Journal of Dynamics, Monitoring and Diagnostics, 2023, 2, 1-12

https://doi.org/10.37965/jdmd.2023.148 FUTURE DIRECTION PAPER

Prognostics and Remaining Useful Life Prediction

of Machinery: Advances, Opportunities, and Challenges
JDMD Editorial Office,1 Nagi Gebraeel,2† Yaguo Lei,3† Naipeng Li,3
Xiaosheng Si,4† and Enrico Zio5,6†
1
Editorial Office of JDMD, Chongqing University of Technology, Chongqing, People’s Republic of China
2
H. Milton Stewart School of Industrial and Systems Engineering,
Georgia Institute of Technology, Atlanta, GA, USA
3
Key Laboratory of Education Ministry for Modern Design and Rotor-Bearing System,
Xi’an Jiaotong University, Xi’an, People’s Republic of China
4
Zhijian Laboratory, Rocket Force University of Engineering, Xi’an, People’s Republic of China
5
MINES Paris, PSL Research University, Sophia Antipolis, France
6
Energy Department, Politecnico di Milano, Milan, Italy

(Received 23 December 2022; Revised 03 January 2023; Accepted 18 January 2023; Published online 18 January 2023)

Abstract: As the fundamental and key technique to ensure the safe and reliable operation of vital systems,
prognostics with an emphasis on the remaining useful life (RUL) prediction has attracted great attention in the last
decades. In this paper, we briefly discuss the general idea and advances of various prognostics and RUL prediction
methods for machinery, mainly including data-driven methods, physics-based methods, hybrid methods, etc. Based
on the observations from the state of the art, we provide comprehensive discussions on the possible opportunities and
challenges of prognostics and RUL prediction of machinery so as to steer the future development.

Keywords: prognostics; remaining useful life; data-driven; machine learning; degradation modeling

I. INTRODUCTION conducted over the past two to three decades has been
recorded in several comprehensive survey papers such as
This paper reflects the important aspects in the field of [1–4]. These papers provide multiple perspectives on how
prognostics and remaining useful life (RUL) prediction of to classify the current literature in prognostics and remain-
machinery. Opportunities and challenges, as well as future ing life predictions. Figure 1 attempts to provide a simple
directions are discussed. Section II on overview of prog- overarching taxonomy of existing works in prognostics.
nostics research and future research opportunities was Our taxonomy is more or less consistent with the tradi-
completed by Professor Nagi Gebraeel from Georgia Insti- tional categorization of modeling approaches into model-
tute of Technology. Section III on opportunities and chal- based frameworks, data-driven models, and hybrid
lenges in RUL prediction of machinery was written by approaches that combine the two. We provide two addi-
Professor Yaguo Lei and Dr. Naipeng Li from Xi’an tional dimensions. The first deals with the modeling
Jiaotong University. Section IV on opportunities and chal- assumptions about the environmental and/or operational
lenges in statistical data-driven prognostics was presented covariates and their impact on remaining life predictions.
by Professor Xiaosheng Si from Rocket Force University of The second focuses on data dimensionality starting with
Engineering. Section V on prediction of RUL: future univariate and multivariate time series and moving on to
directions was written by Professor Enrico Zio from PSL profile and image data.
Research University and Politecnico di Milano.
B. METHODOLOGIES USED DEVELOPING
II. OVERVIEW OF PROGNOSTICS PROGNOSTIC MODELS
RESEARCH AND FUTURE RESEARCH 1) MODEL-BASED PROGNOSTICS. Model-based ap-
OPPORTUNITIES proaches assume the existence of a mathematical model
that exploits physical knowledge of a system to derive
A. OVERVIEW phenomenological equations that characterize system deg-
There is a plethora of work centered on prognostics and radation. In this context, prognostics are typically split into
estimating remaining lifetime. Much of the work two sequential problems: a state-estimation problem where
system health is assessed based on data observations, and a
prediction problem that utilizes filtering techniques, such as
Corresponding author: JDMD Editorial Office (e-mail: dmd@istp- Kalman filter, unscented Kalman filter, and particle filter to
press.com). simulate the state distribution forward in time to RUL [4–6].
†
The joint first authors (Nagi Gerbraeel ([email protected]), Model-based prognostics have been successfully imple-
Yaguo Lei ([email protected]), Xiaosheng Si (sixiaosheng@126. mented in diverse applications ranging from batteries
com), Enrico Zio ([email protected]; [email protected]). [5,7,8] to various types of rotating machinery [9,10].

© The Author(s) 2023. This is an open access article published under the CC BY license (https://creativecommons.org/licenses/by/4.0/). 1
2 JDMD Editorial Office et al.

Fig. 1. Classification of prognostic models.

Model-based prognostics offer some key benefits. They processes are two of the most predominant techniques that
often show good prediction accuracy when the model de- rely on statistical/probabilistic frameworks for modeling
gradation parameters are estimated accurately. They can also degradation. These techniques are based on mapping the
estimate RUL conditional on the future operational and/or state of health of a component (treated as the dependent
environmental conditions by modeling incorporating them variable) to a set of independent variables (e.g., time, usage,
as external inputs to the state-space equations. However, the environment, etc.). The mapping is then used to predict how
model-based approach relies almost exclusively on the the state of health of the system will behave to certain changes
existence and the estimation of a parametric mathematical in the independent variables. A large portion of the literature
representation of the system dynamics and its degradation in this space tries to predict RUL by estimating the time it
process. This is often hard to derive for complex systems and takes to cross some predefined critical degradation threshold
also difficult to generalize. The propagation of the system [3,4]. Many Markov-based techniques usually define some
state forward in time can also become computationally hidden Markov model with a finite number of states and an
prohibitive, especially for systems with high-dimensional observed process that depends on the hidden one [1]. This
state-space representation. Additionally, a substantial approach is suited for applications where degradation states
amount of data is required for parameter estimation. are not directly observable, yet the data still depend on the
2) DATA-DRIVEN PROGNOSTICS. Data-driven ap- degradation state [3].
proaches attempt to learn certain patterns and statistical ML and AI are very powerful tools when it comes to
characteristics present in historical data that can be indica- data-driven prognostic models [3]. They are usually very
tive of the component’s state of health. These patterns and generalizable and can be quite effective when the data are
trends can be utilized to estimate the component’s RUL. abundant and curated properly. Another benefit of ML-/AI-
RUL estimation is generally performed through multivari- based appears clearly when modeling high-dimensional data
ate pattern matching or by extrapolating the current state of such as spectral and image data. Such applications need
health to a predefined threshold [4]. In contrast to model- careful feature extraction to extract informative features that
based approaches, data-driven approaches generally do not are correlated with the underlying physical degradation.
require specific domain knowledge and expertise, or com- There are numerous feature extraction methods and
plex phenomenological models that describe a system’s techniques, some commonly used tools include principal
physics. This makes them a popular choice when modeling component analysis (PCA), functional-PCA, wavelets, con-
complex systems for which creating a physical model might volutional neural networks (CNNs), and variational auto-
be extremely difficult, or even impossible. encoder. It is noteworthy to mention that the accuracy of
Data-driven prognostic models can be loosely classified many RUL predictions is highly dependent on the efficacy
into two types: statistical-based approaches and machine of the feature extraction process. Additionally, the effective-
learning (ML)-/artificial intelligence (AI)-based frameworks. ness of ML/AI models relies on the availability of large
Statistical approaches generally attempt to learn existing volumes of properly curated data – something that is often a
techniques between the variables in the data (such as the major challenge in many industrial applications.
relationship between a degradation signal and operating time) Interpretability is often a key limitation in many data-
and utilize those relationships to make predictions about driven frameworks. Most statistical and ML-based models
future behavior [3]. Regression-based models and Markov generally lack physical interpretability [3]. This poses

JDMD Vol. 2, No. 1, 2023

 Prognostics and Remaining Useful Life Prediction of Machinery 3

serious challenges in building trust between human opera- aspects resulting in reducing the false alarms and improving
tors and predictions generated by these models. the prediction accuracy. Additional works that considered a
similar setup include [28–30].
3) HYBRID PROGNOSTICS MODELS. Hybrid ap-
proaches are usually a combination of model-based and
data-driven approaches. A review on hybrid prognostics is D. DIMENSIONALITY OF DEGRADATION
presented in [11]. One common approach in hybrid prognos- DATA
tics algorithms is the use of data-driven methods (e.g., neural
network [NN], long short-term memory network LSTM, Popular prognostics modeling frameworks involve model-
radial basis function network RBF) to create a mapping ing how degradation signals evolve over time and using that
from sensor measurements to the state space and then using information by estimating the time remaining for the signal
a state-space model to model the evolution of the degradation to cross a predefined failure threshold. Typically, a degra-
state [12,13]. Such approaches are more comprehensive and dation signal is computed from specific features obtained
combine the benefits of both frameworks. However, they are from the raw sensor data. Most prognostic models devel-
still deeply rooted in building customized models that are oped in the literature are designed to model a univariate
relatively application specific. The creation of a truly general- signal representing a time series of degradation-based data
ized hybrid modeling framework is still an illusive endeavor. evolving over time.
1) MULTIVARIATE DEGRADATION SIGNALS. Multi-
sensor application often involved complex equipment that
C. PROGNOSTIC MODELS AND typically undergo multifaceted degradation processes. Using
COVARIATES multiple sensors can potentially capture different aspects of
Generally speaking, harsh environments tend to accelerate the complicated degradation processes that usually involve
the degradation mechanisms that occur prior to failure as different failure modes. Overall, the data in such cases are
compared to milder environments. Yet, as noted by [14], the much richer and can lead to more accurate failure predic-
vast majority of conventional failure models assume that tions. One of the key aspects in this setting is how to
prevailing environmental conditions are temporally invari- systematically combine information from multiple sensors
ant or have no effect on deterioration and failure processes. from the same equipment, otherwise known as fusion. [1]
The limited number of failure models that do consider provides a review of multi-sensor data fusion approaches
environmental effects generally belong to one of two and classifies the techniques based on the level at which
groups: (1) hazard rate function models that treat environ- fusion is performed: data, feature, and decision levels.
mental conditions as model covariates [15–17] and (2) sto- Data-level fusion directly integrates information of the
chastic wear and/or shock models in which the wear and/ raw data from multiple sensors. Feature-level fusion com-
or shock intensities are modulated by the environment bines feature information extracted from the raw data.
(cf. [18,19]). However, even these models have certain Decision-level fusion focuses on integrating different diag-
features that limit their applicability. First, the hazard rate nostic or prognostic results. A large portion of the fusion
function is only useful for making inferences about a large literature utilizes AI approaches, such as NN and fuzzy
population of components, but not about specific compo- logic. However, most of these models have been used for
nents as it cannot be easily observed or measured for fault detection and diagnostics, much less prognostics. A
individual components [20]. The second group of models few examples of recent prognostic models for multi-sensor
is useful for deriving analytical lifetime distributions (or applications can be found in [31–34]. Some of the ap-
their transforms) and assessing, probabilistically, the full or proaches relies on computing an aggregate composite
residual lifetime of the component. However, these models health index [31]. Typically, a univariate aggregate signal
treat failure as a random event and do not provide informa- is constructed by taking a weighted combination of various
tion about the evolution of the physical degradation process degradation signals from individual sensors. In other cases,
that occurs prior to failure [21]. the data fusion is performed through a state-space modeling
framework that is used to represent the overall degradation
1) TIME-VARYING PROGNOSTIC MODELS. Few models
state of a system [32,33]. Work developed by [34] was
have studied predicting RUL under time-varying environ-
among the first to formally leverage the covariance structure
mental and operating conditions; examples include load-
governing a multivariate stream of degradation signals.
dependent degradation models such as batteries [8,22,23]
Other multivariate degradation models rely on clustering
and bearings [24–26]. [27] proposed a methodology for
analysis where historical data are divided into different
modeling degradation signals from components functioning
subsets that characterize the health states at different deg-
under dynamically evolving environment conditions where
radation levels [35,35–37].
in-situ sensor signals related to the degradation process as
well as the future environment conditions were utilized when 2) IMAGE-BASED DEGRADATION SIGNALS. Imaging is
predicting the components RUL. The model assumed that the one of the fastest growing technologies for condition
time-dependent degradation rate where a component's deg- monitoring and industrial asset management. Conventional
radation signal increases (or decreases) is affected by the approaches that utilize random coefficients models, Brow-
severity of the environmental condition. These conditions are nian motion, gamma process, and functional data analysis
assumed to evolve as a continuous-time Markov chain. [26] to model how degradation signals evolve over time are not
extended this work by considering a state-space modeling suitable for characterizing the spatio-temporal correlations
framework where changes in the degradation rate are part of a that exist in image data. One of the key challenges with
state transition function, and jumps in the degradation signals modeling image data revolve around the analytical and
due to environmental changes are part of a measurement computational challenges associated with modeling high-
function. The separate analysis of these two factors made it dimensional data streams. The high dimensionality arises
possible to distinguish the unique contribution of these two from the fact that a single image streams may consist of a

JDMD Vol. 2, No. 1, 2023

4 JDMD Editorial Office et al.

large sequence of images (observed across the life cycle of especially in the context of shocks and load sharing sce-
an equipment) coupled with the large numbers of pixels narios [44]. Other models have studied economic and
embedded in each image. Another challenge is the complex stochastic dependence in the context of opportunistic main-
spatial–temporal structures within these image streams. tenance. A review of optimal maintenance of multi-
Pixels are spatially correlated within a single image and component systems can be found in [45]. The paper states
temporally correlated across sequential images. In recent that interactions between components complicate the
work by [38], degradation-based image data streams were modeling and optimization of maintenance but offer an
modeled as a spatio-temporal process. [39] proposed two opportunity to group maintenance which may save costs.
deep-learning methods for estimating time-to-failure in From the viewpoint of this paper, one of the under-
industrial systems using image sensor data. The authors studied topics in prognostics has to do with predicting RUL
utilized CNN and autoencoder to collect useful information of systems comprised of components with interdependent
from images, which are high dimensional, and then train a degradation processes. One of the key questions is how
LSTM-based regression model to predict time-to-failure. specific levels of degradation in one component affect
[40] developed a methodology to predict residual useful the degradation rates of other components in the system.
lifetime of a system based on a sequence of degradation The characteristics of these interactions may vary from one
images. The methodology has two main steps. First, it system to another. Some interactions may be triggered by
projects the image tensors onto a low-dimensional space. specific degradation levels where others can be more subtle
Next, the projected tensors are regressed against time-to- and evolve continuously over time. Consequently, the abil-
failure via penalized location-scale tensor regression. ity to formally characterize these interdependencies and
The coefficient tensor is decomposed using CANDE- account for their behavior over time and perhaps even
COMP/PARAFAC (CP) and Tucker decompositions, over different levels of component degradation would be
thereby enabling parameter estimation in a high-dimen- worth investigating. Additionally, predicting the RUL of
sional setting. [41] trained different NNs (deep neural each component independently versus the RUL of the
network DNN, LSTM, and CNN) to capture the correlation system is rich topics of research. Model-based approaches
between the degradation image stream and its remaining might prove more successful than data-driven ones in
lifetime. They additionally implemented a multiple performing prognosis on systems with multiple interacting
weighted time window policy to increase the prediction components due to their ability to fully capture the underly-
accuracy of the NN. This policy takes into account not only ing physics phenomena of the system and the interactions
the most recent monitoring data but previous observations between the components. However, such models would be
as well. The proposed image stream-based regressors are extremely challenging, if not impossible, to generate for the
validated by using two datasets of degradation infrared complex systems in consideration.
images, showing that the LSTM achieves the best perfor-
mance on the accuracy. 2) PROGNOSTICS AND DATA SPARSITY. Many prog-
nostic models have been developed on the premise that the
degradation signals are observed with high fidelity at
E. SOME OPEN QUESTIONS IN frequent time steps. In reality, however, degradation ob-
PROGNOSTICS servations often contain outliers as well as missing and
corrupt data that result from the harsh industrial environ-
There have been significant research advances in the field of ments that equipment operate in. For example, there are
prognostics and predictive modeling. Given these ad- numerous issues with data communication, network con-
vances, there are still some open topics that still require nectivity, read/write, storage formats, etc. Developing
more exploration and research. Below are some topics that models that are robust to high levels of missing and corrupt
the authors believe are understudied, yet seem to be impor- data is needed for accurate predictions of RUL.
tant to advancing the field of prognostics. Another important topic related to data sparsity is the
1) PROGNOSTICS OF SYSTEMS WITH INTERDEPEN- lack of failure time data. In many practical applications,
DENT DEGRADATION PROCESSES. Predicting the re- components and systems are repaired preventively before
maining lifetime of multi-component systems requires an any catastrophic failure. In fact, catastrophic failures are so
accurate evaluation of the degradation states of its constituent rare in industry because many critical equipment that
components. More importantly and perhaps significantly, require prognostics often operate in a risk-averse ecosys-
more challenging is the need to characterize failure and tem. As a result, many practical applications only have
degradation interactions among the critical components of partial degradation signals with “censored” failure times.
the system. Characterizing these interdependencies is indeed What makes this problem even more challenging is that
very difficult, so much so that many reliability prediction most of these partial degradation signals are of different
models have circumvented this challenge by assuming that lengths and do not conform to a fixed threshold, since
component lifetimes (within a given system) are indepen- replacement and repairs are not triggered by a predefined
dent. Although such assumptions help to obtain mathemati- threshold. These settings present a unique challenge for
cally tractable models, they remain unrealistic, especially for ML-based prognostics models that rely on supervised
applications where dependencies are indeed present. learning (labeled failure times) that utilize large volumes of
Most conventional models that study component data [46]. Some recent works have approached this problem
dependencies can be divided into two main groups. The from the perspective of few-shot learning [47,48]. How-
first group encompasses models that study how the failure ever, this topic is still one of the open problems that has not
of one component affects the failure rate of other compo- yet been formally investigated.
nents [42,43]. The second group focuses on models that While recently there have been many works that
employ a more statistical approach by developing multi- attempt to develop prognostic models for a wide array of
variate distributions of system component lifetimes, applications using varying techniques, there are still some

JDMD Vol. 2, No. 1, 2023

 Prognostics and Remaining Useful Life Prediction of Machinery 5

areas that have not been fully explored yet. One such area is effectiveness when integrated with decision optimization
the prognosis of systems with multiple interacting, gradually models [53]. In this regards, the development of decision
degrading components. This is because when the compo- models that can leverage RUL predictions is one of the future
nents interact, the degradation of one component can sig- directions that need to be examined and explored carefully.
nificantly affect the rate of degradation of the others, either
increasing or decreasing their individual RULs. This leads to
the utilization of the current state of health to extrapolate the III. OPPORTUNITIES AND
degradation trajectory of the system to predict its RUL CHALLENGES IN RUL
becoming a much more challenging task. Therefore, while
there are works such as [49] where the authors successfully PREDICTION OF MACHINERY
utilize data-driven approaches to diagnose the state of health
of such systems, performing system prognosis utilizing
A. A BRIEF INTRODUCTION
similar approaches remains a largely unexplored area. Operational maintenance plays a major role in keeping the
3) COMPONENTS THAT EXHIBIT INTERMITTENT
safety and reliability of machinery. With the development of
FAULTS. Another area that has not been fully explored the sensor technology and Internet of Things (IoTs), condi-
yet is the prognosis of components that exhibit intermittent tion-based maintenance (CBM) has become the most popular
faults. Intermittent faults are faults that occur randomly and effective maintenance strategy in industrial practice [54].
during the component’s operation and then disappear shortly The basic idea of CBM is to estimate the health state of
after without the need for any repair activities. An example of machinery by capturing on-line monitoring signals using
such fault is a valve that randomly gets stuck in a certain different kinds of sensors, such as vibration signals, tem-
position during its operation and then returns back to the peratures, motor currents, and acoustic emissions, and con-
operational state without any intervention. Such components duct maintenance schedule based on real-time monitoring
do not exhibit the traditional degradation signal where the results. To prepare spare components in advance and sched-
occurrence of a certain fault marks the start of degradation, ule a precise time of repair, industrial managers need to know
and then a sensor reading monotonically and gradually the RUL of the machinery in-service at its early degradation
changes over time to reflect the worsening of said fault. stage. RUL prediction aims to forecast the time left before the
Instead, degradation in these components manifests as a machinery reaches the final failure.
gradual increase in the frequency or the intensity of the RUL prediction is actually a tough issue in most indus-
intermittent faults. It must also be noted that the intermittent trial scenarios, since the damage, degradation, and failure of
faults might not be obviously visible in the sensor data, machinery are usually affected by various uncertainty re-
which introduces another challenge in analyzing it. There are sources, such as the operational conditions, the quality of the
a few examples in literature where the authors successfully product, the working environment, and the service task. It is
utilize either model-based or data-driven approaches to really difficult to forecast the future degradation trend based
perform prognosis on components that exhibit intermittent on historical observations and provide an accurate RUL
faults. Examples include [50] where the authors utilize an prediction result. To deal with this tough issue, lots of
extreme learning machine to predict the RUL of an electrical research work have been conducted in recent years.
connector in vibration environments, [34] where the authors In terms of technical processes, RUL prediction can be
utilized a Bond graph to preform prognosis on an electric divided into the following four steps. The first step is to
scooter, and [51,52] where the authors utilize a linear model capture condition monitoring signals which reflect the deg-
to predict the remaining time until a threshold on the radation behavior of machinery. The second step is
proportion of time that a component spends at a faulty state to construct health indicators (HIs) from monitoring signals
is crossed. While all of these approaches, and many others, to quantify the degradation severity. The failure criterion is
have proven successful in their individual applications, they generally defined based on a specified failure threshold of
all share the same limitation: lack of generalizability. This is corresponding HIs. The third step is to divide the health
because most of the current works utilize specific knowledge stages according to the varying degradation trends of HIs.
and information about the components under study to build The purpose of this step is to identify some important time
stamps including the first degradation time, the first pre-
their prognosis technique, making them only applicable to a
dicting time [55], the stage switching time, etc. The last step
certain component or a class of components. To eliminate the
is to conduct RUL prediction at degradation stages by
need to create a new approach for each individual compo-
mapping different models and the degradation data. More
nents or class of components, there should exist a prognosis
details about the technical processes can be found in the
approach that is generalizable so that it can be utilized for a
systematic review paper regarding to RUL prediction [28].
wide variety of components that exhibit intermittent faults.
In terms of modeling theories, prognostic methods can
Such approach would need to be data-driven instead of
be broadly classified into physics model-based methods,
model based to maximize the generalizability and to ensure
data-driven methods, and their hybrid methods. Physics
applicability to highly complex components.
model-based methods describe degradation processes of
4) PROGNOSTICS-BASED DECISION MODELS. The machinery by constructing functional models on the basis
final goal of prognostics algorithms is to provide insights of the failure mechanisms or the first principle of damage. In
to decision makers, thereby helping them make condition- real practice, it is actually a big challenge to construct a
based decisions regarding system operations, maintenance high-precision physical model to describe the degradation
scheduling, and even spare parts logistics. Existing prog- behavior of machinery. With the increase of complexity
nostics algorithms, however, are usually designed and and integration of mechanical systems, the damage of each
tested without considering their use in decision-making. component will interact with each other. It becomes more
These prognostics algorithms are trained to maximize pre- and more difficult to understand the physics of damage
diction accuracy, which might not necessarily guarantee their and formulate the degradation behavior of machinery.

JDMD Vol. 2, No. 1, 2023

6 JDMD Editorial Office et al.

Conversely, data-driven prognostic methods have been support vector machine, K-nearest neighbor, Gaussian
developed broadly with the advancement of condition process regression, etc. They are named as traditional
monitoring technology. ML techniques hereafter for simplicity. Nowadays, with
Data-driven prognostic methods mainly include statisti- the advances of deep learning (DL) techniques, such as
cal data-driven methods and ML)-based methods [28,56]. A CNNs, recurrent neural networks (RNNs), and Bayesian
systematic review about the basic idea and major processes deep learning, it has attracted more and more attention in
of statistical data-driven methods has been provided in the this research field. Traditional ML-based prognostic meth-
review paper [56]. As a kind of prognostic methods devel- ods generally work together with a feature extraction
oped from statistical theory, they are superior in describing process. A set of features are first extracted from condition
the stochastic characteristics of degradation processes and monitoring data. Then, the features are input into the ML
quantifying the uncertainty of prediction results. Some models to conduct the RUL prediction task. DL techniques
research works have been conducted to deal with different have the capability of analyzing high-dimensional data and
kinds of uncertainty resources, including the unit-to-unit extracting features automatically from data. Thus, they can
variability [57], the temporal variability [58,59], the realize the “end-to-end” RUL prediction, that is, input the
measurement noises [60], the time-varying operational con- original data into the model and output the RUL result
ditions [55,61,62], etc. The major idea of uncertainty descrip- directly. According to data processing strategies, ML-based
tion is to describe the uncertainty resources by introducing RUL prediction approaches can be roughly classified into
random parameters into the degradation model. Then, the the following three categories.
model parameters are estimated according to real-time con- (1) Constructing a fusion HI for RUL prediction. A HI
dition monitoring data. The uncertainty of the degradation with obvious monotonic and stable degradation trend
process is further transmitted into the probability density is helpful to facilitate the RUL prediction process.
function of the predicted RUL. It is of crucial importance to Some researchers [63,64] attempt to construct a
analyze the dominant uncertainty resource in particular cases good HI by employing ML techniques. This strategy
and quantify its contribution to the RUL prediction result. utilities the high capability of ML techniques in non-
For example, both measurement noises and time-varying linear relation mapping to construct HIs from condition
operational conditions can introduce amplitude fluctuation monitoring data for RUL prediction. The major task is
in the condition monitoring data. However, the mechanisms to map the high-dimensional original data into a one-
of these two factors are absolutely different with each other. dimensional HI sequence which is able to represent the
To be specific, the measurement noises are caused by the degradation process of machinery. To ensure the high
inherent characteristic of data acquisition systems and the quality of the HIs, the construction process is generally
interference of environment. They are generally presented as guided by some evaluation criteria such as the mono-
random fluctuation and assumed to be normally distributed. tonicity, trendability, and robustness. New criteria can
The fluctuation caused by the operation conditions is highly also be developed and involved into the model accord-
correlated with the condition profiles, which can be reduced ing to different requirements of prediction tasks.
by using a kind of baseline condition transformation strategy
[55]. As a result, the degradation trend of the health state can (2) Predicting the degradation process using the strategy of
be highlighted while the impact of operational conditions can time series forecasting. In this strategy, the ML tech-
be ruled out, which is helpful for improving the precision and niques are used to learn the recurrent relationship
reliability of prediction result. between the time series data of HIs, which can be
ML-based RUL prediction methods introduce the formulated using the generalized expression xiþ1 =
advanced techniques of ML into the area of RUL predic- f ðxi , xi−1 , · · · , xi−p Þ, where fð·Þ represents the map-
tion. They attempt to learn the degradation patterns from ping relationship function, ðxi , xi−1 , · · · , xi−p Þ is the
available observations automatically without depending on input of the model that is the HI observations of
the first principle of degradation or expert knowledge. previous time steps, and xiþ1 is the output of the model
Therefore, this kind of approaches is more suitable to that is the HI value in the next time step. The RUL can
deal with prognostic issues where no prior knowledge is be predicted by inputting the predicted HI values step
available. The basic requirement is that sufficient data need by step until exceeding a specified failure threshold.
to be provided for model training. However, it is a tough (3) Predicting the RUL using a straightforward mapping
requirement for the task of RUL prediction, since the strategy toward RUL. Different from the time series
collection of whole-life degradation data is time consum- prediction, this strategy maps the relationship between
ing, and it is sometimes impossible in cases where serious current health states to the RUL values. The input
failure is not allowed. The interpretability is also a major variables can be original data or HIs extracted from
concern in the application of ML-based approaches. ML them, which represent the current health state of the
models generally care more about the mapping relationship system. The output is the RUL value or its ratio to the
between the input data and output data. They do not care total lifetime. This strategy can achieve the RUL
about the inherent interpretability of the relation. Because of prediction directly. It is more straightforward than
the lack of transparency, they are always named as “black the previous one in terms of procedure. However, it
boxes”. However, in industrial scenarios, the interpretabil- puts forward higher requirements for the nonlinear
ity of prediction result is significant for maintenance deci- mapping capacity of the prognostic model.
sion. It is difficult to convince operators to accept the
prediction result if it is uninterpretable.
B. OPPORTUNITIES AND CHALLENGES
The development history ML-based RUL prediction
is highly influenced by the advances in ML techniques. There is no doubt that great advancements have been
The early techniques introduced into RUL prediction achieved in the RUL prediction of machinery. However,
include the artificial neural network, neural fuzzy system, most research is conducted in the laboratory environment

JDMD Vol. 2, No. 1, 2023

 Prognostics and Remaining Useful Life Prediction of Machinery 7

wherein the working parameters are controllable. The real development of prognostic research. Thanks to these pub-
practice scenarios are generally more complicated than lished datasets, researchers can develop various prognostic
laboratory environment. Current academic research in approaches and compare their approaches with existing
RUL prediction is still far from practical application. To ones using the same benchmark datasets. This is significant
promote the development of RUL prediction technique in for the development of the prognostic theories and meth-
industrial practice, there are still lots of big challenges in the odologies. However, most of existing accelerated degrada-
future research. Some suggested research directions are tion datasets are generated in the laboratory environment
provided as follows. that are totally different from the industrial scenarios. The
(1) Self-data-driven prognostic approaches degradation behaviors of systems in real industrial cases
suffer from more complicated uncertainties, including the
It is seen from the above introduction that most data- time-varying operational conditions and the interference
driven prognostic approaches need sufficient whole-life from outside environment. Therefore, researchers are
degradation data to train models. Although data acquisition encouraged to conduct more degradation tests under realis-
becomes more easier with the advances of sensor techni- tic operational conditions and publish the degradation
ques, it is still a tough issue to capture high-quality and - dataset. In addition, we also appeal to the companies to
quantity whole-life degradation data in industrial practice. accumulate and share the degradation datasets of their real
The first reason is that the life time of an industrial system is industrial equipment. It would not only promote the devel-
generally many years. It is time-consuming and high cost to opment of prognostic research but also help to advertise and
capture condition monitoring data during the whole life of update their products. Both researchers and companies can
the system. The next reason is that, for some systems with benefit from the voluntary data share.
high requirement on safety, such as aircraft, aerospace
plane, and nuclear power equipment, they are not allowed (3) RUL prediction of machinery with complex degrada-
to operate under serious fault stages. We can only capture tion behaviors
partial degradation data for this kind of systems. Even if the Generally speaking, there are two different kinds of
whole-life data of several failed units are captured, it is prognostic strategies. The first strategy is to forecast the
unable to promise the similarity of the degradation pattern future degradation trend of machinery based on its historical
between the training units and the test unit. The degradation degradation trajectory. This strategy requires that the
behaviors of industrial systems are influenced by many machinery must share the same degradation pattern during
uncertainty resources, leading to the unit-to-unit variability the whole lifetime. Otherwise, the prediction result will
of the degradation pattern among different units. It highly deviate far from the actual curve. The second strategy is to
restricts the applicability and flexibility of training data- predict the RUL of an in-service unit using a prognostic
dependent prediction approaches. An effective strategy to model trained by a group of failure units. This strategy
deal with this issue is to predict the RUL of a system driven requires that the test unit must have the similar degradation
by its own condition monitoring data without depending on pattern with training units. These above two strategies both
training data from failure events, which is also defined as require that the degradation pattern of machinery is simple
self-data-driven RUL prediction [55]. To facilitate the self- enough to be derived from historical observations. We also
data-driven prognostic process, a model base involving find so many demonstration cases in literature with perfect
various degradation models needs to be prepared in gradual degradation trends. These demonstration cases are
advance. During the online prediction process, an optimal generally ideal cases selected from experimental degrada-
model is selected according to the degradation character- tion tests. In most industrial cases, however, the degradation
istics of the in-service unit, and the model parameters are behavior of machinery is complex or even irregular. The
updated according to the real-time data. The major chal- prognostic models will suffer from various strange degra-
lenges in self-data-driven prognostics may include: (1) how dation cases in real practice. It is a big challenge to keep the
to construct a diversified model base and (2) how to select a robust and stable performance of the prognostic models in
suitable model adaptively according to the degradation the RUL prediction of complex degradation behaviors in
pattern of the in-service unit. industrial practice.
(2) Dataset accumulation and publication
As mentioned above, whole-life degradation data are IV. OPPORTUNITIES AND
the basic resource for RUL prediction. It is time consuming CHALLENGES IN STATISTICAL
to accumulate degradation data under normal operational
conditions. Therefore, accelerated degradation tests are DATA-DRIVEN PROGNOSTICS
often employed to accumulate whole-life degradation A. A BRIEF INTRODUCTION ON
data. Some research institutes have published prognostic
STATISTICAL DATA-DRIVEN PROGNOSTICS
datasets on websites, which are free to download for
academic researchers. For example, the prognostic data In engineering practice, particularly lots of systems are
repository of NASA has collected and published many designed to perform particular missions and required to
accelerated degradation datasets including some typical operate safely during their whole life cycle. However, no
mechanical and electrical components and systems [65]. matter how reliable they are, the deterioration of their
In addition, some international societies such as the Prog- quality and performance due to aging, varying loads, and
nostics and Health Management (PHM) society and the operating environments will gradually impair them and
Institute of Electrical and Electronics Engineers (IEEE) finally result in their ultimate failures. Such systems are
reliability society often organize prognostic challenges in also known as the stochastic degrading systems. PHM has
international conferences, which provide valuable datasets emerged as an essential and efficient approach for improv-
and competition opportunities for researchers. Some scho- ing the operating safety and reducing the operational costs
lars [66] also voluntarily share their datasets to promote the for such stochastic degrading systems [67]. In the PHM

JDMD Vol. 2, No. 1, 2023

8 JDMD Editorial Office et al.

framework, prognostics with an emphasis on the RUL of the degradation variable or by the first hitting time
prediction have long been recognized as the fundamental of the degradation process characterized by the sto-
and key technique to implement the health management of chastic model. The difference between the solutions
stochastic degrading systems [68,28]. derived by the distribution of the degradation variable
Driven by the desire to ensure the safe and reliable or by the first hitting time of the degradation process
operation of stochastic degrading systems, especially for can be found in [69].
vital systems, various methods to achieve prognostics have In current studies on statistical data-driven prognostic
been developed. In general, prognostic methods for sto- approaches, the implementation processes of components
chastic degrading systems can be broadly divided into (2) and (3) are basically fixed or seldom changed. In
physics model-based methods, data-driven methods, and contrast, there are significant variants on stochastic models
their fusion. With advances in sensing and condition moni- for characterizing the degradation processes of systems, and
toring techniques, the monitoring data of the system’s thus more discussions are deserved. Despite many variants
performance degradation process can be more easily ob- on stochastic models used for degradation modeling, they
tained. As a result, the data-driven prognostic methods have can be generally described as XðtÞ = x0 þ gðt; θÞ þ εðtÞ,
become the emerging topic in the PHM field. In the past where XðtÞ is the degradation variable of the system
15 years, extensive efforts have been made to developing reflecting the degradation state at time t, x0 is the initial
various data-driven prognostics methods. Data-driven prog- degradation, gðt; θÞ is the time-dependent function with
nostic approaches mainly include ML and statistical data- parameter vector θ to model the time-varying trend of the
driven approaches [28,69]. Due to advantages in reflecting degradation process, and εðtÞ is the random term to model
the uncertainty and randomness of the degradation process the temporal uncertainty or randomness of the degradation
and providing the probability distribution of the RUL to process. According to the modeling principles for the
quantify the prediction uncertainty, significant advances degradation trend gðt; θÞ, statistical data-driven prognostic
have been witnessed in statistical data-driven prognostic approaches can be divided into being parametric, semi-
approaches since this kind of methods can provide a natural parametric, and nonparametric models-based methods.
description of the random failure of the practical system. Based on the functional form of gðt; θÞ, statistical data-
Therefore, we focus mainly on the statistical data-driven driven prognostic approaches include linear models-based
approaches in the following discussions. methods and nonlinear models-based methods, where linear
The basic idea of statistical data-driven prognostic models-based methods adopt linear models with time to
approaches for stochastic degrading systems is as follows: represent the degradation progression and nonlinear mod-
Based on the monitoring data of degrading systems, the els-based methods adopt nonlinear models with time.
RUL of the system can be predicted based on stochastic Besides the degradation trend modeling, modeling the
models by fitting the evolution law of the system perfor- random term εðtÞ is another important aspect in prognostics
mance degradation variable and extrapolating it to the since the degradation process of the system has the inherent
failure threshold. Generally, there are three key components randomness due to the impacts of various uncertain factors.
to achieve statistical data-driven prognostics, respectively, According to difference in modeling εðtÞ by stochastic
described as follows: processes, statistical data-driven prognostic approaches
(1) Stochastic degradation modeling. The performance mainly include random-effect regression models, Gamma
deterioration of degrading systems will be inevitable processes, inverse Gaussian processes, Wiener processes,
due to mutual effects of various random factors and recently developed beta processes, Tweedie exponen-
including aging, loads, and varying environments. tial dispersion process, Student-t processes, etc. The
The deterioration process is accumulated over the detailed discussions on these statistical data-driven prog-
operating time and will lead to the final failures of nostic approaches can be found in several comprehensive
these systems. Therefore, the degradation variable of review papers such as [28,70], and some technical papers
the system will randomly evolve during the system like [71–73].
operating process. As such, adopting stochastic mod-
els to characterize such randomly evolving process is B. DISCUSSIONS ON OPPORTUNITIES
a natural choice. AND CHALLENGES
(2) Parameter estimating of stochastic degradation mod-
els. Because the adopted stochastic models is selected It is observed from the above brief discussions that great
according to the statistical characters of the concerned advances have been made on statistical data-driven prog-
systems, the model parameters are unknown. In this nostic approaches and such methods are still in the stage of
case, to perform the prognostics, the model parame- fast development. Nevertheless, there are some new op-
ters of the used stochastic models should be first portunities and challenges required to be aware and ad-
estimated based on the monitoring degradation dressed in the future, as discussed in the following.
data. The widely used methods for parameter estima- (1) Prognostics via stochastic degradation model calibration
tion include the maximum likelihood estimation Statistical data-driven prognostic approaches generally
method, the Bayesian method, the expectation maxi- adopt the stochastic model to characterize the evolving
mum algorithm, etc. progression of the degradation variable. In existing studies
(3) Solving the probabilistic distribution of the RUL. with such methods for degradation modeling and RUL
Based on the stochastic degradation modeling and prediction, the appropriate functional form of gðt; θÞ should
the associated parameter estimating, to solve the be determined in advances [74]. Then, the model parame-
probabilistic distribution of the RUL is the key ters are estimated or updated by the degradation monitoring
task for prognostics. Generally, the probabilistic dis- data of the concerned system to perform the model calibra-
tribution of the RUL can be solved by the distribution tion. However, selecting the functional form of gðt; θÞ is

JDMD Vol. 2, No. 1, 2023

 Prognostics and Remaining Useful Life Prediction of Machinery 9

itself a challenging problem. More importantly, when the data into the RUL to remove the feature engineering. Nev-
selected functional form of the degradation model is inap- ertheless, the RUL corresponds to predicting the future
propriate, it is difficult and ineffective to calibrate the failure event and thus has the inherent uncertainty. Unlike
degradation model simply by updating the model parame- statistical data-driven prognostic approaches, DL-based
ters, and the prediction accuracy will be thus affected. prognostics methods have the limited capability in charac-
Hence, how to achieve simultaneous calibration of the terizing the prognosis uncertainty. As a result, fusing DL and
functional form and parameters of the degradation model statistical data-driven prognostic approaches will hold great
is an important direction holding promise to improve the promise to pave the way on prognostics for big data cases. A
prognosis accuracy and overcome the difficulty in selecting possible avenue is to apply DL techniques to extract the
the functional form of gðt; θÞ. degradation feature and then model the progression of such
(2) System-level prognostics with multiple degradation feature with stochastic models. To do so, the capability of
variables coupling quantifying prognosis uncertainty and handling the big data
can be jointly achieved. Therefore, fusion of DL and statis-
Although significant advances in statistical data-driven tical data-driven prognostic approaches is an important
prognostic approaches have been witnessed, most of these development direction, but the challenge lies in how to
studies are tailored to component-level prognostics in establish the effective and explainable fusion mechanism.
which an important potential prerequisite is that the health
state of the concerned component can be simply reflected by
a single performance degradation variable. The univariate V. PREDICTION OF RUL: FUTURE
hypothesis provides great convenience and flexibility for DIRECTIONS
the degradation process modeling and RUL prediction [75].
However, this may not be practical for the system-level A. INTRODUCTION
prognostics. As for a complex system, its health state is
often codetermined by multiple variables related to the The RUL of a component or system is the time left before it
system performance and the health state can rarely be will no longer be able to perform its intended function. The
exactly described by a single performance degradation task of predicting the RUL is called prognostics. For RUL
variable. In this case, considering multiple performance prognostics to be adopted in practice, the question of which
variables is a must in prognostics. However, compared with predictive models to use is fundamental. Specifically, the
widely studied component-level prognostics, the system- prediction capability of a prognostic model must be gauged
level prognostics are much more sophisticated. The primary with respect to the ability to provide trustable RUL pre-
challenges in system-level prognostics may include: dictions, which must possess the quality characteristics
(1) how to model the degradation process of the system required and the confidence level necessary for allowing
particularly for coupled multiple performance variables; to use them for taking decisions. Indeed, trust of the
(2) how to define the system failure in the multiple perfor- predictions heavily influences the decision makers’ attitude
mance variables case; and (3) how to estimate the model toward taking the risk of using the predicted RUL values to
parameters by the coupled degradation monitoring data of inform their decisions. The choice of which method to use is
multiple performance variables if indirect component-level typically driven by the data and/or the physics-based mod-
observations are utilized for system-level prognostics. All els available, taking into account the cost–benefit consid-
these aspects introduce the difficulty applying component- erations related to the implementation of the predictive
level prognostic methods to predict the RUL of complex system. A set of Prognostic Performance Indicators, mea-
systems. Thus, new framework for system-level RUL suring different predictive characteristics, must be used to
prediction should be developed in the future by considering guide the choice of the modeling approach.
the above challenges. Eventually, to arrive at taking reliability and safety
decisions based on RUL predictions in practice, it is neces-
(3) Fusion of DL and statistical data-driven prognostic sary to understand and quantify the impacts and benefits of
approaches the development of a predictive system, including avoiding
The success of statistical data-driven prognostic ap- unexpected catastrophic failures, reducing maintenance fre-
proaches is dependent heavily on the performance of the quency, optimizing spare parts and storage, optimizing
degradation feature or the degradation trend of the moni- resources, etc. Clearly, given the increasing complexity,
tored variable. With the good degradation feature, stochas- integration, and informatization of modern engineering com-
tic degradation models can be effectively constructed and ponent and systems, RUL prediction capabilities can no
output the prediction RUL in probabilistic distribution longer be isolated additions in support to maintenance but
forms to quantify the prognosis uncertainty. This is also must be closely linked to the other parts of the overall system
known as the major advantage of statistical data-driven (structural, power, electromechanical, information and com-
prognostic approaches. Recently, with great advances in munication technology, control). Then, such predictive capa-
Industry 4.0 and the IoTs, a large number of monitoring data bilities must be included since the beginning in the system
can be obtained providing abundant information on the conceptualization and design, to meet the overall operation
system’s health state and the RUL. However, statistical and performance requirements [77].
data-driven prognostic approaches are difficult to directly Up to date, the main development efforts in RUL
apply to the big data case unless the additional degradation prediction have been devoted to the hardware (e.g., IoTs,
feature engineering is introduced. Therefore, it is not sur- smart meters, etc.) and to the software for tracking the
prising to observe that DL-based prognostics methods have health state of monitored equipment (e.g., data analytics,
attracted much attention in the big data case [76]. By DL platforms for IoT interconnection and clouding for com-
techniques, the abstract degradation features can be auto- puting, etc.). On the other hand, the full deployment of
matically extracted for prognostics or the end-to-end prog- PHM in practice involves other aspects, including design
nostics can be achieved directly by mapping the original (e.g., the use of smart components may lead to different

JDMD Vol. 2, No. 1, 2023

10 JDMD Editorial Office et al.

reliability allocation solutions), and impacts various work test data distribution and the training data one, which is
units involved in maintenance decisions and actuations a quite common situation in practice
(e.g., workers can use smart systems, maintenance engi- • Bayesian neural networks and deep Gaussian processes
neers can analyze big data), including the supporting to provide the RUL predictions with estimates of their
logistics (spare parts availability and warehouse manage- uncertainty
ment can be driven by the RUL predictions) [78].
• Methods for obtaining transparency and interpretability
of RUL predictions for building trust on their use for
B. CHALLENGES decision making, especially in safety-critical applica-
tions. In particular, methods for injecting physical
A number of challenges still remain to be overcome to information in learning models (e.g., Physics-Informed
render effective the use of RUL predictions in practice. Neural Networks), post-hoc sensitivity approaches,
Some of these challenges are inherent in the complexity of and visualization techniques are being studied to pro-
the components and systems degradation processes, which vide interpretability from different perspectives,
are not fully known, are dynamic, and highly nonlinear; this including explaining the learned input-output relation
makes their understanding, characterization, and modeling representations, explaining the individual model out-
quite difficult. On the other hand, the data used to develop puts, and explaining the way the output is produced by
and calibrate the predictive models are collected in the the model.
field and are affected by inevitable limitations including
missing data and erroneous data from malfunctioning sen-
sors, scarcity, and incompleteness of data, often unlabeled Acknowledgement
with respect to the state of degradation of the component
or system, changing operational and environmental The work in Section III was supported by the National Science
conditions. Foundation of China (NSFC) (Nos. 52025056, 52005387), and the
In practical applications, the RUL predictions, and the work in Section IV was supported by the National Science
models that provide them, must satisfy a number of Foundation of China (NSFC) (Nos. 62233017, 62073336).
requirements to meet diverse objectives. Certainly, accu-
racy and precision are required to the level needed for
the decisions that they support: in some cases, very high References
accuracy and precision are necessary to take confident
decisions (e.g., of anticipating or postponing a scheduled [1] A. K. S. Jardine, D. Lin, and D. Banjevic, “A review on
maintenance based on accurate RUL predictions); in machinery diagnostics and prognostics implementing condi-
other cases, accuracy and precision need not be so high tion-based maintenance,” Mech. Syst. Signal Process.,
and may be compromised for other objectives. For exam- vol. 20, no. 7, pp. 1483–1510, 2006.
ple, transparency, explainability, and interpretability of [2] H. M. Elattar, H. K. Elminir, and A. M. Riad, “Prognostics: a
RUL predictive models can be of particular interest for literature review,” Complex Intell. Syst., vol. 2, no. 2,
decision making or are even demanded as a regulatory pp. 125–154, 2016.
prerequisite in safety-critical applications. Also, security [3] J. Guo, Z. Li, and M. Li, “A review on prognostics methods
issues regarding data integrity, data confidentiality, and for engineering systems,” IEEE Trans. Reliab., vol. 69, no. 3,
authentication exist. Finally, any RUL prediction must pp. 1110–1129, 2020.
be accompanied by an estimate of its uncertainty, in order [4] M. Kordestani, M. Saif, M. E. Orchard, R. Razavi-Far, and K.
to confidently take robust decisions based on such Khorasani, “Failure prognosis and applications—a survey of
prediction. recent literature,” IEEE Trans. Reliab., vol. 70, no. 2, pp. 728–
748, 2021.
[5] M. Daigle, B. Saha, and K. Goebel, “A comparison of filter-
C. FUTURE DEVELOPMENTS based approaches for model-based prognostics,” in 2012
Surely, DL will continue to be developed and used to allow IEEE Aerospace Conf., IEEE, Big Sky, MT, USA, 2012,
incorporating feature engineering in the process of predic- pp. 1–10.
tive model training, by automatic data processing and [6] Y. Lei, J. Lin, Z. He, and M. J. Zuo, “A review on empirical
feature extraction: whereas encouraging results have mode decomposition in fault diagnosis of rotating machin-
been obtained already in the application of DL for fault ery,” Mech. Syst. Signal Process., vol. 35, no. 1–2, pp. 108–
detection and diagnostics, RUL prediction still remains a bit 126, 2013.
of a challenge for DL in practice. [7] W. Yan, B. Zhang, G. Zhao, S. Tang, G. Niu, and X. Wang,
Other developments for the deployment of RUL pre- “A battery management system with a Lebesgue-sampling-
diction in practice include: based extended Kalman filter,” IEEE Trans. Ind. Electron.,
vol. 66, no. 4, pp. 3227–3236, 2018.
• RNNs for time series predictive analysis, combined
[8] C. Díaz, V. Quintero, A. Pérez, F. Jaramillo, C. Burgos-
also with data transformation into images so as to allow
Mellado, H.-a. Rozas, M. E. Orchard, D. Sáez, and R.
exploiting methods of image processing (includ-
Cárdenas, “Particle-filtering-based prognostics for the state
ing CNNs)
of maximum power available in lithium-ion batteries at
• Signal reconstruction methods (including autoenco- electromobility applications,” IEEE Trans. Veh. Technol.,
ders) of unsupervised and semi-supervised learning vol. 69, no. 7, pp. 7187–7200, 2020.
for degradation prediction in practical cases of unla- [9] M. E. Orchard and G. J. Vachtsevanos, “A particle-filtering
beled data approach for on-line fault diagnosis and failure prognosis,”
• Optimal Transport methods and unsupervised adapta- Trans. Inst. Meas. Control, vol. 31, no. 3–4, pp. 221–246,
tion techniques to cope with the difference between the 2009.

JDMD Vol. 2, No. 1, 2023

 Prognostics and Remaining Useful Life Prediction of Machinery 11

[10] Y. Qian and R. Yan, “Remaining useful life prediction dynamic environments,” IIE Trans., vol. 47, no. 5, pp. 471–
of rolling bearings using an enhanced particle filter,” 486, 2015.
IEEE Trans. Instrum. Meas., vol. 64, no. 10, pp. 2696–2707, [28] Y. Lei, N. Li, L. Guo, N. Li, T. Yan, and J. Lin, “Machinery
2015. health prognostics: a systematic review from data acquisition
[11] L. Liao and F. Köttig, “Review of hybrid prognostics ap- to RUL prediction,” Mech. Syst. Signal Process., vol. 104,
proaches for remaining useful life prediction of engineered pp. 799–834, 2018.
systems, and an application to battery life prediction,” IEEE [29] H. Wang, H. Liao, and X. Ma, “Remaining useful life
Trans. Reliab., vol. 63, no. 1, 191–207, 2014. prediction considering joint dependency of degradation
[12] L. Liao and F. Köttig, “A hybrid framework combining data- rate and variation on time-varying operating conditions,”
driven and model-based methods for system remaining useful IEEE Trans. Reliab., vol. 70, no. 2, pp. 761–774, 2020.
life prediction,” Appl. Soft Comput., vol. 44, pp. 191–199, [30] J. A. Flory, J. P. Kharoufeh, and N. Z. Gebraeel, “A switching
2016. diffusion model for lifetime estimation in randomly varying
[13] R. G. Nascimento, M. Corbetta, C. S. Kulkarni, and F. A. C. environments,” IIE Trans., vol. 46, no. 11, pp. 1227–1241,
Viana, “Hybrid physics-informed neural networks for lith- 2014.
ium-ion battery modeling and prognosis,” J. Power Sources, [31] K. Liu, N. Z. Gebraeel, and J. Shi. “A data-level fusion model
vol. 513, p. 230526, 2021. for developing com- posite health indices for degradation
[14] J. H. Cha and J. Mi, “Study of a stochastic failure model in modeling and prognostic analysis,” IEEE Trans. Autom. Sci.
a random environment,” J. Appl. Probab., vol. 44, no. 1, Eng., vol. 10, no. 3, pp. 652–664, 2013.
pp. 151–163, 2007. [32] N. Li, N. Gebraeel, Y. Lei, X. Fang, X. Cai, and T. Yan,
[15] D. R. Cox, “Regression models and life-tables,” J. R. “Remaining useful life prediction based on a multi-sensor
Stat. Soc. Ser. B (Methodol.), vol. 34, no. 2, pp. 187–220, data fusion model,” Reliab. Eng. Syst. Saf., vol. 208,
1972. p. 107249, 2021.
[16] A. K. S. Jardine, D. Banjevic, and V. Makis, “Optimal [33] N. Li, Y. Lei, N. Gebraeel, Z. Wang, X. Cai, P. Xu, and B.
replacement policy and the structure of soft- ware for condi- Wang, “Multi-sensor data-driven remaining useful life
tion-based maintenance,” J. Qual. Maint. Eng., vol. 3, no. 2, prediction of semi-observable systems,” IEEE Trans. Ind.
pp. 109–119, 1997. Electron., vol. 68, no. 11, pp. 11482–11491, 2020.
[17] M. L. T. Lee, G. A. Whitmore, F. Laden, J. E. Hart, and E. [34] C. Xiao, M. Yu, H. Wang, B. Zhang, and D. Wang, “Prog-
Garshick, “Assessing lung cancer risk in railroad workers nosis of electric scooter with intermittent faults: dual degra-
using a first hitting time regression model,” Environmetrics, dation processes approach,” IEEE Trans. Veh. Technol.,
vol. 15, no. 5, pp. 501–512, 2004. vol. 71, no. 2, pp. 1411–1425, 2022.
[18] K. A. Doksum and A. Hóyland, “Models for variable-stress [35] C. Wang, N. Lu, Y. Cheng, and B. Jiang, “A data-driven
accelerated life testing experiments based on wiener pro- aero-engine degradation prognostic strategy,” IEEE Trans.
cesses and the inverse Gaussian distribution,” Technometrics, Cybern., vol. 51, no. 3, pp. 1531–1541, 2019.
vol. 34, no. 1, 74–82, 1992. [36] K. Javed, R. Gouriveau, and N. Zerhouni, “A new multivari-
[19] J. P. Kharoufeh, “Explicit results for wear processes in a ate approach for prognostics based on extreme learning
Markovian environment,” Oper. Res. Lett., vol. 31, no. 3, machine and fuzzy clustering,” IEEE Trans. Cybern.,
pp. 237–244, 2003. vol. 45, no. 12, pp. 2626–2639, 2015.
[20] N. D. Singpurwalla, “Survival in dynamic environments,” [37] H. Zhou, J. Huang, and F. Lu, “Reduced kernel recursive
Stat. Sci., vol. 10, no. 1, 86–103, 1995. least squares algorithm for aero-engine degradation predic-
[21] J. M. Van Noortwijk, “A survey of the application of Gamma tion,” Mech. Syst. Signal Process., vol. 95, pp. 446–467,
processes in maintenance,” Reliab. Eng. Syst. Saf., vol. 94, 2017.
no. 1, pp. 2–21, 2009. [38] X. Liu, K. Yeo, and J. Kalagnanam, “A statistical modeling
[22] Y. Z. Zhang, R. Xiong, H. W. He, X. Qu, and M. Pecht, approach for spatio-temporal degradation data,” J. Qual.
“Aging characteristics- based health diagnosis and remaining Technol., vol. 50, no. 2, pp. 166–182, 2018.
useful life prognostics for lithium-ion batteries,” ETranspor- [39] G. Aydemir and K. Paynabar, “Image-based prognostics
tation, vol. 1, p. 100004, 2019. using deep learning approach,” IEEE Trans. Ind. Inf.,
[23] H. Rozas, D. Troncoso-Kurtovic, C. P. Ley, and M. E. vol. 16, no. 9, pp. 5956–5964, 2019.
Orchard, “Lithiumion battery state-of-latent-energy (sole): [40] X. Fang, K. Paynabar, and N. Gebraeel, “Image-based
a fresh new look to the problem of energy autonomy prog- prognostics using penalized tensor regression,” Techno-
nostics in storage systems,” J. Energy Stor., vol. 40, metrics, vol. 61, no. 3, pp. 369–384, 2019.
p. 102735, 2021. [41] Y. Dong, T. Xia, D. Wang, X. Fang, and L. Xi, “Infrared
[24] N. Gebraeel and J. Pan, “Prognostic degradation models for image stream based regressors for contactless machine prog-
computing and updating residual life distributions in a time- nostics,” Mech. Syst. Signal Process., vol. 154, p. 107592,
varying environment,” IEEE Trans. Reliab., vol. 57, no. 4, 2021.
pp. 539–550, 2008. [42] D. N. P. Murthy and D. G. Nguyen, “Study of a multi-
[25] L. Bian and N. Gebraeel, “Stochastic methodology for component system with failure interaction,” Eur. J. Oper.
prognostics under continuously varying environmental pro- Res., vol. 21, no. 3, pp. 330–338, 1985.
files,” Stat. Anal. Data Mining: ASA Data Sci. J., vol. 6, no. 3, [43] D. N. P. Murthy and R. J. Wilson, “Parameter estimation in
pp. 260–270, 2013. multi–component systems with failure interaction,” Appl.
[26] N. Li, N. Gebraeel, Y. Lei, L. Bian, and X. Si, “Remaining Stoch. Models Data Anal., vol. 10, no. 1, pp. 47–60, 1994.
useful life prediction of machinery under time-varying oper- [44] P. H. Kvam and E. A. Pena, “Estimating load-sharing
ating conditions based on a two-factor state-space model,” properties in a dynamic reliability system,” J. Am. Stat.
Reliab. Eng. Syst. Saf., vol. 186, pp. 88–100, 2019. Assoc., vol. 100, no. 469, pp. 262–272, 2005.
[27] L. Bian, N. Gebraeel, and J. P. Kharoufeh, “Degradation [45] R. P. Nicolai and R. Dekker, “Optimal maintenance of multi-
modeling for real-time estimation of residual lifetimes in component systems: a review,” in Complex System

JDMD Vol. 2, No. 1, 2023

12 JDMD Editorial Office et al.

Maintenance Handbook, Springer Series in Reliability Engi- [63] Y. Lei, N. Li, S. Gontarz, J. Lin, S. Radkowski, and J. Dybala,
neering, Springer, London, 2008, pp. 263–286. “A model-based method for remaining useful life prediction
[46] B. Rezaeianjouybari and Y. Shang, “Deep learning for of machinery,” IEEE Trans. Reliab., vol. 65, pp. 1314–1326,
prognostics and health management: state of the art, 2016.
challenges, and opportunities,” Measurement, vol. 163, [64] L. Guo, N. Li, F. Jia, Y. Lei, and J. Lin, “A recurrent neural
p. 107929, 2020. network based health indicator for remaining useful life
[47] P. Ding, M. Jia, and X. Zhao, “Meta deep learning based prediction of bearings,” Neurocomputing, vol. 240,
rotating machinery health prognostics toward few-shot prog- pp. 98–109, 2017.
nostics,” Appl. Soft Comput., vol. 104, p. 107211, 2021. [65] NASA, Prognostics Center of Excellence Data Set Repository.
[48] Y. Mo, L. Li, B. Huang, and X. Li, “Few-shot RUL estima- Available: https://ti.arc.nasa.gov/tech/dash/pcoe/prognostic-
tion based on model-agnostic meta-learning,” J. Intell. data-repository/.
Manuf., pp. 1–14, 2022, doi: 10.1007/s10845-022-01929-w. [66] Y. Lei, T. Han, B. Wang, N. Li, T. Yan, and J. Yang, “XJTU-
[49] B. Peters, M. Yildirim, N. Gebraeel, and K. Paynabar, SY rolling element bearing accelerated life test datasets: a
“Severity-based diagnosis for vehicular electric systems tutorial,” J. Mech. Eng., vol. 55, pp. 1–6, 2019.
with multiple, interacting fault modes,” Reliab. Eng. Syst. [67] E. Zio, “Prognostics and Health Management (PHM): where
Saf., vol. 195, p. 106605, 2020. are we and where do we (need to) go in theory and practice,”
[50] X. Cheng, K. Lv, G. Liu, and J. Qiu, “Intermittent fault Reliab. Eng. Syst. Saf., vol. 218, p. 108119, 2022.
modeling and RUL prediction for degraded electrical connec- [68] K. T. P. Nguyen, K. Medjaher, and D. Tran, “A review
tors in vibration environments,” IEEE Trans. Compon. Pack- of artificial intelligence methods for engineering prog-
aging Manuf. Technol., vol. 12, no. 5, pp. 769–777, 2022. nostics and health management with implementation guide-
[51] A. Correcher, E. Garcia, F. Morant, R. Blasco-Gimenez, and lines,” Artif. Intell. Rev., 2022, doi: 10.1007/s10462-022-
E. Quiles, “Diagnosis of intermittent fault dynamics,” in 2008 10260-y.
IEEE Int. Conf. Emerg. Technol. Factory Autom., IEEE, [69] X. S. Si, W. B. Wang, C. H. Hu, and D. H. Zhou, “Remaining
Hamburg, Germany, 2008, pp. 559–566. useful life estimation-A review on the statistical data driven
[52] A. Correcher, E. García, F. Morant, E. Quiles, and L. approaches,” Eur. J. Oper. Res., vol. 213, no. 1, pp. 1–14,
Rodríguez, “Intermittent failure dynamics characterization,” 2011.
IEEE Trans. Reliab., vol. 61, no. 3, pp. 649–658, 2012. [70] Z. X. Zhang, X. Si, C. Hu, and Y. G. Lei, “Degradation data
[53] A. N. Elmachtoub and P. Grigas, “Smart ‘predict, then analysis and remaining useful life estimation: a review on
optimize’,” Manag. Sci., vol. 68, no. 1, pp. 9–26, 2022. Wiener process-based methods,” Eur. J. Oper. Res., vol. 271,
[54] A. K. S. Jardine, D. M. Lin, and D. Banjevic, “A review on pp. 775–796, 2018.
machinery diagnostics and prognostics implementing condi- [71] M. Giorgio and G. Pulcini, “A new state-dependent degra-
tion-based maintenance,” Mech. Syst. Signal Process., dation process and related model misidentification pro-
vol. 20, pp. 1483–1510, 2006. blems,” Eur. J. Oper. Res., vol. 267, no. 3, pp. 1027–1038,
[55] N. Li, P. Xu, Y. Lei, X. Cai, and D. Kong, “A self-data-driven 2018.
method for remaining useful life prediction of wind turbines [72] W. A. Yan, X. F. Xu, D. Bigaud, and W. Q. Cao, “Optimal
considering continuously varying speeds,” Mech. Syst. Sig- design of step-stress accelerated degradation tests based on
nal Process., vol. 165, p. 108315, 2022. the Tweedie exponential dispersion process,” Reliab. Eng.
[56] X. S. Si, W. Wang, C. H. Hu, and D. H. Zhou, “Remaining Syst. Saf., vol. 230, p. 108917, 2022.
useful life estimation – a review on the statistical data driven [73] C. Y. Peng and Y. S. Cheng, “Student-t processes for
approaches,” Eur. J. Oper. Res., vol. 213, pp. 1–14, 2011. degradation analysis,” Technometrics, vol. 62, no. 2,
[57] N. Li, Y. Lei, T. Yan, N. Li, and T. Han, “A Wiener-process- pp. 223–235, 2020.
model-based method for remaining useful life prediction [74] X. S. Si, T. M. Li, J. X. Zhang, and Y. G. Lei, “Nonlinear
considering unit-to-unit variability,” IEEE Trans. Ind. Elec- degradation modeling and prognostics: a Box-Cox transfor-
tron., vol. 66, pp. 2092–2101, 2019. mation perspective,” Reliab. Eng. Syst. Saf., vol. 217,
[58] Y. Lei, N. Li, and J. Lin, “A new method based on stochastic p. 108120, 2022.
process models for machine remaining useful life prediction,” [75] T. Li, X. Si, H. Pei, and L. Sun, “Data-model interactive
IEEE Trans. Instrum. Meas., vol. 65, no. 2016, pp. 2671– prognosis for multi-sensor monitored stochastic degrading
2684, 2019. devices,” Mech. Syst. Signal Process., vol. 167, p. 108526,
[59] N. Z. Gebraeel, M. A. Lawley, R. Li, and J.K. Ryan, 2022.
“Residual-life distributions from component degradation [76] T. M. Li, X. S. Si, X. Liu, and H. Pei, “Data-model interactive
signals: a Bayesian approach,” IIE Trans., vol. 37, remaining useful life prediction technologies for stochastic
pp. 543–557, 2005. degrading devices with big data,” Acta Autom. Sin., vol. 48,
[60] X. S. Si, W. B. Wang, C. H. Hu, and D. H. Zhou, “Estimating no. 9, pp. 2119−2141, 2022.
remaining useful life with three-source variability in degradation [77] H. Badihi, Y. M. Zhang, B. Jiang, P. Pillay, and S. Rakheja,
modeling,” IEEE Trans. Reliab., vol. 63, pp. 167–190, 2014. “A comprehensive review on signal-based and model-based
[61] N. Li, N. Gebraeel, Y. Lei, L. Bian, and X. Si, “Remaining condition monitoring of wind turbines: fault diagnosis and
useful life prediction of machinery under time-varying oper- lifetime prognosis,” Proc. IEEE, vol. 110, no. 6, pp. 754–
ating conditions based on a two-factor state-space model,” 806, 2022.
Reliab. Eng. Syst. Saf., vol. 186, pp. 88–100, 2019. [78] Y. Hu, X. W. Miao, Y. Si, E. S. Pan, and E. Zio, “Prognostics
[62] L. Bian and N. Gebraeel, “Stochastic methodology for and health management: a review from the perspectives of
prognostics under continuously varying environmental pro- design, development and decision,” Reliab. Eng. Syst. Saf.,
files,” Stat. Anal. Data Mining, vol. 6, pp. 260–270, 2013. vol. 217, p. 108063, 2022.

JDMD Vol. 2, No. 1, 2023