axis, spindle speed, and hydraulic system. Table I highlights the contribution of some review
works in the last decade.
Unlike the above-mentioned review works, this paper aims to provide an extensive and broad
overview of the most recent advances and trends of data-driven machine prognostics for predictive
maintenance, focusing on their applications in different industrial fields. This article provides a
detailed discussion and recent advances in each of the categories of data-driven approaches for
recent five years. The primary motivation of this survey is to categorize the existing literature and
summarize the latest research progress and directions to assist researchers and practitioners in
acquiring a clear comprehension of the subject area. The paper also discusses applied research
issues when applying current technology and suggests some potentially promising directions for
predictive maintenance.
The remainder of this paper is organized as follows. Section 2 provides a discussion on typical
data-driven predictive models and analysis methods. The applications for each category of data-
driven approaches are summarized in Section 3. Section 4 presents the current challenges and
opportunities of machine prognostics for predictive maintenance. Section 5 concludes this work.

4
 Table I Recent contribution of the review works in the field of PHM
Reference Year Description
[12] 2013 The review only focused on statistical data-driven approaches
[13] 2014 Reviewed based on three categories: statistical process models (SPMs),
general path models (GPMs), and other models beyond SPMs and GPMs
[22] 2014 Summarized the PHM systems for rotating machinery and provides a
systematic design methodology
[14] 2015 Focused on the degradation modeling and RUL estimation for
heterogeneity in the systems
[15] 2017 Reviewed on the health indicators construction for vibration-based bearing
and gears using mechanical signals
[16] 2018 Provided an overview of architectures and theories of artificial
intelligence-based prognostics approaches with plausible advantages and
limitations
[17, 18] 2019 Surveyed the deep learning based prognostics approaches and their
applications
[19, 20] 2019 Reviewed the recent advancement in the field of prognostics and
summarized them into three categories, i.e., the data-driven, physics-
based, and hybrid prognostics.
[21] 2020 Provides a review on diagnostics and prognostics approaches focusing on
the application of machine tools considering the feed axis, spindle speed,
and hydraulic system.

2. Data-driven Prognostic Algorithms

Prognostics algorithms focus on predicting when a system or a component stops to perform its
intended functions. In other words, the prognostic algorithms predict the future performance or the
RUL of a system or component by analyzing the extent of deviation and degradation from its
expected normal operating conditions. In general, the health state of an item degrades linearly with
its usage or operating cycle. However, the task of prognostics is not trivial due to the variation of
operation conditions, environment, and complex nature of different parameters. Prognosis requires
intensive degradation data of an item, such as lifetime data or run-to-failure histories. To make
accurate prognostics, choosing a proper modeling technique is essential. There are mainly three
modeling strategies for predictive maintenance based on degradation data: (1) regression models,
(2) classification models, (3) survival models. A regression model seeks to model the trajectory of
a degradation path and then predict when the system will fail. A classification model tries to predict
if the failure occurs within a given time window. The basic idea of survival models is trying to
answer how the risk of failure changes in time. To implement these strategies, data-driven models
can be classified into three categories: statistical-based models, conventional machine learning

5
based models, and deep learning based models. In this section, we report a systematic overview of
these three categories. The structure of this section is summarized in Figure 1.

Figure 1 Categories of data-driven prognostic models

2.1 Statistical based Models
Typically, a statistical based model for the RUL estimation is constructed via fitting a
probabilistic model to data without relying on any physics or engineering principle. Two broad
categories of statistical based models are general path models (GPMs) and stochastic process
models (SPMs). In the following subsections, a brief review of each type of these models is
provided.
2.1.1 General Path Model
The basic idea of a GPM, which was first introduced by Lu and Meeker in 1993 [23], is to find
an appropriate parametric regression model to capture the degradation trend over time. The general
path model allows the direct use of degradation data and captures the unit-wise fluctuation in

η , ,
degradation data. For any given time , the degradation path of unit is defined as
(1)

~ 0,
where is a vector of fixed (population) effects for all units, is a vector of random (individual)
effects for the unit, and is the normally distributed measurement errors. This

appropriate failure models and mapped to a function of η ∙ . Secondly, the historical data should
model relies on three basic assumptions. Firstly, the degradation data should be captured using any

be collected under similar situations considering a reasonable variation of each individual

6
component. Finally, there exists some defined critical level of degradation, termed as a soft failure,
which indicates component failure. Due to its simplicity and ease of implementation, GPM has

Equation (1) [12, 24, 25]. First, the functional form for the η ∙ could be linear, quadratic,
been well-studied, and various extensions have been developed based on the basic form in

exponential, etc [26]. If the degradation signals are complex and show nonlinear shapes, two or

Second, a variety of distributions for the parameters in the η ∙ can be considered, such as Weibull,
multiple phases, more generally, nonparametric regression forms can be assumed [27, 28].

normal, lognormal, etc. Third, error items that capture the produce and environment noises can be
assumed independently and identically distributed or correlated among different time points. To
predict the RUL for working units at an individual level, the parameters in the model need to be
estimated at the offline stage and updated at the online stage when new observations are available.
For the offline parameter estimation, the empirical two-stage method [29], maximum likelihood
estimation (MLE) and expectation–maximization (EM) algorithm provide reliable estimates. For
the online parameters updating, Bayesian framework is the most natural way, where posterior
distributions of the model parameters are generated based on newly collected data.
2.1.2 Stochastic Process Model
There are four SPMs in the literature which are commonly used for RUL prediction, namely,
Wiener process, Gamma process, Gaussian process and inverse Gaussian process models. A brief
description of these four models is illustrated in the following paragraphs.

In the stochastic process based model, for any time , and ∆ > 0 , the increments ∆
A. Wiener process

∆ −
Wiener process, ∆
of degradation signal in disjoint time intervals are independent. For a
is normally distributed. If we use a wiener process to describe a
degradation trend, The basic form can be written as
(2)
Where is a drift parameter reflecting the degradation rate, is a diffusion coefficient,

~ − , −
represents a standard Brownian motion. So then is a normal distribution with
based on the property of the Wiener
process. As a degradation model, the Wiener process has some unique advantages. A dominant
advantage is that the distribution of the failure time can be formulated analytically by the first

7
passage time (FPT), in which its probability density function (PDF) follows an inverse Gaussian
distribution, namely
,
& & −%
!"# ; %, & ' + -./ 0− 1, >0
2) * 2% (3)

Where & is the mean and % is the shape parameter. Due to its mathematical properties and physical
interpretations, the Wiener process can be easily extended to satisfy different demands. One
alternative is to add an error term into the basic process to capture measurement errors in
degradation signals [30]; the second way is to incorporate random-effects model in dealing with
unobserved heterogeneities, specifically, assume that or or both follow some certain
parametric distributions, see examples [31-33] among others. The third approach is to incorporate
nonlinear structure into this model to make the model more general. In particular, the more

; 2 ;3
generalized model is defined as

; and 2 ; 3 are non-decreasing functions with parameter vectors of and 3 [34].

(4)
Where
Wiener processes have attracted significant attention in modeling several degradation trends
encountered in real systems, such as bridge beams [31], fatigue crack dynamics [35], light-emitting
diodes [36], thrust ball bearings [37], and micro electro mechanical systems (MEMS) [38]. Zhang
[39] provided a comprehensive review of Wiener process methods with application to RUL
prediction.
B. Gamma Process
One of the distinct features of the Wiener process is that it is a non-monotone stochastic process.

trends. Gamma process is an alternative in this regard. If the increment ∆

However, it might not be suitable in many degradation applications that show apparent monotone
follows Gamma
distribution, the process is called the Gamma process. The Gamma process is proved to be
an efficient tool in the stochastic modeling of monotonic and gradual degradation in a sequence of
small increments, such as fatigue, wear, consumption, creep, corrosion, erosion, swell, crack
growth, and so forth [40]. However, Gamma process models have the following shortcomings.
First, gamma process models are constrained by the assumption of Markov property. Second,
gamma process models are only effective in describing the monotonic degradation processes [8].
A survey of the application of gamma processes in degradation modeling can be found in [41].
C. Gaussian Process

8
 The Gaussian process is another emerging approach in the field of prognostics. A Gaussian is

! . 45 6 . , 7 ., .′
defined mathematically as

where 6 . and 7 ., .′ are the mean and covariance functions respectively, denoted by
(5)

6 . 9:! . ;

7 ., . < 9 :! . − 6 . ;:! . < − 6 . < ;

=
(6)
Gaussian process regression is a way to undertake non-parametric regression with Gaussian
processes. The idea is that Gaussian process regression uses conditioning on Gaussian vectors to
find a model that actually passes through the data points. Unlike classical regression models,
Gaussian process regression does not force an analytical formula for the predictor, but a covariance
structure for the outcomes. To accurately reflect the correlations presented in the data, the
covariance functions need to be specified, and the hyperparameter values of the covariance
function need to be optimized. Due to the probabilistic nature of the Gaussian process models, the
classic model optimization approach where model parameters are optimized through the
minimization of a cost function such as mean square error is not readily applicable. A probabilistic
approach to the optimization of the model, such as the maximum likelihood method, is more
appropriate. Some examples of Gaussian process regression applied to RUL prognostics can be
found in [42-45].
D. Inverse Gaussian process

a monotone degradation path. If the increment ∆

The inverse Gaussian (IG) process is another natural choice for degradation data which provides
follows IG distribution, the process is
called inverse Gaussian process. The pdf of an IG distribution is defined as

& ?* & .−%

! .; %, & > . exp − ,. > 0
(7)

2) 2% .

Let C inf{ H I} denotes the failure time. The Failure time distribution is obtained by
N N
5 C< 5 H >I Φ 0M Λ −I 1−e PQ R
Φ 0M− Λ I 1
I I
(8)

Where S is the standard normal cumulative distribution function (CDF). Ye et al. [46] first
justified its physical meaning by exploring the inherent relations between the IG process and the
compound Poisson process.

9
To summarize, SPMs are more favorable than GPMs to account for the randomness in degradation
processes caused by both inherent and environmental factors when a significant fluctuation exists
in the data. However, compared to GPMs, SPMs are often complex and require a more in-depth
statistical and computational ability for the model parameter estimation.

Define a set of states T {U, , U , … , UW }, the Markov process is a process that starts in one of
2.1.3 Markovian-based Model

these states and moves successively from one state to another. Although the Markov process still
belongs to stochastic processes, this model is distinguished from the above stochastic models.
Markov process assumes a finite state of the degradation and the task is to find the transition
probability among those states. The main property of the Markov process is being memoryless,
which states that the future degradation state only relies on the current degradation state. RUL
estimation using Markovian-based models can be captured by computing the amount of time that
the process will take to transit from the current state to the absorbing state for the first time [12].
In real-world applications, however, the transition probabilities may also be related to other
variables, e.g., the level of degradation, the time when the product reached the current state, etc.
Semi-Markovian models extend the application of Markovian-based models by incorporating the
effects of these factors. In practice, the actual degradation level is not accessible due to the
complexity of degradation process or the random nature of the equipment. Hidden Markov Models
(HMM) and Hidden Semi-Markov models (HSMM) [47] can be used to solve this issue. In HMM,
the state of the hidden process can be inferred by the observation sequences, each state is described
by probability density distribution, and each observation vector is generated by the state of the
corresponding probability density distribution. HSMM, as a generalization of HMM, can reflect
gradual changes because of the semi-markovian assumption. Markovian-based models, which are
known for exact and approximate learning and inference, have a strong statistical foundation and
have been well studied. Due to their Markovian nature, they do not take into account the sequence
of states leading into any given state.
2.1.4 Filtering-based Model

degradation models. In the Kalman filtering model, the unobserved degradation . and the
The Kalman filtering model and Particle filters (PFs) are the most popular filtering-based

has the relationship that . X. ?, Y and . N ,

where Y and N are Gaussian noises, X and
observed degradation signal
are the parameters of the state-space model. Unlike

10
Markovian-based models that only depend on the last degradation signal, the Kalman filtering
model takes advantage of all historical data. However, the Kalman filtering model is constrained
by linear assumption and Gaussian noise assumption. To overcome the drawback, PFs are

a PF can be expressed by the state transition function ! and the measurement function ℎ:
particularly useful for linear/nonlinear Gaussian/non-Gaussian state-space models. The process of

.[ ! .[?, , [ , \[

][ ℎ .[ , ^[
where 7 is the time step, .[ is unobserved degradation, is a vector of model parameters, ][ is
(9)
[

observed degradation signal, \[ and ^[ are process and measurement noise, respectively. The
posterior / .[ |],:a can be updated recursively using Bayesian inference based on up-to-date
observations. Once the degradation model is updated, the future degradation magnitudes and RUL
can be predicted based on the updated model. Comparing with the Kalman filter, PF is more elastic
as it does not assume linearity and Gaussian nature of noise in data. Both filters start with a state-
space representation of the stochastic processes of interest. They are robust and scale well in many
applications but at the price of high computational cost.
2.1.5 Covariate based Model
The risk factors that cause the degradation process are called covariates. One of the most
popular covariate-based models is Cox Proportional Hazard (PH) model, which was proposed by
Cox in 1972 [48]. The Cox PH model allows to describe the survival time/RUL as a function of

ℎ ;b ℎc -./ db
multiple prognostic factors. The basic format of the Cox PH model is defined as,

where ℎc
(10)

so the model is often called as a semi-parametric approach. b is a vector of the corresponding

is the baseline hazard rate function, which can be either nonparametric or parametric,

covariates/ prognostic factors. The covariate b is associated with the system. d is the unknown
parameter of the model, which is called regression coefficient, defining the effects of the covariates.
With the hazard function in Equation (10), the pdf of the failure time can be defined as

! ℎ e
S h e
-./ hi −ℎ j kj l (11)
c
where T -./ mnc −ℎ j kj o is the survival function. p denotes an indicator function taking

value 1 if the system is failed at , or taking value 0 if it is censored. Note that censored means that
the equipment doesn’t have a failure event. The Cox PH model is frequently used in medical

11
we can see that if b is extended to b q , the degradation signals can be easily incorporated into the
statistics and has been extended to the manufacturing field in recent years. From the Equation (10)

equation by treating degradation data as a time-varying covariate. It is beneficial in reliability

different degradation level. The functional form for b q can be general path model [49], Wiener
analysis for hard failure systems, where each equipment runs to fail, so that different unit has

process [50, 51], multivariate Gaussian convolution process [52]. The Cox PH model, a semi-
parametric approach, is more robust than other parametric approaches as it is not vulnerable to
misspecification of the baseline hazard. But the proportional hazard assumption may limit its
application to accounting for complex relationships among covariates. Recently, deep learning
based Cox models have been implemented to relax the proportional hazard assumption [53, 54].

2.2 Conventional Machine Learning based Models

Though machine learning has been around for several decades, it has seen a revival in recent
years due to the dominance of data stemming from the information explosion. In the following
subsection, the various machine learning algorithms have been reviewed and discussed from the
predictive maintenance perspective.
2.2.1 Support Vector Machine
Support Vector Machine (SVM) was initially established as a methodology to be used in the
binary classification problem and then is applied to solve the regression problem. When it is
applied to a regression problem it is termed as Support Vector Regression (SVR). The key purpose
for SVR is to get a functional relationship between input and output under the hypothesis that the
joint distribution of the input and output is not defined and unknown. SVR uses a complex penalty
function that a penalty cannot be enacted if the predicted value is farther away from the real value.
The restricted region is called an insensitive tube [55]. Then support vectors are then fitted to
regression models and apply to predict the degradation level and calculate the corresponding RUL
values [56]. Benkedjouh et al. [57] used SVR and the isometric feature mapping reduction
technique to predict the RUL for rotating machines. Hu et al. [58] built an RUL prediction method
based on fuzzy C-mean clustering and wavelet SVM. Shen et al. [59] designed an SVR based on
a generic multi-class solver to recognize the different faults pattern of rotating machinery. Liu et
al. [60] proposed an improved probabilistic SVM regression technique to predict the condition of
Nuclear Power Plant elements. A comprehensive review on SVM-based estimation of RUL can
be found in reference [61]. SVM is very effective in high dimensional spaces and works well in

12
cases where the number of dimensions is greater than the number of samples. More importantly,
SVM is relatively memory efficient, which is a great advantage for online modeling. However,
when the noise level is high, the performance may decrease significantly.
2.2.2 Decision Tree
The Decision Tree (DT) is a non-parametric supervised technique based on a tree-like model
for regression and classification. The key purpose of DT is to predict the value of an objective
variable by establishing a hierarchical structure composed of nodes extracted from a training
dataset. A DT generally consists of one root, several branches, and many interval nodes. Every
path is from the root node to a leaf node through the internal nodes. This path denotes a
classification with the different conditions of the components or systems. Every leaf node
represents a response for regression or a class label for classification. To extend the power of DT,
some variants have been developed, such as gradient boosting decision tree (GBDT), random
forest (RF). The RF is a term for a collaborative approach of DT, which consists of numerous trees.
Unlike classical methods that build a single tree on a whole dataset, RF randomly chooses the
features and instances to build multiple trees. Each DT then votes for a particular target class and
a class having the bulk votes is the model’s prediction with a certain probability. In contrast to a
traditional DT, RF demonstrates good predictive performance with considerable noise made by
random selection of instances and features. Furthermore, it can deal with large datasets having
numerous features with diverse data types, e.g., continuous or categorical values. Kundu et al. [62]
presented an RF regression methodology for RUL prediction for spur gears depending on pitting
failure mode. GBDT is an iteratively accumulative decision tree method. The algorithm
accumulates the results of multiple decision trees as the final prediction output by creating a group
of weak learners. Wang et al. [63] developed a GBDT model to estimate the RUL by choosing
fault features and measuring fault severity subjected to relative entropy distance in fault prediction
of electronic circuits. DT requires less effort for data preparation during pre-processing, and it is
very intuitive and easy to explain. But we need to be careful that a slight change in the data can
cause a significant change in the structure of the decision tree, which makes the model instable.
The calculation of the tree can go far more complex compared to other algorithms.
2.2.3 Back Propagation Neural Network
Back Propagation Neural Network (BPNN) is a supervised-learning method implemented by
iterative optimization to solve the classification or regression problem. Usually, it takes a vector

13
as the input, and outputs is a label representing the information of corresponding classes or function
value. It firstly calculates the model results through the forward propagation step and then tunes
the network's weights through the back propagation step. The two steps above can be executed
iteratively until the errors between the model results and the label reduce to a desired threshold.
The ultimate target of the BPNN is to get the network parameters representing the relation between
the input and output by minimizing a corresponding loss function. BPNN usually used the squared
error sum (SES) for the network as an objective function and applied the gradient descent
technique to get the objective function’s minimum value. The BPNN, just like other NNs, is
flexible and powerful to find the nonlinear mapping between inputs and outputs, and it doesn’t
require prior knowledge about the network. But back propagation is notorious for the easily getting
stuck in “local minima”.

2.3 Deep Learning based Model

In recent years, deep learning approaches have shown excellent performance in various
applications ranging from feature extraction, defect detection, segmentation, medical imaging,
additive manufacturing, and many more [64-70]. Realizing the promising ability, researchers have
experimented on various deep architectures to develop the solution approach in remaining useful
life prediction. In the following subsection, we discussed the architecture of deep Convolutional
Neural Networks (CNNs) and their variants from the predictive maintenance point of view.

2.3.1 Convolutional Neural Network

Due to the ability to generalize the local and global features, CNNs turn out to be the most
popular deep learning methods. CNNs are exceptionally successful in extracting features from
input data and using them to make a trustworthy prediction. A basic CNN structure mainly has an
input layer, convolution layer, pooling layer, and fully connected layer as shown in Figure 2.

14
 Figure 2 Architecture of CNN
The input data could be either two-dimensional or one-dimensional such as time-frequency
spectrum or time series data, respectively. The convolution layer uses a set of weights and

of the convolutional layer is calculated as: Hr ! ∗ tr &r , where ∗ represents an operator

convolutes at each layer to form the layer-wise features, which are called a feature map. The output

of the convolution, u denotes the number of convolution filters, tr is the weight matrix, &r is the
filter kernel bias. Following the convolution, the model parameters are reduced by subsampling,
named as pooling process. After the pooling layer, multiple fully connected layers are used to
convert the matrix to a row or a column. Finally, a classification or regression layer is added to get
the predictions or results. To predict the RUL, CNN can be used to extract useful and robust
features from data. The number of processing units or the CNN structure greatly depends on the
nature of problems and datasets. To expand the power of CNN, several variants of CNN-based
models have been introduced in the literature as reported in Table II. The key benefit of using
CNNs is to extract complex, non-linear, non-handcrafted features because of the superior feature
extraction and object recognition performances of CNN. Based on our observations, only a few
research works focus on pure CNNs for RUL prediction as listed in Table II. The reason is that,
CNNs may not sufficiently model the temporal characteristics of time series data. Moreover, CNN
is significantly slower due to a convolutional operation and requires a lot of data to train
effectively. Also, the tuning to find the proper learning rate for the CNN methods on real-world
applications is difficult [71].

15
 Table II Variants of CNNs and their use for RUL prediction
Variants Distinctions References
• Babu et al. [72] combined a regressor
with a deep CNN architecture to estimate
• It consists of different processing units at multiple
the RUL from multivariate time series
layers (usually have 5 to 10 layers, even more)
Deep CNN data.
• Effective in capturing the salient patterns in the
• Ren et al. [73] fused a smoothing method
signals
with a CNN built a CNN for predicting
the bearing RUL.
• Multiple CNN architectures are stacked together
• Yang et al. [74] proposed a double-CNN
• The output of the previous CNN becomes input of
model architecture to predict RUL using
other CNNs
Deep Multi CNN original vibration signals without
• Effective in dealing with raw signals, instead of
resorting to any feature extractor.
depending on the feature extractor
• Kiranyaz et al. [75] utilized the MSCNN
• MSCNN framework has three sequential stages:
for fault detection and identification for a
transformation, local convolution, and full
circuit monitoring system
convolution.
• The transformation stage applies transformations • Zhu et al. [76] used the wavelet transform
to propose Time Frequency
on the input time series.
Representation (TFR), then applied this
Deep Multi-scale • In the local convolution stage, extract the features
TFR to MSCNN to perform RUL
CNN (MSCNN) for each branch.
estimation.
• The full convolution stage concatenates all
• Li et al. [77] used MSCNN for RUL
extracted features and applies several more
prediction, the model has three multi-
convolutional layers to generate the final output.
scale blocks, where three different sizes
• Effective to keep the multiple levels of abstraction
of convolution operations are put on each
for the prediction
block in parallel.
• This CNN architecture is mainly the combination
of above-mentioned CNN along with additional
• Wen et. al [85] proposed a new residual
supporting layers.
Hybrid CNN CNN (ResCNN) by adding a skip
• Incorporates the advantages of different
(HCNN) connection between convolution blocks
methodologies by their integration to improve the
prediction performance.

2.3.2 Recurrent Neural Network

The underlying motivation behind Recurrent Neural Network (RNN) is to mine the sequential
information for any given dataset. It creates memory cells that capture the past and predict the
future sequence based on the previous computation. A typical RNN is shown in Figure 3. As shown
in Figure 3, the structure of the RNN constitutes a deep network with one layer per time step and
shares the parameters across the layers. The concept of parameter sharing is a useful way to capture
the relationship between one input item and its neighboring context. This makes the RNNs very
successful over the traditional NNs and CNNs. The network can be trained in a similar fashion of
backpropagation across the time steps. However, the training process is especially challenging due
to the problem of gradients vanishing or exploding. To overcome this issue, long short-term
memory networks (LSTM) is constructed [78]. LSTMs are a special kind of RNN for remembering

16
information for long periods (long-term dependencies) and are explicitly designed to avoid the
problem of standard RNN. Similar to standard RNNs, LSTMs also possess chain-like structures,
but they differ in the structure of the memory cell. Instead of having a single neural network layer,
LSTMs have four interacting networks connected in a very tricky way to remove or add
information to the cell state by regulating the structure of different gates. Gates indicate a special
setup to control the information passing to the cell state and output at each repeating module. Gated
Recurrent Unit (GRU) is another modified LSTM cell, which was introduced by Cho et al. [79].
Recently, this architecture showed its promising application in the field of RUL prediction [80-
82]. GRU combines the input gate and the forget gate into the update gate. It also merges with
cellular sate and hidden state. A comparison of the memory cell in standard RNN, LSTM and GRU
LSTM is shown in Figure 4.

Figure 3 Structure of a typical RNN

(a) (b) (c)

Figure 4 Memory cell structure, (a) standard RNN; (b) LSTM; (c) GRU LSTM

17
 Table III Variants of RNNs and their use for RUL prediction
Variants Distinctions References
• Heimes [83] utilized the RNN incorporating with
• Can utilize the sequential Kalman filtering training and evolutionary algorithm
information for prognostics problem.
Recurrent
• Able to retain the short-term • Liu et al. [84] proposed an adaptive RNN for dynamic
Neural Network
information state forecasting to leverage the RUL prediction for
(RNN)
• Able to capture the temporal Lithium-ion-batteries.
correlations in sequence data • Liang et al. [85] proposed a RNN based health
indicator for RUL prediction of bearings.
• Zhang et al. [86] employed LSTM RNN to learn the
• Solve the vanishing gradient or
long-term dependencies among the degraded
exploding problem
capacities and construct a RUL predictor.
Long Short- • Able to retain both Long and
• Zhao et al. [87] developed an RUL predictor using the
Term Memory short term information
LSTM, which can evaluate the trend features.
(LSTM) • It is easy to detect and capture
• Xiang et. al. [88] introduced a new type of LSTM with
important features over a long
weight amplification for accurate prediction of gear
distance
remaining life.
• Can deal with long term
• Song et al. [80] proposed a battery RUL prediction
relationship in RNN
Gated approach based on the RNN with gated recurrent unit.
• Able to capture dependencies
Recurrent Unit • Chen et al. [81] incorporated kernel principal
at different time scales
(GRU) component analysis and GRU RNN to predict RUL.
• Able to capture the inherent
RNN/LSTM • Wang et al. [82] presented a hybrid RUL prediction
relation for long-term
model by adapting a deep heterogeneous GRU model.
prediction
• Utilize the information in both • Zhang et al. [89] proposed a transfer learning based
Bi-directional
forward and backward BLSTM network for turbofan engine RUL prediction.
Long Short-
direction • Wang et al. [90] proposed a data-driven approach with
Term Memory
• Suitable for intermediate BLSTM network for RUL prediction, which can make
(BLSTM)
prediction full use of sensor data sequence in bidirectional.
Bi-directional • Can utilize two sets of LSTM
• Elsheikh et al. [91] proposed the BHLSTM to predict
Handshaking cells in reverse order
the RUL of a system, which is capable to process
Long Short- • Allow forward and backward
maximum information for any given subset of
Term Memory unit collaboration in the
sequence.
(BHLSTM) learning process

The horizontal line on the top of the repeating module indicates the cell state, which passes
through the entire network chain with some minor liner interactions. They are comprised of a
sigmoid neural net layer and a pointwise multiplication and addition operation. The sigmoid layer
maps numbers between zero and one, where zero means no information will pass through and the
value of one allows all information. Recently many researchers have introduced several variations
based on the original RNN and LSTM. Among them, some major variants are Gated Recurrent
Unit (GRU) RNN, Bi-Directional LSTM (BiLSTM), and Bi-Directional Handshaking LSTM
(BHLSTM). Table III summarizes the major distinction for these variants and includes some
representative publications for each variant as references. Several studies have proven that RNNs
and LSTMs are outstanding and perform better than many conventional machine learning

18
approaches and even better than CNNs for the RUL prediction tasks [92]. However, as some
researchers pointed out that the LSTM network may not be robust when processing raw time series
data directly since the sensor data usually contains noise [93].
2.3.3 Autoencoder
An autoencoder is an unsupervised neural network. The main idea of autoencoder is to train the
model to reconstruct the original input at the output layer [94]. The autoencoder network consists
of three layers: an input layer, a hidden layer for encoding, and an output layer for decoding as

the network is forced to learn a compressed representation of the input. Define the input as v, an
shown in Figure 5. The size of the hidden neurons is usually smaller than the input. In this way,

encoder function w . parameterized by y , a decoder function ! . parameterized by z , the

output as v′, then the reconstructed input is v′ !{ w| v . The parameters z, y are trained to
minimize the reconstruction error so that the output is similar to the original input, i.e., v ≈ v′.
Several variants have been developed to solve different issues based on the basic mode. If the
number of network parameters is larger than the number of inputs, the basic autoencoder will face
overfitting issues. To avoid overfitting and improve the robustness, denoising autoencoder is
developed by randomly changing some of the input values to zero, then loss function is designed
in a way that the output values are compared with the tampered input instead of original input.
Another tactic to control the number of hidden nodes is sparse autoencoder, where a sparse
constraint is added to limit the activation of its nodes. With sparse constraint, some nodes in the
hidden layer are active and the other nodes are inactive. This constraint is achieved by adding a
penalty term into the loss function. The third variant is called variational autoencoder [88]. The
basic idea is that instead of mapping an input to a fixed vector, input is mapped to a distribution.
When decoding from the layer, samples from each distribution are randomly selected to generate
a vector. Variational autoencoder provides a probabilistic manner for describing the input. In the
predictive maintenance field, autoencoders currently are mainly used to reduce the dimension and
eliminate the redundancy of the data. Autoencoders are commonly employed as a feature extractor
or a tool for health index construction. Some examples can be found in [95-99]. To learn sequential
information from input signals, LSTM autoencoder is a preferable choice. As an autoencoder
learns to capture as much information as possible rather than as much relevant information as
possible, it may misunderstand important variables of the input data.

19
 Figure 5 Architecture of an autoencoder
2.3.4 Bayesian Deep Learning
Bayesian Deep learning method is an extension of deep learning in a probabilistic manner. The
fundamental idea is to adopt Bayesian inference as the learning tool for quantifying uncertainty of
the model by treating the deep learning architectures as probabilistic models. Typically, a Bayesian
Deep learning places prior distributions over the network's weights and then learn the
corresponding posterior distributions over the weights. Then each forward pass will have different
weights and therefore providing potentially different outputs. As exact Bayesian inference is
computationally intractable for most of the NN structures, some sampling strategies are used to
learn the parameters of Bayesian deep learning models, which is often computationally expensive.
From the literature review, we observed that some popular approximations have been proven
effective to alleviate the computational burden, such as variational inference, expectation
propagation, Laplace approximation, Hamiltonian methods, bootstrapping, Monte Carlo dropout,
etc. Among these, Monte Carlo dropout, which combines approximate Bayesian inference with
dropout, has drawn considerable attention in many research fields due to its simplicity, scalability,
and computational efficiency. It is well known that data-driven models inevitably face two types
of uncertainties: aleatoric uncertainty, reflecting the noise pollution in data collection and
transmission, and epistemic uncertainty, reflecting the ignorance of model property. To address
these issues, a variety of researchers have been exploring Bayesian Deep learning to account for
the uncertainties to improve model accuracy. References [100, 101] use Bayesian deep learning
based methods to quantify the uncertainty of point prediction for bearings and Gas turbine engine.
Li et al. [102] proposed a Bayesian deep learning based methods, in which a sequential Bayesian
boosting algorithm was executed to improve the prediction accuracy. A Bayesian deep learning
model can be treated as an ensemble of multiple models, which may naturally reduce the risks of

20
over-fitting issues. Another benefit of Bayesian deep learning models is that they allow to quantify
the uncertainty, which is very important for RUL prediction considering the limited data
availability and the stochastic nature of degradation processes in the manufacturing field. But the
computational cost is heavier for online inference comparing with other deep learning models as
the model needs to run multiple times to get the distribution of outputs.
2.3.5 Transfer Learning
Deep learning models excel at learning from a large number of labeled examples, learn a very
accurate mapping from the inputs to outputs. But it lacks capability to generalize to different
application scenarios. The reason is that many machine learning algorithms assume that the
training and test data are in the same feature space and have the same distribution. This assumption
may not hold in many real-world applications. Transfer learning is developed to tackle this issue
through storing knowledge learned from one domain (called the source domain) and transferring

can be defined as I {~, 5 }, where ~ represents a feature space, 5

it to a different but related problem (called the target domain). Mathematically speaking, a domain

{., , . … , .r } ∈ ~. A task can be defined as C { , ! . }, where

is a marginal

label space, ! ⋅ is a predictive function. Given a source domain I• and learning task C• , a target
distribution where is a

domain I and learning task C , where I• ≠ I or C• ≠ C . Transfer learning aims to help improve
the learning of the target predictive distribution based on I• and C• . Based on what to transfer,
transfer learning can be conducted at several levels: instance-transfer, feature-representation-
transfer, parameter-transfer, and rational-knowledge-transfer [103]. Instance-transfer tries to
reweight some the samples from the source domain in an attempt to correct for the distribution
difference, then apply them in the target domain for training. Feature-representation-transfer tries
to get good feature representations that can reduce the difference between the source domain and
the target domain. Parameter-transfer discovers shared parameters or prior knowledge between the
source domain and the target domain. Parameter-transfer models believe that a well-trained model
on the source domain has learned a well-defined structure, and if two tasks are related, this
structure can be transferred to the target model. Relational-knowledge-transfer works by mapping
some similar patterns from the inputs to the outputs between both domains. Examples for RUL
prediction based on transfer learning can be found from [89, 98, 104-106]. Transfer learning is
especially useful in the situation where source data and target data are in different feature spaces
or have different distributions in which training data for the target problem are limited but data for

21
a related problem are abundant. Transfer learning provides an effective way for RUL prediction
with limited historical failure data. However, transfer learning only performs better under the
condition that the domain and target problems of both models are similar enough. Otherwise, it
will end up with a negative transfer. Currently, it is still challenging to find solutions to negative
transfer.

3. Applications in Predictive Maintenance

In this section, we first provide an overview of the workflow of the predictive maintenance and
how the prognostic approaches are applied for predictive maintenance. Then we share some
statistics and give a comprehensive review of applying RUL predictions to various fields.

Figure 6 Workflow of predictive maintenance

Figure 6 summarizes the workflow of the predictive maintenance. First, data are collected
intermittently or continuously from an interactive physical system of interest. Various sensors are
installed to collect the degradation signals in semi-observable or fully online systems. Some
commonly used sensors are pressure sensors, force sensors, speed sensors, temperature sensors,
torque sensors, proximity probes, accelerometers, etc. Apart from the sensor data, quantitative data
are sometimes collected based on the purpose and application domain. For example, the RUL of
many power storage systems depends on the charging and discharging cycles. In such cases, a
number of cycles are recorded to collect the data. Following the data collection and processing,
feature extraction plays a critical role in model development and RUL prediction. The
straightforward use of raw data is inconvenient due to the high complexity and nonlinearity of
sensor signals. Hence, the underlying motivation of feature extraction is to utilize the patterns and
trends in the sensor signals to predict RUL. Literature in RUL prediction has focused on many
time-domain and frequency-domain features such as root mean square, kurtosis, short-time Fourier

22
transform [107], wavelet transform [108], empirical mode decomposition [109]. Recently, a
number of machine learning based approaches have been utilized to learn the learn the features
and mapping the raw signals to the associated RUL. Several such machine learning algorithms are
described in Subsection 2.2. These machine learning algorithms are that they are capable to extract
the useful features and information within the data with a very limited human intervention.
Subsection 2.3 demonstrates several deep learning architectures, which are emerging and highly
effective techniques for patterns and trends recognition. Their deep networks are capable of
obtaining high-level abstractions of data to improve the performance in intelligent prognostics.
The best part of these deep learning architectures is that they are capable of feature extraction
without human intervention. The extracted features are then used as inputs of developed RUL
prediction model. Once the predicted RUL is obtained, the last step is to decide the optimal time
to send out an alert of failure to help maintenance decision making. If the predicted time is shorter
than the actual time, maintenance will be implemented earlier, the benefits of more extended usage
are lost. If the prediction is too late, the equipment may fail and result in a more significant loss.
To determine the optimal maintenance policy, the inventory management associated with the spare
parts cost also needs to be considered in practice. Timely acquisition of the inventory number and
status of spare parts is challenging. Some researchers have investigated this issue by joint
optimization of maintenance and inventory management [110-112]. Another exigent issue in the
decision making process is that spare parts inevitably deteriorate over time due to the inner
mechanism and imperfect storage conditions, which will shorten the storage lifetime and
eventually affect inventory management. How to estimate the storage lifetime with arbitrary
number of spare parts based on the operating and storage degradation processes is attracting
researchers’ attention [113, 114]. Interested readers are referred to the recent articles for more
details [113, 114]. In this paper, we mainly focus on the RUL prediction in predictive maintenance.
Over the last decade, researchers showed intensive interest in RUL prediction for predictive
maintenance and published excellent research papers in this area. However, we only focus on the
recent advancement and applications within the time frame of the year 2015 to 2020. We reviewed
253 published papers under a broad point of view of data-driven approaches such as statistical
approaches, deep learning, conventional artificial intelligence and hybrid approaches (combination
of statistical and AI-based methods). The papers were collected from Google scholar using a
thorough key word search of “remaining useful life (RUL)”, “predictive maintenance (PM)”,

23
“prognostic and health management (PHM)”, “health index (HI)” and “machine health”. It is
observed that the concept of predictive maintenance and RUL prediction are widely applied for
aircraft engines, bearings, gears, motors, machine tools, wind turbines, batteries, computer hard
disks, and many more. Based on our observations, we categorized all of these applications into
four broad categories as “aircrafts”, “rotating machinery”, “power systems” and “electronic
systems”. We also break down the proportion of published papers in these four categories as shown
in Figure 7. Surprisingly, most of the researchers showed their interest in the field of rotating
machineries (48%) and aircraft system (23%) applications. One potential reason for this is the
availability of standard datasets in these two fields. We found that the NASA C-MAPSS,
PRONOSTIA, and Center of Intelligence at the University of Cincinnati dataset are widely used
in this area. However, some researchers are also trying to generate the dataset from their own lab
and focusing on the field of the power and electronic system components and products.

Figure 7 List of recent application fields of predictive maintenance

We also capture the trend (number of published papers) of different data-driven methods for
the individual applications as shown in Figure 8. We can see that the number of publications based
on statistical approaches is notably large except for hard drives. Due to persuasive mathematical
properties and physical interpretations, also the capabilities to capture the uncertainty of
parameters, statistical approaches have been attracted widespread attention. However, some
statistical models are complex and can be computationally expensive. Due to the fact that deep
learning provides good functional mappings between inputs and outputs, which is powerful to
capture dynamic information, the number of deep learning prognostic models is also escalating.

24
The main disadvantage of deep learning is the operating and training processes are “black boxes”
and a majority of proposed approaches are only focused on prediction on a population level, it is
hard to catch the uncertainty and individual heterogeneity. It would be interesting to combine the
advantages of each approach in the future. In the following subsection, we will provide details
review of research of RUL prediction for those applications (i.e., aircraft, rotating machinery,
electronic systems, power systems) of predictive maintenance. For each field, we first describe the
popular and open access dataset, then summarize the contemporary research work in the recent
five years in the corresponding field. It is worth noting that the number of papers on data-driven
prognostics is enormous. Consequently, some papers would be omitted inevitably.

Figure 8 Recent research trends (2015-2020) on different application fields & methods

3.1 Application in Rotating machinery

Bearing, gears, motor, and shaft are the most common rotating parts which are widely used in
machinery and industry. Rotating machinery is critical for the machine heath and ensures safety
as any rotating parts' flaws can lead to significant consequences and failure. The identified
significant causes of bearing failures are generally subject to excessive balance, improper
installations, poor lubrication practices, alignment tolerances, poor storage, and handling
techniques. RUL prediction of bearings and gears is derived based on the vibration or mechanical
signals obtained from the bearings. In literature, we found two experimental setups for the run-to-
failure and historical data collection of bearings and gears: The Center of Intelligence at the
University of Cincinnati [134] and PRONOSTIA [115]. Most of the researches in the field of
bearings and gears used the datasets from these two experimental setups. Lubricant is another

25
important substance for rotating machinery. It helps to operate machines in safe and healthy
conditions. The deterioration trends can be described by observing the characteristics of lubricants
and the operation conditions of the machine. Over time and usage, lubricants degrade and produce
acidic substances, moisture, and insoluble deposits, such as carbon deposits, sludges [116], which
lead to machine failure. Hence, it is necessary to understand the deterioration trend in advance to
reduce wear and friction from the mobile components and avoid machine failure. Recently,
condition monitoring of lubricating oil has attracted considerable attention in research. The
lubricant data could be obtained from either a physical experiment or simulation procedure by
mimicking the real environment under a certain assumption. One of the procedures of collecting
lubricant data is the four-ball test found in [117]. To obtain real-time data of lubricating oil, a wear
debris sensor and an oil property sensor were employed in the oil cycling line. Besides, a
temperature control system was set to offer a fixed temperature for the oil property sensor. By
doing this, wear conditions and dynamic viscosity and permittivity of lubricating oil can be
monitored simultaneously. To accelerate oil degradation, a test composed of various loads and
speeds was carried out. The test was stopped and restarted when working conditions were changed.
Another lubricant data was found in [118]. The lubricating oil used for large machines was
collected from the engine of a loader. In the following paragraphs, we articulate different research
methodologies and approaches win this application field along with the description of the
commonly used datasets in this area.
Lei et al. proposed a two-stage method based on a particle filtering algorithm to predict the
bearing RUL [119]. They fused multiple features to construct new health indications and then used
the maximum-likelihood estimation to initialize the model parameters. Gaussian Process
Regression (GPR) is another effective method that was used in [120] with the integration of
composite kernels. RMS, Kurtosis and Crest factor are used for feature fusion by self-organizing
map. It is experimentally demonstrated that integration of composite kernels improves the
prediction accuracy than particle filter method. Liu et al. [121] divided the entire bearing life and
built an individual local regression model to get the multiple health states to leverage the RUL
prediction. This constituted a semi-supervised approach that can be utilized without having any
prior knowledge. Wang et al. [122] proposed a Wiener process model with stage correlation and a
Bayesian approach to utilize the prior distribution information into the model parameter. Ahmad
et al. [123] proposed a dynamic regression model to capture the trend of bearing health indicator,

26
which is later used to project the future health indicator value and estimate the bearing useful life.
The adaptive regression model can determine an appropriate time to start prediction which yields
excellent prognostics performance. Kundu et al. [124] used a clustering and change point detection
algorithm to identify the failure behaviors and predict the RUL. Sometimes bearing could be
operated under multiple conditions. Integrating these multiple conditions would provide better
prognosis result.
Literature also states several artificial intelligence based approaches to construct the health
indicators for bearing and gear remaining life prediction. Zhao et al. [125] employed the principal
component analysis (PCA) and linear discriminant analysis for dimensionality reduction. Then
they used a multiple linear regression model to estimate the RUL. In some applications, it is hard
to obtain the failure or suspension histories which make the RUL prediction tasks more
challenging. To address this issue, Xiao et al. [126] developed an inference method using recorded
condition monitoring data. An adaptive time windows was employed to divide the extracted
features and to train an ANN for intelligent prognosis. Bastami et al. [127] used the wavelet packet
transform to extract signal features and later trained an ANN to estimate the RUL of rolling
bearings. The nonlinear nonparametric approach is also considered as a very appealing technique
in predicting the RUL of bearings. The ensemble technique is another effective machine learning
paradigm that could be used in RUL prediction. Such an ensemble approach, decision tree-based
random forest, was proposed in [62] by Kundu et. al. for monitoring and detecting the pitting
progression in spur gears to predict the RUL of gear.
In recent years, deep learning becomes very effective in machine health monitoring and
prognostics due to its capability of learning representation from raw data. With the development
of deep learning methods, Guo et al. [85] proposed a deep neural network structure named
recurrent neural network based health indicator (RNN-HI), where several classical time-frequency
features are combined with the original feature set to get the most sensitive features as the input of
RNN-HI model to leverage the RUL prediction of bearings. Deep learning methods can effectively
extract the discriminative features for monitoring bearing fault. However, temporal information
also plays a critical role in the fault degradation process, which was not considered in many cases.
Mao et al. [128] first considered this temporal information in bearing RUL prediction and proposed
the LSTM. In another research, Tang et al. [129] proposed an LSTM approach combining the
bottleneck features to develop a novel prediction method of bearing performance degradation. In

27
LSTM approaches, first, the feature parameters are extracted from the different domains such as
time domain, frequency domain, time-frequency domain. Then the important features are extracted
from the original feature set that could better represent the degradation process of bearings. Finally,
the selected features are used to train the LSTM network to predict the bearing RUL. Traditionally
the feature extraction is derived from prior knowledge and is separated from the RUL models. Ren
et al. [130] proposed a Multi-scale Dense Gated Recurrent Unit Network (MDGRU) to combine
the feature extraction into the RUL model by pre-trained Restricted Boltzmann Machine network,
multi-scale layers, skip gate recurrent unit layers, dense layers. In [131], Li et al. used a CNN to
explore the time-frequency domain information and to extract multi-scale features. Deep learning
approaches showed the limitation on predicting less stability for the single sensory information.
To address this issue, Wu and Zhang [132] proposed a new cascade fusion convolutional long-
short term memory to fuse the information streams in the form of an ensemble model. Lo et al.
[133] proposed a one-dimensional CNN for the prognosis of bearing and gear. The network was
trained in a hybrid fashion where both the classification loss and clustering loss were combined to
estimate the status of prognosis. Xiang et al. [88] proposed an attention based LSTM named
LSTM-A. This special type of network utilizes an attention mechanism to amplify the input and
hidden layer weights at different degrees to accurately predict the gear remaining life.
The RUL prediction of lubricant is accomplished based on the parameters obtained from oil and
degradation trends. Tanwar et al. [134] proposed a degradation model based on continuous time
stochastic process, i.e., the Wiener process for lubricating oil degradation tracking and RUL
prediction under regular oil top-up effects. In this research, the Oil Replenished Effect was
neglected in the prediction of lubricant remaining life. Tanwar and Raghavan addressed this issue
in [135] and proposed the use of the GPR model as a non-parametric Bayesian method. Recently,
the machine learning approach is also being incorporated for condition monitoring from lubricant
oil. In [136], the researcher used the machine learning approach to classify the engine lubricant
into three conditions as normal, degraded, and unsuitable. They used a cohort of military land
vehicles to collect the data from laboratory test results of lubricants and monitoring system of
vehicle health. The proposed machine learning procedure used feature selection methods to
identify the best feature set for representing the lubricant oil condition.

28
3.2 Application in Aircrafts
An aircraft engine is the power component of an aircraft propulsion system. Most aircraft
engines are either piston engines or gas turbines. An aircraft engine produces thrust to propel an
aircraft. Aircraft engine failures may result in significant economic losses and even accidents in
extreme cases. Except for the aircraft engine, another two of the most important systems in the
aircraft are aircraft auxiliary power unit (APU) and actuators [114]. APU is a small turbine engine
installed under the tail of an aircraft. Instead of providing propulsion, its main function is to supply
power at a certain flight altitude and provide bleed air for the cabin air condition system on the
ground. For some aircrafts, APU can also provide compressed air and backup electric power to
compensate for the effect of dead engines. Thus, monitoring of the health state to ensure safety
and operation efficiency is essential. Actuators (e.g., Electro-Mechanical Actuators (EMA) and
Electro-Hydraulic Actuators (EHA)) play an active role in control systems in aircrafts. They are
used to convert electrical signals to mechanical movement or other physical variables, such as
pressure or temperature. In this research field, a majority of publications use the Commercial
Modular Aero-Propulsion System Simulation (C-MAPSS), which is collected from National
Aeronautics and Space Administration (NASA) [137]. The C-MAPSS dataset includes four sub-
datasets. All engines work in normal condition first and then degrade continuously until a failure
criterion is reached. Every record of the engine state is generated by a set of 24 variables, three of
which are operational settings and the other 21 are for engine performance measurements.
Currently, there is no public dataset available. Most researchers investigate the performance based
on data collected from commercial aircraft fleets. The research in this filed has been summarized
in the following paragraphs.
Chehade et al. [138] predicted RUL through individual failure threshold distribution estimation.
They developed a convex quadratic formulation that combines the historical population
information and the condition monitoring data of an operating unit to online estimate its failure
threshold. Some efforts have focused on developing data fusion methodologies for prognostics
[139-141]. The main idea is to construct a health index via selecting and fusing multiple
degradation signals to track the trajectories of the degradation process. After that, the constructed
health index was treated as another sensor signal and then was used for degradation modeling and
prognostics. Song and Liu [140] solved the HI construction by the quantile regression technique.
Kim et al. [139] proposed a latent linear model for HI construction and a systematic sensor

29
selection procedure for RUL prediction. Chehade et al. [141] extended the data-level fusion
techniques to multiple failure mode scenarios. Using constructed HI, Li et al. [142] developed an
age- and state-dependent Wiener-process model for RUL prediction with the consideration of the
unit-to-unit variability. Son et al. [143] proposed a non-homogeneous gamma process based RUL
prediction method. The model considered noisy degradation data and by using the Gibbs sampling
technique, the hidden degradation states were approximated by using the Gibbs sampling
technique. All of the above mentioned methods were developed based on statistical approaches.
Researchers also investigated the machine learning approaches to develop more sophisticated
methods. For example, Ordóñez Celestino et al. [144] proposed a hybrid ARIMA (auto-regressive
integrated moving average)-SVM model for estimating the RUL. ARIMA model is utilized to
estimate the values of the predictor variables in advance. Then, the result of ARIMA is applied as
the input of a support vector regression model. Al-Dulaimi et al. [145] proposed a hybrid of LSTM
and CNN framework. In their model, LSTM and CNN are constructed in parallel followed by a
fully connected multilayer fusion neural network. Zheng et al. [146] proposed an LSTM approach
for RUL estimation, which fully utilizes the sensor sequence information and uncovers hidden
patterns. Badu et al. [72] developed a deep CNN-based regression method to predict the RUL. The
convolution and pooling filters were used along the temporal dimension over the multi-channel
sensor data to integrate automated feature learning from raw sensor signals. Wen et al. [147]
proposed a residual convolutional neural network (ResCNN), it can help overcome vanishing or
exploding gradient problem of deep learning algorithms. Song et al. [96] developed a autoencoder-
BLSTM hybrid model. racy of RUL. Autoencoder was used as a feature extractor to reduce the
dimension of data. BLSTM was designed to capture the bidirectional long-range dependencies. It
showed that the hybrid model had better prediction performance comparing with most existing
methods including CNN and LSTM.
To predict the RUL of APU, Chen et al. [148] developed a Gaussian process regression model
combined with ensemble empirical mode decomposition. Liu et al.[149] utilized an Extreme
learning machine (ELM) to predict the degradation of an APU. They employed a restricted
Boltzmann machine (RBM) to optimize the ELM. Wang et al. [150] derived a health index to
characterize the APU degradation and then used a Bayesian framework for the RUL prediction.
Zhang et al. [151] utilized a Weibull-based generalized renewal process to implement failure rate
prediction of APU. Researchers also put their effort for actuators prognostics. Zhang et al. [152]

30
proposed a weighted bagging GPR algorithm. With the idea of ensemble learning, the weighted
bagging GPR algorithm uses a series of subsets to train the GPR model. They found that the
proposed method can take the randomness of data into consideration. Then Zhang et al. [153]
proposed a feature-aided Kalman Filter method for motor voltage estimation, which is an essential
parameter for performance degradation assessment of EMA. The dataset they collected through
Flyable Electromechanical Actuator, which was made by NASA Ames Research Center. Guo et
al. [154] presented an optimized incremental learning and on-line training algorithm based on the
relevance vector machine for EHA RUL prediction. In their research, sample entropy was
introduced as an effective signature of the EHA’s health.

3.3 Application in Power systems

Wind energy generated by wind turbines is a growing and reliable renewable energy source in
the world. However, the wind energy industry experiences increasing operation & maintenance
costs because of main components failures. The temperature stress caused by the temperature
difference along with the machine, e.g., shafts’ and gears’ temperature, together with lubrication
problems accelerate wind turbine faults. To monitor the health of wind turbines, various
monitoring techniques have been used, such as acoustic measurement, electrical effects monitoring,
equipment vibration monitoring, power temperature monitoring, oil debris monitoring, etc.
SCADA (Supervisory Control and Data Acquisition) is a commonly used tool for data collection,
which is a system built into turbines to control electricity generation. This system use sensors to
collect various functional parameters and data, such as temperature, bearing vibration, wind speed,
and phase currents of wind turbines [155]. Carroll et al. [156] ensembled ANN, SVM, and logistic
regression to predict wind turbine gearbox failure using SCADA data. This methodology appears
to be effective in predicting the failure up to a month before it occurs. Inclusion of high frequency
vibration data could extend that prediction capability to 5-6 months before failure occurs with
reasonable accuracy. Song et al. [157] introduced a Bayesian framework with three different
methods, namely, the bin method, the multivariate normal distribution based method, and the
copula method to identify wind turbine health states based on their SCADA data. The results
showed that copula method has the best prediction performance. Chen et al. [158] proposed an
enhanced particle filtering algorithm for wind turbine drivetrain gearboxes RUL prediction using
vibration data, In their method, an adaptive neuro-fuzzy inference system was used to learn the
health state transition. Hu et al. [159] explored the Wiener process for the prediction of wind

31
turbine health status using the temperature characteristics of operational SCADA data. Nielsen and
Sørensen [160] proposed a Markov deterioration model to predict the deterioration and RUL of
wind turbine blades. In their model, a dynamic Bayesian network was used to obtain probabilities
of inspection outcomes and the maximum likelihood method was applied to estimate the transition
probabilities for a hidden Markov model. Saidi et al. [161] proposed a vibration-based prognostic
and health monitoring methodology for wind turbine high-speed shaft bearing using a spectral
kurtosis and SVR. Reviews about wind turbine condition monitoring can be found from [155, 162,
163].
Electric valves, power transformers, reactor coolant pumps, etc. are also widely used
components in many power systems. Though the RUL prediction of these individual components
is not trivial, in recent years, researchers showed their interest in prognosis of these components.
For example, Wang et al. [164] applied a convolution kernel combined with LSTM for feature
extraction. Then, LSTM is utilized for predicting RUL of electric valves. Later, Wang et al. [165]
improved the RUL prediction method by combining LSTM and convolutional auto-encoder
(CAE). They combined deeper features extracted by CAE and the original features to enrich the
dimension of features, and the case study showed an improved predictive capability. Aizpurua et
al. [166] focused on lifetime predictions of power transformers in NPPs. They proposed a Bayesian
Particle Filtering framework by integrating model-based experimental models, forecasting models
and uncertainty modelling concepts together for condition assessment of transformers. Nguyen et
al. [167] combined ensemble empirical mode decomposition and LSTM for the prognostics of
reactor coolant pumps of NPPs. They observed that multi-step-ahead predictions obtained by an
ensemble of separate prediction models are more accurate and less noisy than the predictions
obtained by a single model.

3.4 Application in Electrical and Electronic Components

Many electronic systems including consumer electronics, electric vehicles, airplanes, and
renewable energy devices use lithium-ion batteries as the main sources of energy storage. The
performance of lithium-ion batteries deteriorates over the service time in terms of capacity loss
and resistance increase. To ensure the safety and reliability of lithium-ion batteries, accurate
estimation of the health state and RUL prediction are essentials to track the actual performance of
batteries. RUL of batteries can be determined by the number of charges and discharge cycles to
reduce its capacity from the known current value to the threshold value. In the papers [1, 168],

32
they provided comprehensive reviews on data-driven methods for battery health diagnostics and
prognostics estimation. We found two public datasets available: NASA Ames Prognostic enter of
Excellence [169] and Center for Advanced Life Cycle Engineering (CALCE) of University of
Maryland [170]. NASA Ames Prognostic enter of Excellence dataset contains 4 batteries’ aging
processes which were tested under certain conditions. The batteries were run through different
charge, discharge, and impedance operational profiles at room temperature. The CALCE provides
multiple batteries dataset. For the RUL prediction, the battery capacity was tested as an indicator
of battery status. To measure the capacity, all batteries were fully charged under the constant-
current/constant-voltage mode. In the discharge period, the cells were applied to a specific load to
maintain at a constant current until the voltage was reduced to 2.7V. Then the discharge capacity
was recorded after each full charge-discharge process. The details of the experiments to generate
data can be found in [171].
A number of researches has been reported for RUL prediction based on these two public datasets.
Zhai and Ye [172] studied a Wiener process model with an adaptive drift for RUL prediction of
batteries. They concluded that the proposed model fixed the deficiency of conventional Wiener
process models that ignoring the variability of drifts. Shen et al. [173] proposed a Wiener-based
model with measurement errors, which were assumed to be a logistic distribution with zero means.
They adopted the Monte Carlo expectation-maximization method together with the Gibbs
sampling for parameter estimation. Wang et al. [174] proposed a mixed-effects model based on
the Wiener process to capture the two-phase degradation pattern. This model accommodated two
significant aspects: phase correlation and unit heterogeneity. Zhang et al. [175] presented a
stochastic modeling method and took the recovery phenomenon into consideration, which is a
common phenomenon for batteries that the system performance degrades with usage and recovers
in storage. Si [176] proposed a generic nonlinear stochastic modeling framework, they utilized a
time-dependent drift coefficient to characterize the nonlinearity and dynamics of the degradation
signals. Chen et al. [177] employed a hybrid method based on SVR and error compensation
methods for RUL prediction. They used genetic algorithm to optimize the hyper-parameters of
SVR to achieve better accuracy. Xue et al. [178] proposed an integrated algorithm that combines
unscented Kalman filter and SVR. Similar to Chen et al. [177], they used a genetic algorithm to
optimize parameters of SVR. Khumprom and Yodo [179] proposed a Deep Neural Network (DNN)
and compared with other machine learning algorithms, including SVM, k-NN, ANN, and Linear

33
Regression. The results showed that the DNN algorithm could be comparable and outweigh
conventional machine learning algorithms. Ren et al. [180] integrated autoencoder with DNN, in
which autoencoder was used for multi-dimensional feature extraction and DNN is trained for RUL
prediction. Jiao et al. [181] tried to combine both statistical method and deep learning model, they
proposed a Particle Filtering framework based on conditional variational autoencoder and a
reweighting strategy to predict the RUL. Since battery powered electric vehicles are starting to
play a significant role in today's automotive industry. The reliability and safety of batteries are
critical. Robust RUL prediction methods for batteries are desirable and the number of publications
are expected to increase rapidly.
Through the literature search, it is also observed that the prognosis methodologies are widely
being applied for another electronic component namely hard disk drive (HDD). The Hard disk
drive (HDD) is a complex system integrating mechanical, electricity and magnetism, which is the
most important and robust data storage device for major data storage services. Self-monitoring,
analysis and reporting technology (SMART) is a commercial health monitoring system, which can
detect and report various indicators of HDD reliability to facilitate the HDD prognosis. However,
the conventional SMART can only provide a basic evaluation, and the failure detection rate (FDR)
is 3-10% [182]. To improve the accuracy of proactive failure prediction, in recent years, statistical
and machine learning methods have been adopted to build prediction models based on the SMART
attributes. There are two public datasets used in literature: Baidu Inc. [183] and Backblaze Inc
(available at: https://www.backblaze.com/b2/hard-drive-test-data.html). The dataset of Baidu Inc.
was collected from a total of 23,395 drives, which had the same initial mode. The attributes of
those drives were sampled at every hour using the SMART and labeled as good or failed, with
only 433 drives in the failed class and the rest of 22,962 drives in the good class. The Backblaze
company also collect the dataset in the similar fashion. They gathered the SMART attributes on
daily basis. In 2013, the company made the dataset available for the research community and
provides update quarterly [184]. Using Baidu dataset, Xu et al. [185] introduced a RNN based
approach to assessing the health status of hard drives based on the gradually changing sequential
SMART attributes. Li et al. [186] proposed two hard drive failure prediction models based on
Decision Trees (DTs) and Gradient Boosted Regression Trees. Both prediction models showed
steady prediction performance, with high failure detection rates (80% to 96%) and low false alarm
rates (0.006% to 0.31%). Using Backblaze dataset, Lima et al. [187] evaluated the performance of

34
both LSTM and CNN architecture to predict the hard drive failure. The results of this study showed
that deep learning models could be the effective alternative for failure prediction.

3.5 Performance Analysis

This section first describes some performance evaluation metrics for RUL prediction. Prediction
of RUL is a vast research field where researchers develop and proposed a variety of methodologies
and algorithms. Hence, different researchers used different performance evaluation metrics.
Following our pursuit, we found the four most used evaluation metrics, i.e., Mean Absolutes Error
(MAE), Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE) and
Scoring Function (SF). These metrics are defined as,

1
ˆ

ƒ„9 †| ‡ − | (12)
‰,

1
ˆ

ŠƒT9 ‹ † ‡ − (13)
‰,

1 |‡ − |
ˆ

ƒ„59 † (14)
‰,
ˆ •‡‘ ?•‘
•† '- ? ,* − 1+ , ! ‡ − <0
TŒ ‰,
Ž † '- •‡‘,c
ˆ ?•‘
− 1+ , ! ‡ − ≥ 0
(15)
• ‰,
In the above equations, and ‡ represent true RUL and predicted RUL, respectively. is the
total number of units or systems. All of these metrics are negatively oriented scores, which means

prediction error to the model. The value of these two metrics could range between 0 and ∞. MAPE
a lower value indicates a better model. Both the MAE and RMSE report the average RUL

is a variant of MAE which is the absolute error normalized over the data. This metric is useful
when the errors need to be compared across data with different scales. The SF, as shown in
Equation (15), is asymmetric around the true time of failure. It is defined in a way that late
predictions are more heavily penalized than early predictions [137]. We only list a subset of metrics
in evaluating the performance of methodologies for different applications or datasets. For example,
C-MAPSS dataset, researchers who developed deep learning approaches for the dataset commonly
used the scoring and RMSE metrics to quantitively evaluate the proposed methodologies in this
field. This provides us an opportunity to compare different deep learning approaches directly. It is

35
worth mentioning that, the literature states a wide variety of performance criteria to evaluate
different methods. This is usual and expected, as the datasets used in this research field are
characterized by different parameters, scales, variations and operating conditions. Sometimes, a
dataset has multiples of sub-datasets and different experimental setups. Considering this limitation,
in the comparison table, we only report the common evaluation metrics for any particular dataset
having the same parameters, scales, and operating conditions. IEEE PHM2012 challenge dataset,
provided by the FEMTO-ST institution in France, is a popular dataset used for experimenting and
predicting the bearing remaining useful life. For this dataset, MAE and RMSE are often used to
evaluate the proposed methodologies. A comparison table for C-MAPSS and IEEE PHM 2012
challenge dataset is reported in Table IV and Table V, respectively. From Table IV, we can observe
that deep learning approaches applied to the C-MAPSS dataset are only started in 2016. Since
then, the number of publications has grown exponentially, which has been demonstrated in Figure
8. The best three performances in each column of Table IV are labeled in bold. Due to the
calculation biases from different authors, the RMSE and Scoring are not consistent. Generally, we
can conclude that capsule NN [188] and MSCNN [77] show superior performance and a good
generalization capability for all sub-datasets. Bayesian Deep learning [101] also demonstrates
satisfying results. It is worth mentioning that unlike other deep learning approaches only focusing
on point estimation of RUL, Bayesian Deep learning enhanced the model interpretability by
providing non only point estimation, but also uncertainty quantification, which is highly desirable
in practice. For IEEE PHM2012 challenge dataset, the performance metrics, dataset and data
prediction points vary, we only provide a rough comparison based on the work from [224]. As we
can see, DNN performs the best comparing other traditional machine learning approaches.

36
 Table IV Performance comparison for C-MAPSS dataset
F001 F002 F003 F004
Method year
Scoring RMSE Scoring RMSE Scoring RMSE Scoring RMSE
CNN [72] 2016 1287 18.45 13570 30.29 1596 19.82 7886 29.16
Multi-objective Deep
2017 640 17.96 10851 28.06 683 19.41 7210 29.45
belief NN[189]
Deep LSTM [129] 2017 338 16.14 4450 24.49 852 16.18 5550 27.17
BiLSTM [22] 2018 295 13.65 4130 23.18 317 13.74 5430 24.86
LSTM-FNN [190] 2018 481 14.89 7982 26.86 493 15.11 5200 27.11
Deep CNN [191] 2018 274 12.61 10400 22.36 284 12.64 12500 23.31
hybrid LSTM [87] 2019 262 14.72 6953 29.00 452 17.72 15069 33.43
Ensemble ResCNN [147] 2019 212 12.16 2087 20.85 180 12.01 3400 24.97
capsule NN [188] 2020 276 12.58 1229 16.30 283 11.71 2625 18.96
MSCNN [77] 2020 196 11.44 3747 19.35 241 11.67 4844 22.22
CNN+LSTM [192] 2020 231 12.56 3366 22.73 251 12.10 2840 22.66
Generative Adversarial
2020 174 10.71 2982 19.49 273 11.48 3874 19.71
Networks [193]
Bayesian Deep learning
2020 267 12.19 2007 18.49 409 12.07 2415 19.41
[194]

Table V Performance comparison for PHM 2012 challenge dataset

Dataset Method MAE (80% training) RMSE (80% training)
DNN 0.0333 0.0467
GBDT (Gradient boosting
0.0333 0.0600
FEMTO-ST institution at decision tree)
PRONOSTIA platform Gaussian 0.0633 0.0933
(PHM 2012 challenge
SVM 0.0600 0.0800
dataset)
BP Neural Network 0.1133 0.1533
Bayesian Ridge 0.1400 0.1833
**The metrics values are approximated from Ren et al. [195]

4. Challenges and Future Trends

4.1 Challenges and Future Trends for the Predictive Maintenance

With advancements in the industrial Internet of things and artificial intelligence, predictive
maintenance has become more and more efficient. The core point to apply predictive maintenance
strategy successfully is to model and predict failure patterns accurately. Based on our studies, it
can be seen that this area has been well studied using many methodologies ranging from statistical
approaches to machine learning based approaches. Recently, many researchers have focused on
machine learning approaches to explore the field of RUL prediction. Deep neural networks such
as CNNs, auto-encoder, RNN, LSTM are getting popularity for learning features from raw data.
Some researchers integrate the statistical and traditional machine learning approach to explore this

37
field [196, 197]. However, the task of RUL is still challenging due to the complex, uncertain,
nonlinear features and operational conditions. The main challenges can be summarized in several
aspects: (1) Data insufficiency and imbalanced classes: most data-driven models, especially
machine learning approaches, predict the RUL based on the extracted features learned from the
data. The predictive performance largely relies on the volume and quality of the datasets. However,
with the enhanced quality of engineering systems, the probabilities of failure are extremely low.
Data collection in the engineering field is time-consuming and costly. Moreover, the collected data
usually face the class imbalance issue. As stated in Section 3, most datasets are generated in a
simulated environment for case studies. The developed approaches may show poor performance
when adapting to real-world scenarios. (2) Poor generalization ability of developed models: many
developed models show poor performance when applying to different types or even the same type
of equipment under different operational environments. Although transfer learning has been
attempted to solve this problem, in practice remains a challenging issue to achieve satisfactory
performance in many scenarios. Moreover, the degradation mechanism may be affected by various
environmental factors (e.g., speed, loading) in the real world. How to accurately model these
variables are still challenging and critical since they will impact the degradation trend which in
turn will impact the prediction performance. (3) Late prediction caused by poor predictive
capability: if the predicted failure is later than the actual fault occurrence, it may not allow adequate
time for remedies, thus causing damage or losses. However, current research pay little attention to
time complexity analysis. (4) Noise associated with real-time/online prognostics for in-situ
components: the operating components may be affected by random disturbance from the
environment. The extensive research on how to incorporate data heterogeneity and uncertainty for
real-time prognostics are expected in the future. Moreover, online monitoring and prognostics
require fast and efficient algorithms, only a limited research report the computational time. (5)
Manual assignment of hyper-parameters estimation and tuning: many algorithms require to assign
and tune hyper-parameters, especially for deep learning models, the accuracy of the model is
largely dependent on the choice of hyper-parameters. However, the majority of the research only
assigns parameters manually. (6) Discrepancy of cross-domain prognosis: it is generally assumed
that the training and testing data are from the same distribution. In real world, distribution
discrepancy exists between training and testing data due to variation of machine working

38
condition, interference of environmental noise, etc. This may lead to significant prediction
performance deterioration.
Opportunities always come along with challenges. These opportunities include, but are not
limited to, (1) Dataset enrichment: currently, most of the data-driven algorithms only utilize
numeric data to predict the remaining life. It is well known that deep learning algorithms are
efficient at extracting features and learn from images. Thus, integration of images and numerical
data could bring an effective way to dig into the task of RUL prediction. In addition, the rise and
recent development of transfer learning and data augmentation techniques (e.g., Generative
adversarial networks) can also alleviate the demand for large datasets and boost the prognostic
accuracy. However, currently only a very limited number of research have been conducted for
RUL prediction. (2) Uncertainty quantification: a majority of machine learning based approaches
are only focused on predicting the mean values of RUL. Combining statistical approaches and then
quantifying uncertainty for more informed decision-making will be an interesting direction to
investigate. (3) Development of robust yet effective prediction algorithms for online monitoring:
with the emerging of machine learning techniques in predictive maintenance field, the proposed
models are becoming more and more complex, researchers may focus on how to control the
computation time through designing efficient models. (4) Model generalization: most prognostics
approaches are typically designed for a specific system or domain. It would be interesting to
develop general prognostic methods, which can be implemented in any systems, and to promote
in practical applications. (5) Real-time maintenance strategies development: basically, real-time
schemes for decision making based on RUL prediction that can deal with more complex and
dynamic scenarios are anticipated in society for maintenance and reliability. By utilizing advanced
cloud computing, AI and machine learning algorithms, we believe that more automated, intelligent
tools will be developed and deployed to improve performance in the industry soon. (6) Integration
of attention mechanism into the predictive model: attention mechanism is able to model the
dependencies between the target output and the input sequences and has now become an important
element in deep learning area, we’ll expect to see more and more publications that attention
mechanism to RUL prediction in the near future.

4.2 Guidelines for Model Selection of Predictive Maintenance Implementation

After a thorough review of data-driven algorithms with different applications, we believe the
selection of a concrete data-driven model is application-dependent. Also, different models have

39
been their own pros and cons. We would like to provide our perception on how to select and
construct an appropriate data-driven model for a specific application scenario. If the data size is
small and degradation data show similar degradation forms, GPMs perhaps are the most suitable
and simplest models to use. However, their inability to capture the temporal variability and the
uncertainty inherent in the progression of deterioration over time, which is common in practice,
limits their engineering applications. In other words, GPMs are applicable only when the
unexplained randomness is sufficiently small. To deal with the randomness caused by inherent
variability and environmental factors, SPMs are a natural choice. If degradation processes are
monotonic and evolving only in one direction, Gamma process and inverse Gaussian process are
appropriate to model this type of degradation data. Wiener process is suitable in modeling
deterioration which is not monotone. To relax the assumption of parametric forms, Gaussian
process models can be well adapted to model the complex data, where do not involve parameters.
Both GPMs and SPMs have well-established statistical properties, where a closed form of PDF of
RUL is usually available. If not in some cases, filtered based methods or some sampling methods
have to be used for finding an approximated RUL. One limitation of both GPMs and SPMs is that
they require a pre-defined failure threshold. However, this may not be available or accurate since
it requires knowledge from domain experts. Moreover, a fixed failure threshold may not be
sufficient to characterize the health status of all products due to their heterogeneous features. In
this case, covariate based models are efficient without needing failure threshold assumption. The
rapid development of sensing and computing technologies has enriched degradation data
significantly. This data-rich environment for degradation modeling and prognostics that could
potentially lead to an accurate inference about RUL of products. However, RUL prediction with
multi-sensor signals is a more challenging issue than the cases of a single degradation signal. One
way to deal with this issue is to combine multi-sensor signals into a composite health index or
mapping the correlation between signals and RUL values, then widely-used GPMs and stochastic
process models are still applicable. Another option is to use machine learning, which has gradually
become a mainstream for RUL prediction. The goal of conventional machine learning and deep
learning for RUL prediction is to learn the non-linear mapping between the sensor data and RUL
using different network architecture. Among those machine learning models, LSTMs have
attracted great attention and presented an outstanding ability in the application of RUL prediction,
as they have the capability to learn dependencies of sequential data. While machine models can

40
provide better performance for RUL prediction, they do not have a probabilistic orientation,
namely, uncertainty quantification, and therefore, no PDF of the RUL is available. Bayesian neural
networks have been used to cover the shortage. If there is a need for cross-domain prognosis,
transfer learning is a preferable way to provide a better performance. Attention mechanism can
also assist the learning model in yielding potential improvements in the learning tasks.

5. Conclusion
In the context of Industry 4.0, predictive maintenance is transforming the way of thinking
maintenance: from cost to business opportunity in the industry. Based on this rationale, predictive
maintenance is attracting considerable investment from industries and increasing attention from
research societies. Many predictive maintenance techniques have been developed up to now to
respond to the demand of high reliability of facilities and equipment but more studies are still
required to improve their predictive accuracies and efficiencies. This review provides a
comprehensive overview of the most recent data-driven prognostic techniques, which is the
indispensable process for predictive maintenance. Specifically, this paper reviews the
methodologies, best practices, current challenges, and future trends of machine prognostics. To
make accurate prognostics, choosing a proper modeling technique is essential. We provide a
detailed summary of statistical based models and machine learning based models. Then, their
applications based on these models are demonstrated in detail, which provide a good reference for
selecting an appropriate model for a specific application scenario. Moreover, we investigate and
pinpoint some challenges and promising directions and opportunities of prognostics for future
studies. Lastly, we provide some constructive resolutions to mitigate the predictive maintenance
challenges (e.g., data insufficiency and imbalanced classes; poor generalization ability of
developed models; late prediction caused by poor predictive capability; noise associated with real-
time/online prognostics for in-situ components; manual assignment of hyper-parameters
estimation and tuning; and discrepancy of cross-domain prognosis) and guidance for the user to
choose appropriate models to support predictive maintenance implementation. This research effort
would lead to develop machine learning based predictive maintenance system that enables to
sustain effective and accurate preventive maintenance. In summary, this review provides an
indication of how to study predictive maintenance problems from data-driven machine prognostics
perspective and pave a path for effective further investigation. It can be foreseen that more and
more advanced predictive models will be developed in the near future, which will boost predictive

41
maintenance, improve reliability, enhance productivity and achieve intelligent decision-making in
the industry.

Acknowledgments
This work was partially supported by the National Science Foundation (ECR-PEER-1935454),
(ERC-ASPIRE-1941524) and Department of Education (Award # P120A180101). The authors
wish to express sincere gratitude for their financial support.

References
[1] H. Meng, Y.-F. Li, A review on prognostics and health management (PHM) methods of lithium-ion
batteries, Renewable and Sustainable Energy Reviews, 116 (2019) 109405.
[2] P.G. Ramesh, S.J. Dutta, S.S. Neog, P. Baishya, I. Bezbaruah, Implementation of Predictive
Maintenance Systems in Remotely Located Process Plants under Industry 4.0 Scenario, Advances in
RAMS Engineering, Springer, 2020, pp. 293-326.
[3] N. Sakib, T. Wuest, Challenges and Opportunities of Condition-based Predictive Maintenance: A
Review, Procedia CIRP, 78 (2018) 267-272.
[4] R. Ahmad, S. Kamaruddin, An overview of time-based and condition-based maintenance in industrial
application, Computers & industrial engineering, 63 (2012) 135-149.
[5] A. Jezzini, M. Ayache, L. Elkhansa, B. Makki, M. Zein, Effects of predictive maintenance(PdM),
Proactive maintenace(PoM) & Preventive maintenance(PM) on minimizing the faults in medical
instruments, 2013 2nd International Conference on Advances in Biomedical Engineering, 2013, pp.
53-56.
[6] A.K.S. Jardine, D. Lin, D. Banjevic, A review on machinery diagnostics and prognostics implementing
condition-based maintenance, Mechanical Systems and Signal Processing, 20 (2006) 1483-1510.
[7] K.L. Tsui, N. Chen, Q. Zhou, Y. Hai, W. Wang, Prognostics and health management: A review on data
driven approaches, Mathematical Problems in Engineering, 2015 (2015).
[8] Y. Lei, N. Li, L. Guo, N. Li, T. Yan, J. Lin, Machinery health prognostics: A systematic review from
data acquisition to RUL prediction, Mechanical Systems and Signal Processing, 104 (2018) 799-834.
[9] M.S. Kan, A.C.C. Tan, J. Mathew, A review on prognostic techniques for non-stationary and non-linear
rotating systems, Mechanical Systems and Signal Processing, 62 (2015) 1-20.
[10] M.G. Pecht, A prognostics and health management roadmap for information and electronics-rich
systems, IEICE ESS Fundamentals Review, 3 (2010) 4_25-24_32.
[11] Y. Wen, J. Wu, Q. Zhou, T.-L. Tseng, Multiple-Change-Point Modeling and Exact Bayesian Inference
of Degradation Signal for Prognostic Improvement, IEEE Transactions on Automation Science and
Engineering, (2018) 1-16.
[12] X.-S. Si, W. Wang, C.-H. Hu, D.-H. Zhou, Remaining useful life estimation – A review on the
statistical data driven approaches, European Journal of Operational Research, 213 (2011) 1-14.
[13] Z.S. Ye, M. Xie, Stochastic modelling and analysis of degradation for highly reliable products, Applied
Stochastic Models in Business and Industry, 31 (2015) 16-32.
[14] Z. Zhang, X. Si, C. Hu, X. Kong, Degradation modeling–based remaining useful life estimation: A
review on approaches for systems with heterogeneity, Proceedings of the Institution of Mechanical
Engineers, Part O: Journal of Risk and Reliability, 229 (2015) 343-355.
[15] D. Wang, K.-L. Tsui, Q. Miao, Prognostics and health management: A review of vibration based
bearing and gear health indicators, Ieee Access, 6 (2017) 665-676.
[16] S. Khan, T. Yairi, A review on the application of deep learning in system health management,
Mechanical Systems and Signal Processing, 107 (2018) 241-265.

42
[17] L. Zhang, J. Lin, B. Liu, Z. Zhang, X. Yan, M. Wei, A review on deep learning applications in
prognostics and health management, IEEE Access, 7 (2019) 162415-162438.
[18] R. Zhao, R. Yan, Z. Chen, K. Mao, P. Wang, R.X. Gao, Deep learning and its applications to machine
health monitoring, Mechanical Systems and Signal Processing, 115 (2019) 213-237.
[19] M. Kordestani, M. Saif, M.E. Orchard, R. Razavi-Far, K. Khorasani, Failure Prognosis and
Applications—A Survey of Recent Literature, IEEE transactions on reliability, (2019).
[20] J. Guo, Z. Li, M. Li, A Review on Prognostics Methods for Engineering Systems, IEEE Transactions
on Reliability, (2019) 1-20.
[21] M. Baur, P. Albertelli, M. Monno, A review of prognostics and health management of machine tools,
The International Journal of Advanced Manufacturing Technology, 107 (2020) 2843-2863.
[22] J. Lee, F. Wu, W. Zhao, M. Ghaffari, L. Liao, D. Siegel, Prognostics and health management design
for rotary machinery systems—Reviews, methodology and applications, Mechanical systems and
signal processing, 42 (2014) 314-334.
[23] C.J. Lu, W.O. Meeker, Using degradation measures to estimate a time-to-failure distribution,
Technometrics, 35 (1993) 161-174.
[24] N. Gebraeel, A. Elwany, J. Pan, Residual life predictions in the absence of prior degradation
knowledge, IEEE Transactions on Reliability, 58 (2009) 106-117.
[25] H. Kim, J.T. Kim, G. Heo, Prognostics for integrity of steam generator tubes using the general path
model, Nuclear Engineering and Technology, 50 (2018) 88-96.
[26] N. Gebraeel, Sensory-updated residual life distributions for components with exponential degradation
patterns, IEEE Transactions on Automation Science and Engineering, 3 (2006) 382-393.
[27] Y. Wen, J. Wu, Y. Yuan, Multiple-phase modeling of degradation signal for condition monitoring and
remaining useful life prediction, IEEE Transactions on Reliability, 66 (2017) 924-938.
[28] R. Zhou, N. Serban, N. Gebraeel, Degradation-based residual life prediction under different
environments, The Annals of Applied Statistics, (2014) 1671-1689.
[29] N. Chen, K.L. Tsui, Condition monitoring and remaining useful life prediction using degradation
signals: Revisited, IIE Transactions, 45 (2013) 939-952.
[30] G.A. Whitmore, Estimating degradation by a Wiener diffusion process subject to measurement error,
Lifetime data analysis, 1 (1995) 307-319.
[31] X. Wang, Wiener processes with random effects for degradation data, Journal of Multivariate Analysis,
101 (2010) 340-351.
[32] Y. Wen, J. Wu, D. Das, T.-L.B. Tseng, Degradation modeling and RUL prediction using Wiener
process subject to multiple change points and unit heterogeneity, Reliability Engineering & System
Safety, 176 (2018) 113-124.
[33] X.-S. Si, W. Wang, C.-H. Hu, D.-H. Zhou, M.G. Pecht, Remaining useful life estimation based on a
nonlinear diffusion degradation process, IEEE Transactions on Reliability, 61 (2012) 50-67.
[34] Z.-S. Ye, Y. Wang, K.-L. Tsui, M. Pecht, Degradation data analysis using Wiener processes with
measurement errors, IEEE Transactions on Reliability, 62 (2013) 772-780.
[35] X.-S. Si, W. Wang, C.-H. Hu, M.-Y. Chen, D.-H. Zhou, A Wiener-process-based degradation model
with a recursive filter algorithm for remaining useful life estimation, Mechanical Systems and Signal
Processing, 35 (2013) 219-237.
[36] C.-Y. Peng, S.-T. Tseng, Mis-specification analysis of linear degradation models, IEEE Transactions
on Reliability, 58 (2009) 444-455.
[37] H. Wang, X. Ma, Y. Zhao, An improved Wiener process model with adaptive drift and diffusion for
online remaining useful life prediction, Mechanical Systems and Signal Processing, 127 (2019) 370-
387.
[38] Z.-S. Ye, Y. Shen, M. Xie, Degradation-based burn-in with preventive maintenance, European journal
of operational research, 221 (2012) 360-367.
[39] Z. Zhang, X. Si, C. Hu, Y. Lei, Degradation data analysis and remaining useful life estimation: A
review on Wiener-process-based methods, European Journal of Operational Research, 271 (2018) 775-
796.

43
[40] Q. Dong, L. Cui, A study on stochastic degradation process models under different types of failure
thresholds, Reliability Engineering & System Safety, 181 (2019) 202-212.
[41] J.M. van Noortwijk, A survey of the application of gamma processes in maintenance, Reliability
Engineering & System Safety, 94 (2009) 2-21.
[42] R.R. Richardson, M.A. Osborne, D.A. Howey, Gaussian process regression for forecasting battery
state of health, Journal of Power Sources, 357 (2017) 209-219.
[43] P. Boškoski, M. Gašperin, D. Petelin, Đ. Juričić, Bearing fault prognostics using Rényi entropy based
features and Gaussian process models, Mechanical Systems and Signal Processing, 52 (2015) 327-337.
[44] S.A. Aye, P.S. Heyns, An integrated Gaussian process regression for prediction of remaining useful
life of slow speed bearings based on acoustic emission, Mechanical Systems and Signal Processing, 84
(2017) 485-498.
[45] D. Yang, X. Zhang, R. Pan, Y. Wang, Z. Chen, A novel Gaussian process regression model for state-
of-health estimation of lithium-ion battery using charging curve, Journal of Power Sources, 384 (2018)
387-395.
[46] Z.-S. Ye, N. Chen, The inverse Gaussian process as a degradation model, Technometrics, 56 (2014)
302-311.
[47] F. Cartella, J. Lemeire, L. Dimiccoli, H. Sahli, Hidden semi-Markov models for predictive
maintenance, Mathematical Problems in Engineering, 2015 (2015).
[48] C.R. David, Regression models and life tables (with discussion), Journal of the Royal Statistical
Society, 34 (1972) 187-220.
[49] Q. Zhou, J. Son, S. Zhou, X. Mao, M. Salman, Remaining useful life prediction of individual units
subject to hard failure, IIE Transactions, 46 (2014) 1017-1030.
[50] J. Man, Q. Zhou, Prediction of hard failures with stochastic degradation signals using Wiener process
and proportional hazards model, Computers & Industrial Engineering, 125 (2018) 480-489.
[51] J. Hu, Q. Sun, Z. Ye, Q. Zhou, Joint Modeling of Degradation and Lifetime Data for RUL Prediction
of Deteriorating Products, IEEE Transactions on Industrial Informatics, (2020) 1-1.
[52] X. Yue, R.A. Kontar, Joint Models for Event Prediction from Time Series and Survival Data,
Technometrics, (2020) 1-26.
[53] J.L. Katzman, U. Shaham, A. Cloninger, J. Bates, T. Jiang, Y. Kluger, DeepSurv: personalized
treatment recommender system using a Cox proportional hazards deep neural network, BMC medical
research methodology, 18 (2018) 24.
[54] H. Kvamme, Ø. Borgan, I. Scheel, Time-to-event prediction with neural networks and Cox regression,
arXiv preprint arXiv:1907.00825, (2019).
[55] P.J.G. Nieto, E. García-Gonzalo, F.S. Lasheras, F.J. de Cos Juez, Hybrid PSO–SVM-based method
for forecasting of the remaining useful life for aircraft engines and evaluation of its reliability,
Reliability Engineering & System Safety, 138 (2015) 219-231.
[56] T. Qin, S. Zeng, J. Guo, Robust prognostics for state of health estimation of lithium-ion batteries based
on an improved PSO–SVR model, Microelectronics Reliability, 55 (2015) 1280-1284.
[57] T. Benkedjouh, K. Medjaher, N. Zerhouni, S. Rechak, Remaining useful life estimation based on
nonlinear feature reduction and support vector regression, Engineering Applications of Artificial
Intelligence, 26 (2013) 1751-1760.
[58] Y. Hu, C. Hu, X. Kong, Z. Zhou, Real-time lifetime prediction method based on wavelet support vector
regression and fuzzy c-means clustering, Acta Automatica Sinica, 38 (2012) 331-340.
[59] C. Shen, D. Wang, F. Kong, W.T. Peter, Fault diagnosis of rotating machinery based on the statistical
parameters of wavelet packet paving and a generic support vector regressive classifier, Measurement,
46 (2013) 1551-1564.
[60] J. Liu, R. Seraoui, V. Vitelli, E. Zio, Nuclear power plant components condition monitoring by
probabilistic support vector machine, Annals of Nuclear Energy, 56 (2013) 23-33.
[61] H.-Z. Huang, H.-K. Wang, Y.-F. Li, L. Zhang, Z. Liu, Support vector machine based estimation of
remaining useful life: current research status and future trends, Journal of Mechanical Science and
Technology, 29 (2015) 151-163.

44
[62] P. Kundu, A.K. Darpe, M.S. Kulkarni, An ensemble decision tree methodology for remaining useful
life prediction of spur gears under natural pitting progression, Structural Health Monitoring, 19 (2020)
854-872.
[63] L. Wang, D. Zhou, H. Zhang, W. Zhang, J. Chen, Application of relative entropy and gradient boosting
decision tree to fault prognosis in electronic circuits, Symmetry, 10 (2018) 495.
[64] M. Ferguson, R. Ak, Y.-T.T. Lee, K.H. Law, Automatic localization of casting defects with
convolutional neural networks, 2017 IEEE international conference on big data (big data), IEEE, 2017,
pp. 1726-1735.
[65] M.K. Ferguson, A. Ronay, Y.-T.T. Lee, K.H. Law, Detection and segmentation of manufacturing
defects with convolutional neural networks and transfer learning, Smart and sustainable manufacturing
systems, 2 (2018).
[66] M.F. Rahman, J. Wu, T.L.B. Tseng, Automatic morphological extraction of fibers from SEM images
for quality control of short fiber-reinforced composites manufacturing, CIRP Journal of Manufacturing
Science and Technology, 33 (2021) 176-187.
[67] W. Hou, Y. Wei, J. Guo, Y. Jin, Automatic detection of welding defects using deep neural network,
Journal of Physics: Conference Series, IOP Publishing, 2017, pp. 012006.
[68] Z. Huang, Z. Pan, B. Lei, Transfer learning with deep convolutional neural network for SAR target
classification with limited labeled data, Remote Sensing, 9 (2017) 907.
[69] M.F. Rahman, T.-L.B. Tseng, M. Pokojovy, W. Qian, B. Totada, H. Xu, An automatic approach to
lung region segmentation in chest x-ray images using adapted U-Net architecture, Medical Imaging
2021: Physics of Medical Imaging, International Society for Optics and Photonics, 2021, pp. 115953I.
[70] M.F. Rahman, Y. Wen, H. Xu, T.-L.B. Tseng, S. Akundi, Data mining in telemedicine, Advances in
Telemedicine for Health Monitoring, (2020) 103.
[71] L. Wen, L. Gao, X. Li, B. Zeng, Convolutional neural network with automatic learning rate scheduler
for fault classification, IEEE Transactions on Instrumentation and Measurement, 70 (2021) 1-12.
[72] G.S. Babu, P. Zhao, X.-L. Li, Deep convolutional neural network based regression approach for
estimation of remaining useful life, International conference on database systems for advanced
applications, Springer, 2016, pp. 214-228.
[73] L. Ren, Y. Sun, H. Wang, L. Zhang, Prediction of bearing remaining useful life with deep convolution
neural network, IEEE Access, 6 (2018) 13041-13049.
[74] B. Yang, R. Liu, E. Zio, Remaining useful life prediction based on a double-convolutional neural
network architecture, IEEE Transactions on Industrial Electronics, 66 (2019) 9521-9530.
[75] S. Kiranyaz, A. Gastli, L. Ben-Brahim, N. Al-Emadi, M. Gabbouj, Real-time fault detection and
identification for MMC using 1-D convolutional neural networks, IEEE Transactions on Industrial
Electronics, 66 (2018) 8760-8771.
[76] J. Zhu, N. Chen, W. Peng, Estimation of bearing remaining useful life based on multiscale
convolutional neural network, IEEE Transactions on Industrial Electronics, 66 (2018) 3208-3216.
[77] H. Li, W. Zhao, Y. Zhang, E. Zio, Remaining useful life prediction using multi-scale deep
convolutional neural network, Applied Soft Computing, 89 (2020) 106113.
[78] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural computation, 9 (1997) 1735-1780.
[79] K. Cho, B. Van Merriënboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, Y. Bengio,
Learning phrase representations using RNN encoder-decoder for statistical machine translation, arXiv
preprint arXiv:1406.1078, (2014).
[80] Y. Song, L. Li, Y. Peng, D. Liu, Lithium-Ion Battery Remaining Useful Life Prediction Based on
GRU-RNN, 2018 12th International Conference on Reliability, Maintainability, and Safety (ICRMS),
IEEE, 2018, pp. 317-322.
[81] J. Chen, H. Jing, Y. Chang, Q. Liu, Gated recurrent unit based recurrent neural network for remaining
useful life prediction of nonlinear deterioration process, Reliability Engineering & System Safety, 185
(2019) 372-382.
[82] J. Wang, J. Yan, C. Li, R.X. Gao, R. Zhao, Deep heterogeneous GRU model for predictive analytics
in smart manufacturing: Application to tool wear prediction, Computers in Industry, 111 (2019) 1-14.

45
[83] F.O. Heimes, Recurrent neural networks for remaining useful life estimation, IEEE, pp. 1-6.
[84] J. Liu, A. Saxena, K. Goebel, B. Saha, W. Wang, An adaptive recurrent neural network for remaining
useful life prediction of lithium-ion batteries, National Aeronautics And Space Administration Moffett
Field CA Ames Research …, 2010.
[85] L. Guo, N. Li, F. Jia, Y. Lei, J. Lin, A recurrent neural network based health indicator for remaining
useful life prediction of bearings, Neurocomputing, 240 (2017) 98-109.
[86] Y. Zhang, R. Xiong, H. He, M.G. Pecht, Long short-term memory recurrent neural network for
remaining useful life prediction of lithium-ion batteries, IEEE Transactions on Vehicular Technology,
67 (2018) 5695-5705.
[87] S. Zhao, Y. Zhang, S. Wang, B. Zhou, C. Cheng, A recurrent neural network approach for remaining
useful life prediction utilizing a novel trend features construction method, Measurement, 146 (2019)
279-288.
[88] S. Xiang, Y. Qin, C. Zhu, Y. Wang, H. Chen, Long short-term memory neural network with weight
amplification and its application into gear remaining useful life prediction, Engineering Applications
of Artificial Intelligence, 91 (2020) 103587.
[89] A. Zhang, H. Wang, S. Li, Y. Cui, Z. Liu, G. Yang, J. Hu, Transfer learning with deep recurrent neural
networks for remaining useful life estimation, Applied Sciences, 8 (2018) 2416.
[90] J. Wang, G. Wen, S. Yang, Y. Liu, Remaining useful life estimation in prognostics using deep
bidirectional lstm neural network, 2018 Prognostics and System Health Management Conference
(PHM-Chongqing), IEEE, 2018, pp. 1037-1042.
[91] A. Elsheikh, S. Yacout, M.-S. Ouali, Bidirectional handshaking LSTM for remaining useful life
prediction, Neurocomputing, 323 (2019) 148-156.
[92] S. Xiang, Y. Qin, J. Luo, H. Pu, B. Tang, Multicellular LSTM-based deep learning model for aero-
engine remaining useful life prediction, Reliability Engineering & System Safety, 216 (2021) 107927.
[93] Q. An, Z. Tao, X. Xu, M. El Mansori, M. Chen, A data-driven model for milling tool remaining useful
life prediction with convolutional and stacked LSTM network, Measurement, 154 (2020) 107461.
[94] Y. Wang, H. Yao, S. Zhao, Auto-encoder based dimensionality reduction, Neurocomputing, 184
(2016) 232-242.
[95] J. Ma, H. Su, W.-l. Zhao, B. Liu, Predicting the remaining useful life of an aircraft engine using a
stacked sparse autoencoder with multilayer self-learning, Complexity, 2018 (2018).
[96] Y. Song, G. Shi, L. Chen, X. Huang, T. Xia, Remaining useful life prediction of turbofan engine using
hybrid model based on autoencoder and bidirectional long short-term memory, Journal of Shanghai
Jiaotong University (Science), 23 (2018) 85-94.
[97] C. Su, L. Li, Z. Wen, Remaining useful life prediction via a variational autoencoder and a time‐
window‐based sequence neural network, Quality and Reliability Engineering International, (2020).
[98] W. Mao, J. He, M.J. Zuo, Predicting Remaining Useful Life of Rolling Bearings Based on Deep
Feature Representation and Transfer Learning, IEEE Transactions on Instrumentation and
Measurement, 69 (2020) 1594-1608.
[99] M. Xia, T. Li, T. Shu, J. Wan, C.W. De Silva, Z. Wang, A two-stage approach for the remaining useful
life prediction of bearings using deep neural networks, IEEE Transactions on Industrial Informatics, 15
(2018) 3703-3711.
[100] W. Peng, Z.-S. Ye, N. Chen, Bayesian Deep-Learning-Based Health Prognostics Toward Prognostics
Uncertainty, IEEE Transactions on Industrial Electronics, 67 (2019) 2283-2293.
[101] M. Kim, K. Liu, A Bayesian Deep Learning Framework for Interval Estimation of Remaining Useful
Life in Complex Systems by Incorporating General Degradation Characteristics, IISE Transactions,
(2020) 1-23.
[102] G. Li, L. Yang, C.-G. Lee, X. Wang, M. Rong, A Bayesian deep learning RUL framework integrating
epistemic and aleatoric uncertainties, IEEE Transactions on Industrial Electronics, (2020).
[103] S.J. Pan, Q. Yang, A survey on transfer learning, IEEE Transactions on knowledge and data
engineering, 22 (2009) 1345-1359.

46
[104] C. Sun, M. Ma, Z. Zhao, S. Tian, R. Yan, X. Chen, Deep transfer learning based on sparse autoencoder
for remaining useful life prediction of tool in manufacturing, IEEE Transactions on Industrial
Informatics, 15 (2018) 2416-2425.
[105] Y. Fan, S. Nowaczyk, T. Rögnvaldsson, Transfer learning for remaining useful life prediction based
on consensus self-organizing models, Reliability Engineering & System Safety, 203 (2020) 107098.
[106] H. Zhang, Q. Zhang, S. Shao, T. Niu, X. Yang, H. Ding, Sequential Network with Residual Neural
Network for Rotatory Machine Remaining Useful Life Prediction Using Deep Transfer Learning,
Shock and Vibration, 2020 (2020).
[107] Z. Zhang, Y. Wang, K. Wang, Fault diagnosis and prognosis using wavelet packet decomposition,
Fourier transform and artificial neural network, Journal of Intelligent Manufacturing, 24 (2013) 1213-
1227.
[108] Y. Wang, G. Xu, L. Liang, K. Jiang, Detection of weak transient signals based on wavelet packet
transform and manifold learning for rolling element bearing fault diagnosis, Mechanical Systems and
Signal Processing, 54 (2015) 259-276.
[109] G. Bin, J. Gao, X. Li, B. Dhillon, Early fault diagnosis of rotating machinery based on wavelet
packets—Empirical mode decomposition feature extraction and neural network, Mechanical Systems
and Signal Processing, 27 (2012) 696-711.
[110] J.-X. Zhang, D.-B. Du, X.-S. Si, C.-H. Hu, H.-W. Zhang, Joint optimization of preventive
maintenance and inventory management for standby systems with hybrid-deteriorating spare parts,
Reliability Engineering & System Safety, 214 (2021) 107686.
[111] J. Cai, Y. Yin, L. Zhang, X. Chen, Joint optimization of preventive maintenance and spare parts
inventory with appointment policy, Mathematical Problems in Engineering, 2017 (2017).
[112] Y. Jiang, M. Chen, D. Zhou, Joint optimization of preventive maintenance and inventory policies for
multi-unit systems subject to deteriorating spare part inventory, Journal of manufacturing systems, 35
(2015) 191-205.
[113] J.-X. Zhang, X.-S. Si, D.-B. Du, C.-H. Hu, C. Hu, A novel iterative approach of lifetime estimation
for standby systems with deteriorating spare parts, Reliability Engineering & System Safety, 201 (2020)
106960.
[114] H. Jia, Y. Ding, R. Peng, Y. Song, Reliability evaluation for demand-based warm standby systems
considering degradation process, IEEE Transactions on Reliability, 66 (2017) 795-805.
[115] P. Nectoux, R. Gouriveau, K. Medjaher, E. Ramasso, B. Chebel-Morello, N. Zerhouni, C. Varnier,
PRONOSTIA: An experimental platform for bearings accelerated degradation tests.
[116] R.M. Mortier, S.T. Orszulik, M.F. Fox, Chemistry and technology of lubricants, Springer, 2010.
[117] Y. Du, T. Wu, J. Cheng, R. Gong, Lubricating oil deterioration on a four-ball test rig via on-line
monitoring, Proceedings of Malaysian international tribology conference, 2015, pp. 185-186.
[118] Y. Du, T. Wu, S. Zhou, V. Makis, Remaining useful life prediction of lubricating oil with dynamic
principal component analysis and proportional hazards model, Proceedings of the Institution of
Mechanical Engineers, Part J: Journal of Engineering Tribology, 234 (2020) 964-971.
[119] Y. Lei, N. Li, S. Gontarz, J. Lin, S. Radkowski, J. Dybala, A model-based method for remaining
useful life prediction of machinery, IEEE Transactions on Reliability, 65 (2016) 1314-1326.
[120] S. Hong, Z. Zhou, C. Lu, B. Wang, T. Zhao, Bearing remaining life prediction using Gaussian process
regression with composite kernel functions, Journal of Vibroengineering, 17 (2015) 695-704.
[121] Z. Liu, M.J. Zuo, Y. Qin, Remaining useful life prediction of rolling element bearings based on health
state assessment, Proceedings of the Institution of Mechanical Engineers, Part C: Journal of Mechanical
Engineering Science, 230 (2016) 314-330.
[122] H. Wang, Y. Zhao, X. Ma, Remaining Useful Life Prediction Using a Novel Two-Stage Wiener
Process With Stage Correlation, IEEE Access, 6 (2018) 65227-65238.
[123] W. Ahmad, S.A. Khan, M.M. Islam, J.-M. Kim, A reliable technique for remaining useful life
estimation of rolling element bearings using dynamic regression models, Reliability Engineering &
System Safety, 184 (2019) 67-76.

47
[124] P. Kundu, S. Chopra, B.K. Lad, Multiple failure behaviors identification and remaining useful life
prediction of ball bearings, Journal of Intelligent Manufacturing, 30 (2019) 1795-1807.
[125] M. Zhao, B. Tang, Q. Tan, Bearing remaining useful life estimation based on time–frequency
representation and supervised dimensionality reduction, Measurement, 86 (2016) 41-55.
[126] L. Xiao, X. Chen, X. Zhang, M. Liu, A novel approach for bearing remaining useful life estimation
under neither failure nor suspension histories condition, Journal of Intelligent Manufacturing, 28 (2017)
1893-1914.
[127] A.R. Bastami, A. Aasi, H.A. Arghand, Estimation of remaining useful life of rolling element bearings
using wavelet packet decomposition and artificial neural network, Iranian Journal of Science and
Technology, Transactions of Electrical Engineering, 43 (2019) 233-245.
[128] W. Mao, J. He, J. Tang, Y. Li, Predicting remaining useful life of rolling bearings based on deep
feature representation and long short-term memory neural network, Advances in Mechanical
Engineering, 10 (2018) 1687814018817184.
[129] G. Tang, Y. Zhou, H. Wang, G. Li, Prediction of bearing performance degradation with bottleneck
feature based on LSTM network, 2018 IEEE International Instrumentation and Measurement
Technology Conference (I2MTC), IEEE, 2018, pp. 1-6.
[130] L. Ren, X. Cheng, X. Wang, J. Cui, L. Zhang, Multi-scale dense gate recurrent unit networks for
bearing remaining useful life prediction, Future Generation Computer Systems, 94 (2019) 601-609.
[131] X. Li, W. Zhang, Q. Ding, Deep learning-based remaining useful life estimation of bearings using
multi-scale feature extraction, Reliability Engineering & System Safety, 182 (2019) 208-218.
[132] Q. Wu, C. Zhang, Cascade Fusion Convolutional Long-Short Time Memory Network for Remaining
Useful Life Prediction of Rolling Bearing, IEEE Access, 8 (2020) 32957-32965.
[133] C.-C. Lo, C.-H. Lee, W.-C. Huang, Prognosis of bearing and gear wears using convolutional neural
network with hybrid loss function, Sensors, 20 (2020) 3539.
[134] M. Tanwar, N. Raghavan, Lubricating oil degradation modeling and prognostics using the Wiener
process, 2019 International Conference on Sensing, Diagnostics, Prognostics, and Control (SDPC),
IEEE, 2019, pp. 601-605.
[135] M. Tanwar, N. Raghavan, Lubricating Oil Remaining Useful Life Prediction Using Multi-Output
Gaussian Process Regression, IEEE Access, 8 (2020) 128897-128907.
[136] V.T. Le, C.P. Lim, S. Mohamed, S. Nahavandi, L. Yen, G.E. Gallasch, S. Baker, D. Ludovici, N.
Draper, V. Wickramanayake, Condition monitoring of engine lubrication oil of military vehicles: A
machine learning approach, 17th Australian International Aerospace Congress: AIAC 2017, Engineers
Australia, Royal Aeronautical Society, 2017, pp. 718.
[137] A. Saxena, K. Goebel, D. Simon, N. Eklund, Damage propagation modeling for aircraft engine run-
to-failure simulation, 2008 international conference on prognostics and health management, IEEE,
2008, pp. 1-9.
[138] A. Chehade, S. Bonk, K. Liu, Sensory-based failure threshold estimation for remaining useful life
prediction, IEEE Transactions on Reliability, 66 (2017) 939-949.
[139] M. Kim, C. Song, K. Liu, A Generic Health Index Approach for Multisensor Degradation Modeling
and Sensor Selection, IEEE Transactions on Automation Science and Engineering, 16 (2019) 1426-
1437.
[140] C. Song, K. Liu, Statistical degradation modeling and prognostics of multiple sensor signals via data
fusion: A composite health index approach, IISE Transactions, 50 (2018) 853-867.
[141] A. Chehade, C. Song, K. Liu, A. Saxena, X. Zhang, A data-level fusion approach for degradation
modeling and prognostic analysis under multiple failure modes, Journal of Quality Technology, 50
(2018) 150-165.
[142] N. Li, Y. Lei, T. Yan, N. Li, T. Han, A Wiener-process-model-based method for remaining useful
life prediction considering unit-to-unit variability, IEEE Transactions on Industrial Electronics, 66
(2018) 2092-2101.
[143] K. Le Son, M. Fouladirad, A. Barros, Remaining useful lifetime estimation and noisy gamma
deterioration process, Reliability Engineering & System Safety, 149 (2016) 76-87.

48
[144] C. Ordóñez, F.S. Lasheras, J. Roca-Pardiñas, F.J. de Cos Juez, A hybrid ARIMA–SVM model for
the study of the remaining useful life of aircraft engines, Journal of Computational and Applied
Mathematics, 346 (2019) 184-191.
[145] A. Al-Dulaimi, S. Zabihi, A. Asif, A. Mohammadi, A multimodal and hybrid deep neural network
model for remaining useful life estimation, Computers in Industry, 108 (2019) 186-196.
[146] S. Zheng, K. Ristovski, A. Farahat, C. Gupta, Long short-term memory network for remaining useful
life estimation, 2017 IEEE international conference on prognostics and health management (ICPHM),
IEEE, 2017, pp. 88-95.
[147] L. Wen, Y. Dong, L. Gao, A new ensemble residual convolutional neural network for remaining
useful life estimation, Math. Biosci. Eng, 16 (2019) 862-880.
[148] X. Chen, H. Wang, J. Huang, H. Ren, APU degradation prediction based on EEMD and Gaussian
process regression, IEEE, pp. 98-104.
[149] X. Liu, L. Liu, L. Wang, Q. Guo, X. Peng, Performance sensing data prediction for an aircraft
auxiliary power unit using the optimized extreme learning machine, Sensors, 19 (2019) 3935.
[150] F. Wang, J. Sun, X. Liu, C. Liu, Aircraft auxiliary power unit performance assessment and remaining
useful life evaluation for predictive maintenance, Proceedings of the Institution of Mechanical
Engineers, Part A: Journal of Power and Energy, 234 (2020) 804-816.
[151] Y. Zhang, Y. Peng, P. Wang, L. Wang, S. Wang, H. Liao, Aircraft APU failure rate prediction based
on improved Weibull-based GRP, IEEE, pp. 1-6.
[152] Y. Zhang, D. Liu, J. Yu, Y. Peng, X. Peng, EMA remaining useful life prediction with weighted
bagging GPR algorithm, Microelectronics Reliability, 75 (2017) 253-263.
[153] Y. Zhang, L. Liu, Y. Peng, D. Liu, An Electro-Mechanical Actuator motor voltage estimation method
with a feature-aided Kalman Filter, Sensors, 18 (2018) 4190.
[154] R. Guo, Z. Liu, J. Wang, Remaining useful life prediction for the electro-hydraulic actuator based on
improved relevance vector machine, Proceedings of the Institution of Mechanical Engineers, Part I:
Journal of Systems and Control Engineering, 234 (2020) 501-511.
[155] A. Stetco, F. Dinmohammadi, X. Zhao, V. Robu, D. Flynn, M. Barnes, J. Keane, G. Nenadic, Machine
learning methods for wind turbine condition monitoring: A review, Renewable energy, 133 (2019) 620-
635.
[156] J. Carroll, S. Koukoura, A. McDonald, A. Charalambous, S. Weiss, S. McArthur, Wind turbine
gearbox failure and remaining useful life prediction using machine learning techniques, Wind Energy,
22 (2019) 360-375.
[157] Z. Song, Z. Zhang, Y. Jiang, J. Zhu, Wind turbine health state monitoring based on a Bayesian data-
driven approach, Renewable energy, 125 (2018) 172-181.
[158] F. Cheng, L. Qu, W. Qiao, L. Hao, Enhanced Particle Filtering for Bearing Remaining Useful Life
Prediction of Wind Turbine Drivetrain Gearboxes, IEEE Transactions on Industrial Electronics, 66
(2019) 4738-4748.
[159] Y. Hu, H. Li, P. Shi, Z. Chai, K. Wang, X. Xie, Z. Chen, A prediction method for the real-time
remaining useful life of wind turbine bearings based on the Wiener process, Renewable Energy, 127
(2018) 452-460.
[160] J.S. Nielsen, J.D. Sørensen, Bayesian estimation of remaining useful life for wind turbine blades,
Energies, 10 (2017) 664.
[161] L. Saidi, J.B. Ali, E. Bechhoefer, M. Benbouzid, Wind turbine high-speed shaft bearings health
prognosis through a spectral Kurtosis-derived indices and SVR, Applied Acoustics, 120 (2017) 1-8.
[162] H.D.M. de Azevedo, A.M. Araújo, N. Bouchonneau, A review of wind turbine bearing condition
monitoring: State of the art and challenges, Renewable and Sustainable Energy Reviews, 56 (2016)
368-379.
[163] J.P. Salameh, S. Cauet, E. Etien, A. Sakout, L. Rambault, Gearbox condition monitoring in wind
turbines: A review, Mechanical Systems and Signal Processing, 111 (2018) 251-264.

49
[164] H. Wang, M.-j. Peng, Y.-k. Liu, S.-w. Liu, R.-y. Xu, H. Saeed, Remaining Useful Life Prediction
Techniques of Electric Valves for Nuclear Power Plants with Convolution Kernel and LSTM, Science
and Technology of Nuclear Installations, 2020 (2020).
[165] H. Wang, M.-j. Peng, Z. Miao, Y.-k. Liu, A. Ayodeji, C. Hao, Remaining useful life prediction
techniques for electric valves based on convolution auto encoder and long short term memory, ISA
transactions, (2020).
[166] J.I. Aizpurua, S.D.J. McArthur, B.G. Stewart, B. Lambert, J.G. Cross, V.M. Catterson, Adaptive
power transformer lifetime predictions through machine learning and uncertainty modeling in nuclear
power plants, IEEE Transactions on Industrial Electronics, 66 (2018) 4726-4737.
[167] H.-P. Nguyen, P. Baraldi, E. Zio, Ensemble empirical mode decomposition and long short-term
memory neural network for multi-step predictions of time series signals in nuclear power plants,
Applied Energy, (2020) 116346.
[168] Y. Li, K. Liu, A.M. Foley, A. Zülke, M. Berecibar, E. Nanini-Maury, J. Van Mierlo, H.E. Hoster,
Data-driven health estimation and lifetime prediction of lithium-ion batteries: A review, Renewable
and Sustainable Energy Reviews, 113 (2019) 109254.
[169] B. Saha, K. Goebel, Battery data set, NASA AMES prognostics data repository, (2007).
[170] P. Michael, Battery Data Set, CALCE Battery Research Group, Maryland, MD,2017, 2017, pp.
https://web.calce.umd.edu/batteries/index.html.
[171] W. He, N. Williard, M. Osterman, M. Pecht, Prognostics of lithium-ion batteries based on Dempster–
Shafer theory and the Bayesian Monte Carlo method, Journal of Power Sources, 196 (2011) 10314-
10321.
[172] Q. Zhai, Z.-S. Ye, RUL prediction of deteriorating products using an adaptive Wiener process model,
IEEE Transactions on Industrial Informatics, 13 (2017) 2911-2921.
[173] Y. Shen, L. Shen, W. Xu, A Wiener‐based degradation model with logistic distributed measurement
errors and remaining useful life estimation, Quality and Reliability Engineering International, 34 (2018)
1289-1303.
[174] H. Wang, X. Ma, Y. Zhao, A mixed-effects model of two-phase degradation process for reliability
assessment and RUL prediction, Microelectronics Reliability, 107 (2020) 113622.
[175] Z.-X. Zhang, X.-S. Si, C.-H. Hu, M.G. Pecht, A prognostic model for stochastic degrading systems
with state recovery: Application to Li-ion batteries, IEEE Transactions on Reliability, 66 (2017) 1293-
1308.
[176] X.-S. Si, An adaptive prognostic approach via nonlinear degradation modeling: Application to battery
data, IEEE Transactions on Industrial Electronics, 62 (2015) 5082-5096.
[177] L. Chen, Y. Zhang, Y. Zheng, X. Li, X. Zheng, Remaining useful life prediction of lithium-ion battery
with optimal input sequence selection and error compensation, Neurocomputing, 414 (2020) 245-254.
[178] Z. Xue, Y. Zhang, C. Cheng, G. Ma, Remaining useful life prediction of lithium-ion batteries with
adaptive unscented kalman filter and optimized support vector regression, Neurocomputing, 376 (2020)
95-102.
[179] P. Khumprom, N. Yodo, A data-driven predictive prognostic model for lithium-ion batteries based
on a deep learning algorithm, Energies, 12 (2019) 660.
[180] L. Ren, L. Zhao, S. Hong, S. Zhao, H. Wang, L. Zhang, Remaining Useful Life Prediction for
Lithium-Ion Battery: A Deep Learning Approach, IEEE Access, 6 (2018) 50587-50598.
[181] R. Jiao, K. Peng, J. Dong, Remaining Useful Life Prediction of Lithium-Ion Batteries Based on
Conditional Variational Autoencoders-Particle Filter, IEEE Transactions on Instrumentation and
Measurement, (2020).
[182] J.F. Murray, G.F. Hughes, K. Kreutz-Delgado, Machine learning methods for predicting failures in
hard drives: A multiple-instance application, Journal of Machine Learning Research, 6 (2005) 783-816.
[183] B. Zhu, G. Wang, X. Liu, D. Hu, S. Lin, J. Ma, Proactive drive failure prediction for large scale
storage systems, IEEE, pp. 1-5.
[184] N. Aussel, S. Jaulin, G. Gandon, Y. Petetin, E. Fazli, S. Chabridon, Predictive models of hard drive
failures based on operational data, IEEE, pp. 619-625.

50
[185] C. Xu, G. Wang, X. Liu, D. Guo, T. Liu, Health Status Assessment and Failure Prediction for Hard
Drives with Recurrent Neural Networks, IEEE Transactions on Computers, 65 (2016) 3502-3508.
[186] J. Li, R.J. Stones, G. Wang, X. Liu, Z. Li, M. Xu, Hard drive failure prediction using decision trees,
Reliability Engineering & System Safety, 164 (2017) 55-65.
[187] F.D.S. Lima, F.L.F. Pereira, L.G.M. Leite, J.P.P. Gomes, J.C. Machado, Remaining useful life
estimation of hard disk drives based on deep neural networks, IEEE, pp. 1-7.
[188] A. Ruiz-Tagle Palazuelos, E.L. Droguett, R. Pascual, A novel deep capsule neural network for
remaining useful life estimation, Proceedings of the Institution of Mechanical Engineers, Part O:
Journal of Risk and Reliability, 234 (2020) 151-167.
[189] C. Zhang, P. Lim, A.K. Qin, K.C. Tan, Multiobjective Deep Belief Networks Ensemble for
Remaining Useful Life Estimation in Prognostics, IEEE Transactions on Neural Networks and
Learning Systems, 28 (2017) 2306-2318.
[190] Y. Liao, L. Zhang, C. Liu, Uncertainty prediction of remaining useful life using long short-term
memory network based on bootstrap method, 2018 IEEE International Conference on Prognostics and
Health Management (ICPHM), IEEE, 2018, pp. 1-8.
[191] X. Li, Q. Ding, J.-Q. Sun, Remaining useful life estimation in prognostics using deep convolution
neural networks, Reliability Engineering & System Safety, 172 (2018) 1-11.
[192] A.L. Ellefsen, E. Bjørlykhaug, V. Æsøy, S. Ushakov, H. Zhang, Remaining useful life predictions
for turbofan engine degradation using semi-supervised deep architecture, Reliability Engineering &
System Safety, 183 (2019) 240-251.
[193] G. Hou, S. Xu, N. Zhou, L. Yang, Q. Fu, Remaining Useful Life Estimation Using Deep
Convolutional Generative Adversarial Networks Based on an Autoencoder Scheme, Computational
Intelligence and Neuroscience, 2020 (2020).
[194] M. Kim, K. Liu, A Bayesian deep learning framework for interval estimation of remaining useful life
in complex systems by incorporating general degradation characteristics, IISE Transactions, 53 (2020)
326-340.
[195] L. Ren, J. Cui, Y. Sun, X. Cheng, Multi-bearing remaining useful life collaborative prediction: A
deep learning approach, Journal of Manufacturing Systems, 43 (2017) 248-256.
[196] L. Liao, F. Köttig, A hybrid framework combining data-driven and model-based methods for system
remaining useful life prediction, Applied Soft Computing, 44 (2016) 191-199.
[197] M. Djeziri, S. Benmoussa, R. Sanchez, Hybrid method for remaining useful life prediction in wind
turbine systems, Renewable Energy, 116 (2018) 173-187.

51