Artificial Intelligence Review (2023) 56:3659–3709

https://doi.org/10.1007/s10462-022-10260-y

A review of artificial intelligence methods for engineering

prognostics and health management with implementation
guidelines

Khanh T. P. Nguyen1   · Kamal Medjaher1 · Do T. Tran2

Published online: 9 September 2022

© The Author(s), under exclusive licence to Springer Nature B.V. 2022

Abstract
The past decade has witnessed the adoption of artificial intelligence (AI) in various appli-
cations. It is of no exception in the area of prognostics and health management (PHM)
where the capacity of AI has been highlighted through numerous studies. In this paper, we
present a comprehensive review of AI-based solutions in engineering PHM. This review
serves as a guideline for researchers and practitioners with varying levels of experience
seeking to broaden their know-how about AI-based PHM. Specifically, we provide both a
broad quantitative analysis and a comprehensive qualitative examination of the roles of AI
in PHM. The quantitative analysis offers an insight into the research community’s inter-
est in AI-based approaches, focusing on the evolution of research trends and their devel-
opments in different PHM application areas. The qualitative survey gives a complete pic-
ture on the employment of AI in each stage of the PHM process, from data preparation to
decision support. Based on the strengths and weaknesses of existing methods, we derive
a general guideline for choosing proper techniques for each specific PHM task, aiming to
level up maintenance practitioners’ efficiency in implementing PHM solutions. Finally, the
review discusses challenges and future research directions in the development of autono-
mous intelligent PHM solutions.

Keywords  Artificial intelligence · Prognostics and health management · Condition

monitoring · Machine learning · Predictive maintenance · Decision-making support

* Khanh T. P. Nguyen
[email protected]; [email protected]
Kamal Medjaher
[email protected]
Do T. Tran
[email protected]
1
Production Engineering Laboratory, ENIT, Toulouse INP, 47 Avenue d’Azereix, Tarbes 65016,
France
2
Decision Science CoE, Tetra Pak, Ruben Rausings gata, Lund 22186, Sweden

13
Vol.:(0123456789)
 3660 K. T. P. Nguyen et al.

1 Introduction

During the last decade, prognostics and health management (PHM) has emerged in engi-
neering as a promising scientific discipline and has become an indispensable component
in product life cycles. The PHM procedure consists of seven following tasks: data acquisi-
tion, data processing, health indicator construction, fault detection, diagnostics, prognos-
tics, and decision-making (Gouriveau et al. 2016; Atamuradov et al. 2017). In other words,
PHM uses condition monitoring data and the information provided by operational practi-
tioners to assess systems health, detect potential anomalies, diagnose any incoming faults,
and predict the remaining useful life (RUL). The information and indications provided by
PHM are then used to schedule maintenance activities, which in turn maintain the system’s
healthy operation and thus ensure the overall system availability, reliability, and safety
(Gouriveau et al. 2016).
Along with the increasing industrial adoption of PHM, the engineering literature has
also seen many publications that review the diverse applications of PHM. Early reviews
include the work by Katipamula and Brambley (2005) that synthesized generic fault detec-
tion and diagnostics (FDD) with prognostics methods. Another study by Jardine et  al.
(2006) summarized the development of diagnostics and prognostics in mechanical sys-
tems within the framework of condition-based maintenance (CBM). Since the concept of
prognostics was not yet widely developed in the middle of the 2000s, both of the reviews
focused mainly on FDD methods. They provided methodological descriptions as well as
identified the strengths and weaknesses of FDD techniques.
The last years of the 2000s have witnessed the development of studies on prognostics
methods in the context of reliability theory and CBM. Those methods were classified into
three groups (Heng et al. 2009): conventional reliability, condition-based prognostics, and
integrated reliability and prognostics. From an industrial viewpoint, the authors in Sikor-
ska et  al. (2011) categorized the RUL estimation and prognostic modeling approaches
into knowledge-based, life expectancy, artificial neural networks, and physics models. The
abundant publications on the different prognostics methods developed in each of these cat-
egories have completed the picture of a conventional PHM process.
The first half of the 2010s has seen the emergence of data science and cloud computing
as new enabling paradigms which revolutionize many business and industrial areas includ-
ing predictive maintenance. These new domains have quickly established as mainstream
industries and become the new catalysts for technological advances in prognostics since
the second half of the 2010s. In line with the development of modern data-driven prog-
nostics methods, the authors in An et al. (2015) and Kan et al. (2015) proposed classify-
ing prognostics methodologies into three categories: (i) model-based or physics-based, (ii)
data-driven, and (iii) hybrid approaches. This new taxonomy has become more and more
popular, and can be generalized for PHM processes (Lee et  al. 2014; Atamuradov et  al.
2017; Javed et al. 2017; Saufi et al. 2019).
Also since mid-2010s, information and communications technology together with cloud
computing have undergone excellent development. The tangible achievements in those
areas quickly triggered the explosion of the Internet of things (IoT) and big data technol-
ogy. These new technological pillars in turn has brought about a new revolution in industry
and society, which is widely known as the fourth industrial revolution or Industry 4.0. With
the ability to gather and analyze data across clusters of connected machines, Industry 4.0
offers unprecedented promising perspectives on PHM, but at the same time it also intro-
duces real challenges to the development and application of modern PHM.

13
A review of artificial intelligence methods for engineering… 3661

Indeed, besides the added safety of the working environment, modern industries have
imposed an increased demand for production systems availability and reliability. One of
the levers to achieve this goal is to develop an autonomous system from the viewpoint of
systems health management. An autonomous PHM system requires continuous monitoring
of multiple connected devices via the IoT and cloud-powered database systems. Advanced
analytics and intelligent algorithms are then used to automatically exploit the acquired
data for RUL prediction. In addition to prognostics, decision making is another key com-
ponent of an autonomous PHM system. In order to enhance the overall system resilience,
advanced algorithms are employed to take autonomous real-time decisions on the basis of
prognostics results. To this end, it is recommended to replace conventional PHM systems
with artificial intelligence (AI) powered solutions.
AI has been demonstrated to increase the automation level not only in data processing
and analytics but also in condition monitoring and decision making. It is therefore neces-
sary to review the current state-of-the-art AI methods used in the whole PHM process and
pinpoint essential research directions. However, most of the surveys in the literature dedi-
cated to the application of AI in PHM have focused exclusively on fault diagnostics (Urai-
kul et al. 2007; Saufi et al. 2019). Moreover, they mainly concerned AI-based FDD meth-
ods in specific domains. Prominent examples include oil-immersed power transformers
(Sun et al. 2012), power distribution systems (Gururajapathy et al. 2017), fracture mechan-
ics (Nasiri et al. 2017), rotating machinery (Liu et al. 2018), building energy systems (Zhao
et al. 2019), and chemical process systems (Md et al. 2020).
Nevertheless, we have noticed that there are some lately published articles that review
the application of deep learning, a modern AI technology, in certain aspects of prognos-
tics (Khan and Yairi 2018; Hamadache et al. 2019; Fink et al. 2020). Virtually all of these
reviews examine AI-based methods for three principal PHM tasks, i.e., fault detection,
diagnostics, and prognostics, while skipping over other important tasks such as data prepa-
ration, health indicator construction, and decision-making support. Furthermore, existing
reviews, including the latest ones, performed the surveys that aim primarily to synthesize
the technical strengths and limitations of existing methods. They all miss a detailed and
in-depth quantitative analysis of the development trends of PHM methods in general and of
AI-based approaches in particular.
This review article attempts to fill the above-mentioned literature gap and would there-
fore fit perfectly the missing piece of the AI and PHM trans-domain puzzle. The paper
would also serve as an instructive guideline for academics and industrial practitioners with
varying levels of experience seeking to broaden their skills and knowledge about AI-based
PHM implementation. The major contributions of the paper are summarized in Fig.  1.
Unlike existing reviews which largely focus on FDD and prognostics, our work offers a
comprehensive view of AI-based methods for the whole PHM process from data acquisi-
tion and health indicator construction to decision-making support. Moreover, the exten-
sive and up-to-date quantitative analysis performed in this paper provides a global insight
into the interest of the PHM community in AI solutions, especially those belonging to the
artificial narrow intelligence branch which mainly involves machine learning, evolution-
ary algorithms, and expert or fuzzy inference systems. The theory of mind or self-aware
AI, which is promising and yet unexplored for PHM, is outside the scope of this review. In
addition to the quantitative analysis, we also provide a comprehensive qualitative analysis
of existing AI-based methods for each of the steps in the PHM implementation. This thor-
ough analysis per se forms a general guideline for maintenance practitioners on how to
properly employ AI-based solutions and how to choose suitable methods for accomplishing
a specific PHM task.

13
 3662 K. T. P. Nguyen et al.

Fig. 1  A graphical summary of the article’s contributions

Considering the article’s extensive coverage, in order to improve its readability we have
logically organized the article and provided a structured outline to help readers efficiently
navigate the parts of their interest. The rest of the paper is therefore organized as follows.
Section 2 introduces principal AI algorithms used in PHM and their role in each specific
PHM task. In Sect. 3, we quantitatively analyze the growth trend of AI-based PHM meth-
ods and their potential applications. Section 4 is dedicated to a comprehensive survey of
the deployment of AI-based PHM solutions, and accordingly provides a guideline on how
to choose proper AI techniques for a given PHM task. Finally, Sect. 5 concludes the paper
and discusses challenges and future research directions of intelligent PHM solutions.

2 Principal AI approaches and their applications in PHM

According to one of the ‘founding fathers’ of artificial intelligence (AI), John McCa-
rthy, AI is the science and engineering of making intelligent machines, especially intel-
ligent computer programs. Its primary goal, in theory, is to reach human cognitive abili-
ties, i.e., understanding, learning, and performing any intellectual tasks that a human being
can. However, the current development of AI only reaches a narrow level. This means that
based on specific software, AI is designed to accomplish only some particular tasks. There-
fore, AI algorithms and their classifications are usually investigated within the boundaries
of a specific discipline.
The raise of AI provides numerous opportunities for the development of PHM solutions.
Indeed, it offers various algorithms contributing to the success of all PHM tasks, from data
acquisition to decision making. An overview of the contributions of major AI approaches
to the PHM field is presented in Fig. 2. Details on how to apply these approaches to each
of the PHM tasks and why some of them are more preferable than others in a specific task
will be presented in Sect. 4.
The first class of approaches is unsupervised learning (UL), which is one of main pil-
lars of machine learning — a key AI discipline. UL aims to automatically draw pattern
characteristics of a dataset without pre-existing labels. Some UL techniques can effec-
tively contribute to the data processing task in the PHM process. For example, princi-
pal component analysis (PCA) and Isomap are among the most used techniques for data

13
A review of artificial intelligence methods for engineering… 3663

Fig. 2  Contribution of AI approaches to different PHM tasks

dimensionality reduction (Sorzano et  al. 2014), whereas auto-encoder–decoder network

(AED) is a useful tool to reduce signal noises (Hasani et al. 2017). These techniques also
often contribute to the success of the health indicator construction phase (Thieullen et al.
2012). In addition, clustering algorithms, such as kNN, t-SNE, and UMAP, which can clus-
ter observations into different categories, may also be used to find anomalies in a given
dataset for the fault detection purpose (He and Wang 2007; Md et al. 2020).
The second class of AI approaches and also the vital constituent of machine learning is
supervised learning (SL). SL is used to infer a mapping function between a pair of input
object and target value based on a fully labeled training dataset. The association of every
input observation and its attached target value is the main factor driving the operation of an
SL algorithm on how it interprets the data. Depending on the data interpretation manner,
SL algorithms can be divided into two categories:

• Classification is algorithms that predict the target class to which an observation of data
would belong. In the PHM field, these algorithms are usually applied to learn the pat-
tern characteristics in condition monitoring data, which are then used to identify health
states of a system. Classification algorithms have considerably contributed to the suc-
cess of automatic fault detection and diagnostics in recent years (Zhao et al. 2019; Sub-
ramaniyan et al. 2020).
• Regression is algorithms that aim to model the relationship between a given observa-
tion of data and its associated numerical target value. In PHM, regression algorithms
can be used to construct health indicators representing the system degradation evolution
over time (Thieullen et al. 2012). The major role of regression algorithms is however
prominently recorded through the excellent development of data-driven prognostics
techniques (Diez-Olivan et al. 2019). They allow for effectively predicting the remain-
ing useful life time of a given system based on its monitoring data.

The third class of AI approaches is semi-supervised learning (SSL), which is a mixed

approach between unsupervised learning with unlabeled training data and supervised learning
with labeled training data. During the training phase, SSL uses a small amount of labeled data
to draw more information about the pattern characteristics of the dataset for further elabora-
tion, e.g., clustering or classification. In PHM, the nominal system condition can be recorded
as labeled data, while it is nearly impossible to sample all types of failures. Consequently, fault

13
 3664 K. T. P. Nguyen et al.

detection and diagnostics can be investigated from the viewpoint of SSL algorithms (Yan et al.
2018). Among those, transductive learning, which aims to separate and identify the given
unlabeled data only, is suitable for early fault detections, whereas inductive learning, which is
dedicated to derive the correct mapping from input to output, is more appropriate for diagnos-
tics purposes.
Reinforcement learning (RL) is also a machine learning approach that aims to guide an
agent on how to perform the most suitable action in a given situation to maximize its total
cumulative reward. The agent’s decision making is trained based on its experiences, i.e., tri-
als, errors, and reward feedback from its iterative environment. Thanks to its flexibility and
capacity to take into account the stochastic behavior of the agent and the uncertain impact of
the environment, RL is a powerful tool for decision-making support in PHM, especially for
maintenance scheduling tasks (Kuhnle et al. 2019; Skordilis and Moghaddass 2020).
In addition, evolutionary and nature-inspired search and optimization algorithms are
another class of AI approaches, which may be recognized as intelligent search, intelligent opti-
mization, or evolutionary computing. They allow for deriving decision variable values in an
intelligent way to optimize the studied problem without any requirement of its exact math-
ematical formulation. These algorithms are thus able to deal with very complex problems that
normally cannot be solved by traditional approaches with reasonable computational resources.
Given their merits, these algorithms have widely been used in almost all PHM tasks. A pleth-
ora of examples are reported in the literature on optimization of monitoring/inspection poli-
cies (Korytkowski et al. 2009; Van Weele and Ramirez-Marquez 2012), feature selection (Cer-
rada et  al. 2015), health indicator construction (Nguyen and Medjaher 2021), optimization
of hyper-parameters in diagnostics and prognostics models (Saridakis et al. 2008; Jiang et al.
2016), and maintenance and production scheduling (Marseguerra et al. 2002; Yu et al. 2017).
Finally, it is necessary to mention fuzzy inference systems (FIS), which is a class of soft
computing approaches used in AI in over three decades. They use fuzzy sets and fuzzy logics
to simulate expert judgment into a comprehensible expression that can be analyzed and evalu-
ated for further elaboration. Built on top of flexible fuzzy rules, FIS can cope with uncertainty
and imprecision, and therefore may be able to efficiently interpret complex data without loos-
ing a significant amount of information. In PHM, FIS techniques have been widely applied
in numerous tasks, such as fault detection (Evsukoff and Gentil 2005), diagnostics (Viharos
and Kis 2015), prognostics (Mahdaoui and Mouss 2012), and maintenance decision making
(Baban et al. 2016).
In summary, the integration of AI algorithms in PHM allows for not only improving its
efficiency and robustness but also enhancing its autonomous capacity. Indeed, AI has enabled
intelligent health condition monitoring, as well as preemptive and automated fault detection,
diagnostics, and prognostics. Furthermore, it links the diagnostics and prognostics results to
automatically, suitably, and promptly making maintenance decisions to avoid imminent sys-
tems failure and minimize damage losses. The important role of AI in the development of
PHM techniques is therefore analyzed quantitatively in the next section.

3 Quantitative trend analysis of AI‑based methods in PHM

This section aims to provide an overview of research interests and potential practical appli-
cations of AI-based PHM methods through a broad quantitative analysis of published PHM
studies. Section  3.1 will start with presenting a literature research methodology and the
results will be discussed in Sect. 3.2.

13
A review of artificial intelligence methods for engineering… 3665

Fig. 3  Overview of the methodology used for the quantitative analysis

3.1 Quantitative analysis methodology

Figure  3 shows an overview of the quantitative analysis methodology used in this paper
to study the development trends in the PHM field, with an emphasis on solutions that are
based on AI methods.
To ensure the reliability of conclusions drawn from this critical review, we employed
the Web of Science (WoS) platform, which provides a variety of electronic databases for
many different academic disciplines, to acquire an abundant, comprehensive, and repre-
sentative list of relevant published materials. The detailed search string is presented in the
Annex. The literature searches were then refined by the areas where PHM solutions are
usually implemented. After combining the search results and removing all duplicates, we
obtained a list of 25,800 documents including journal articles, proceeding papers, and edi-
torial materials published in the period from 1967 until February 13, 2022. This list is
structured in a data-frame and saved in the TSV format that can be easily manipulated for
further exploitation and analysis. The data-frame consists of 25,800 rows, each represent-
ing a publication, and 67 columns capturing all relevant information of the material. Each
column is named by a two-letter field tag, for example, TI = document title, DE = authors
keywords, ID = keyword plus, AB = abstract, WC = WoS categories. More information can
be found on the WoS website 1.
To yield an insight into the published material list, we implemented a text data process-
ing procedure using Python’s regex and the well-known pandas package. The main steps of
this procedure are described as follows:

1. Check the suitability of the search results. A list of unique values of the research
categories with their occurrence frequency from the WC column was first extracted.
Then we digged into this list with unsuitable categories being identified and recorded in
a separate list. In addition, a similar list of inappropriate keywords from the DE and ID
columns were also created. These two particular lists were subsequently used to refine
the search results data in the next step.
2. Remove inappropriate materials. We removed all publications that belongs to the
unsuitable categories list identified in the previous step. The remaining data were then

1
  Link to the Web of Scien​ce page that descr​ibes the meani​ng of the 67 colum​ns

13
 3666 K. T. P. Nguyen et al.

filtered one more time by the list of inappropriate keywords. After this refinement, we
obtained a reliable data set consisting of 21,013 papers related to PHM.
3. Classify papers into ‘AI-related’ and ‘non-AI-related’ lists. Since this work focuses
on AI-based methods in PHM, it is essential to extract papers related to AI from the data
set refined in step 2. To do this, a list of AI-related keywords is cautiously created, com-
bining both the authors’ knowledge and the pertinent keywords extracted from the DE
and ID columns of the refined data set. Using this list, the database is then consecutively
filtered by the TI, DE, ID, WC, and AB columns. That means if one of the keywords
in the AI-related keywords list is found in the title, keywords, or abstract session, the
corresponding paper is extracted and saved to another data-frame named ‘AI-related’.
The remaining papers will then be classified into the ‘non-AI-related’ data-frame.
4. Data mining to extract useful information. Both ‘AI-related’ and ‘non-AI-related’
data-frames were exploited to draw critical remarks on the PHM research trend as well
as an overview of practical PHM applications, especially using AI-based techniques.
Details of the analysis are presented in next subsection.
Note that a detailed description on how to create the lists of unsuitable categories, inap-
propriate keywords, and AI-related keywords are presented in the Annex.

3.2 Discussion on the development of PHM and AI‑based solutions

In this section, the first part Sect.  3.2.1 provides an insight into the PHM research com-
munity’s interest in AI-based approaches. Then the developments of non-AI and AI-based
methods in different PHM application areas will be discussed in Sect. 3.2.2.

3.2.1 Overview of PHM researches

During more than five decades of development, researches related to PHM have appeared
in a wide range of scientific journals with good reputation in various disciplines. Figure 4
presents the top 10 journals that host most of the papers related to PHM. The percentage of
publications related to AI or non-AI method within each journal is also added to the top of
the bar representing the respective category. We notice that Mechanical Systems and Signal
Processing (MSSP) has published a large number of studies in PHM with more than 680
papers; the number of AI-related articles (51%) is higher than that of the non-AI category
(49%).
To have an insight into the interest of the research community on AI-based PHM, the
authors evaluate the normalized citation number (by publication time) of total papers pub-
lished in each year:
∑ni
TCij
(1)
j=1
ACi =
tc − i

where ACi and ni are the normalized citation number and the total number of papers pub-
lished in year i, respectively; TCij is the WoS times cited count of paper j published in year
i; and tc is the current time.
Considering Fig.  5, we can notice that the normalized citation rate of PHM papers
related to AI-based techniques has increased exponentially during one decade from 2010
to 2020 (over 20 times), whereas its non-AI counterpart has grown at a pretty lower rate
(only about 4 times) during the same period. It is worth noting that the decreased trend

13
A review of artificial intelligence methods for engineering… 3667

Fig. 4  Top relevant journals for PHM publications. The number atop each bar is the percentage of publica-
tions in the respective category of a discipline. MSSP: Mechanical Systems and Signal Processing, IEEE
Trans IE: IEEE Transactions on Industrial Electronics, IEEE Trans IM: IEEE Transactions on Instrumenta-
tion and Measurement, ASB: Applied Sciences-Basel, SAV: Shock And Vibration, ESA: Expert Systems
with Applications

Fig. 5  Normalized citation number (by publication time) of total papers published in every year ( ACi)

recognized since 2021 may likely be due to the short existence period of those articles pub-
lished from that year until now—they may not have enough exposure to attract the com-
munity’s attention which therefore results in a lower number of citations from subsequent
publications.
In order to give an overview of the evolution process of AI methods, we extracted
the most used AI-related keywords from PHM publications. Those keywords together
with their co-occurrence information are graphically visualized in Fig. 6 using the open-
source tool Gephi. We can observe that before 2000 not so many AI-related keywords

13
 3668 K. T. P. Nguyen et al.

Fig. 6  Evolution of AI-related keywords extracted from publications in PHM

were recorded. The most popular keywords of that time (Fig. 6a) show that studies dur-
ing this period were mostly related to the basics of genetic algorithms, machine learn-
ing, and neural networks. Then researches in the early 2000s (Fig. 6b) focused mainly
on the combination of statistical projection methods, such as principal component anal-
ysis or fisher discriminant analysis, and supervised learning methods, such as support
vector machines or decision trees, for fault diagnostics (Chiang et  al. 2004). We can
also recognize the initial PHM applications of unsupervised learning methods such as
clustering and self-organizing maps to fault detection. Besides, neural networks and its
hybrid variants such as neuro-fuzzy systems or ANFIS also started to attract the atten-
tion of researchers in PHM. The research trends of the period 2000–2007 continued to
be maintained and developed in the next period 2008–2014 as seen in Fig. 6c. In addi-
tion, emergent prognostics studies since the early 2010s, especially those using fuzzy
inference, Bayesian approaches, and neural networks, have also attracted significant
attention (Sikorska et  al. 2011). More recently, along with the successful application
and outstanding development of deep learning from 2015, a large number of deep neu-
ral network variants have been proposed in all aspects of PHM studies (Fig. 6d). This
important mega trend can be illustrated by a notable article (Jia et al. 2016) with more
than 890 citations.

13
A review of artificial intelligence methods for engineering… 3669

Fig. 7  Top popular application areas of PHM. The number on the top of each bar is the percentage of publi-
cations in the respective category within a discipline

3.2.2 Overview of practical PHM applications

When inspecting a large amount of literature related to PHM, one can easily notice an
extensive coverage of PHM applications across a wide range of engineering areas. Among
those, 14 disciplines that have the highest numbers of publications related to AI- and non-
AI-based methods are highlighted in Fig. 7.
We can realize the dominating number of publications related to applications of PHM
in electrical and electronic engineering. These articles have made up of around 38% of the
total number of publications in PHM. This is quite explainable given the fact that it is a
very important discipline which is connected to nearly all types of today’s machinery sys-
tems. We also notice a greater number of publications with AI methods compared to those
using non-AI methods in most of the popular PHM application areas thanks to the aston-
ishing development of deep learning in recent years.
Considering the growth of PHM in each area over time, Fig.  8 presents the develop-
ment trend of six major disciplines with PHM applications over the last three decades, in
which the vertical axis of subfigures indicates the cumulative number of publications. We
see that PHM applications in these six broadest engineering areas (Fig. 8a, b) have almost
exponential growths. In particular, PHM applications in electrical, electronic, and mechan-
ical engineering began to grow rapidly after 2000. In control engineering, the difference in
the publication numbers between AI- and non-AI-related PHM applications was marked in
2005, and since then the superiority of AI-based methods has been more recognizable over
time (Fig. 8a).
Figure 8b illustrates the development trends of PHM in multidisciplinary, energy, and
industrial engineering, in which PHM applications involving non-AI approaches was
recorded started in 1997 and those considering AI approaches was not accelerated until
ten years later. Multidisciplinary engineering exhibits an interesting pattern in which the
utilization of both AI- and non-AI-related techniques has been growing significantly since
2012; this might likely connect to the concept of cyber-physical systems becoming popular
at that same time. Another prominent trend worth mentioning in multidisciplinary engi-
neering is that despite started slowly AI-based solutions for multidisciplinary systems have

13
 3670 K. T. P. Nguyen et al.

(a) (b)

Fig. 8  Cumulative numbers of PHM publications in different application areas over time

been grown tremendously and showed their outstanding evolution when compared to the
non-AI category since 2019. This detail greatly aligns with our perception of the wide-
spread use of AI in every aspect of the modern life.

4 Qualitative analysis of AI‑based methods: guidelines for PHM

implementation

In previous sections, the key role of AI and its contribution to the development of the PHM
field have been highlighted through a broad quantitative analysis of the literature. An inter-
esting question that can be raised here is how AI-based PHM solutions are developed in the
literature and implemented in practice? To answer this question, a comprehensive survey of
the deployment of AI-based PHM solutions is presented from data acquisition (Sect. 4.1)
and high-frequency signal processing (Sect.  4.2) to decision support (Sect.  4.6). In par-
ticular, Sect. 4.3 is dedicated to the HI construction task, in which Sects. 4.3.1 and 4.3.2
will present two sub-tasks: feature selection and fusion for HI construction. Then Sect. 4.4
will provide details on fault detection and diagnostics (FDD) and Sect.  4.5 will discuss
failure prognostics. For principal PHM tasks (i.e. FDD and prognostics) and their crucial
sub-tasks (i.e. feature selection and fusion for HI construction), researchers have shown a
special attention and developed a variety of methods. We will therefore provide general
guidelines to help maintenance practitioners choose the appropriate ones. The reader will
find the guideline paragraph at the end of each method review associated to each task or
sub-task. For FDD methods, due to the huge number of methods available in the literature,
it is particularly hard to present the guideline in a single diagram. However, it is possible to
divide that diagram into separate parts corresponding to each group of methods and present
a guideline at the end of the subsection on each group.

4.1 Data acquisition and preprocessing prior to using AI

Data acquisition, including the collection and storage tasks, is essential for a better under-
standing of the system behavior and thus for developing appropriate PHM algorithms.
Industrial practitioners can consult the standard ISO 13374 which provides general guide-
lines for data processing, communication, and presentation in condition monitoring and
diagnostics of machines. In this section, we focus on the definition of a pertinent strategy

13
A review of artificial intelligence methods for engineering… 3671

Fig. 9  Overview of the recommended data collection scenario

for data acquisition in AI-related PHM projects. In most practical cases, AI-related PHM
projects are defined and launched independently of data acquisition, or once the data have
already been collected. This can subsequently lead to numerous difficulties and limitations
for data scientists when trying to select appropriate and representative data portions and
remove inconsistent data for a PHM purpose. Hence, we propose a recommended data col-
lection scenario as illustrated in Fig. 9 to ensure the obtained data are valuable and suitable
for building AI models.
In the recommended scenario, both domain experts and data scientists should be
involved in planning the condition monitoring process. Indeed, while the data scientists
investigate the operating and maintenance historical data to identify critical components
and important failure types, the experts can provide their valuable knowledge about the
system’s characteristics and also verify the useful information extracted by the data scien-
tists. Then, depending on the characteristics of the system and on the information available,
a proper definition of the monitoring process, such as the studied system, its monitoring
parameters and the appropriate sensors, can be defined.
Afterwards, to ensure the credibility, validity, and reliability of the collected data, the
monitoring process must be set up in a careful way, which attentively considers external
information such as system operating conditions, according to the expert advices. Finally,
in addition to CM data, other data sources including historical operating & maintenance
data, external information, and expert knowledge should be considered as a complete use-
ful dataset for the next steps of AI-based PHM implementation. However, as most human
knowledge is available only implicitly, it must be recorded in a machine-readable format
before exploring and exploiting it.
After data acquisition, it is essential to perform appropriate preprocessing techniques to
change raw data into a representation suitable for subsequent modeling tasks. These tech-
niques including the following steps:

1. The first step, i.e., data inspection, aims to detect incorrect, unexpected, and inconsistent
data. For this purpose, it is necessary to understand the underlying characteristics of a
data frame through summary statistics and data visualization methods.

13
 3672 K. T. P. Nguyen et al.

2. The second step, i.e., data cleaning, is an important step to correct and delete inaccurate
records. It consists of filtering noisy data, filling missing data, and removing irrelevant
data, duplicates and syntax errors.
3. The last step, i.e., data transformation, aims to transform the data in a format that is
ready to use and most convenient to a specific purpose. It includes standardizing, scal-
ing/normalizing, resampling/discretizing and dimension reduction.

For PHM purposes, the most commonly manipulated type of CM data is high-frequency
measurements such as vibration, current, and voltage signals. This format normally
requires particular treatment techniques to extract useful indicators for further studies.
More details will be presented in the next subsection.

4.2 High‑frequency signal processing

In the literature, high-frequency signal processing techniques can be classified into three
categories: time, frequency, and time–frequency domain (Nguyen et al. 2018). The proper-
ties of these techniques are summarized in Table 1.
Among the three categories, the time domain is classical, fast, and simple and therefore
can be generally applied for various systems to extract the properties of their raw measure-
ments, i.e., overall form and statistical values (Mahamad and Hiyama 2008; Shukla et al.
2015) for prognostic and diagnostic purposes. However, it is not viable for noisy signals
and requires pre-treatment techniques before further analysis.
In the second category, frequency domain analysis is effective to detect system anoma-
lies when knowing their fault characteristic frequencies. Specifically, fast Fourier Trans-
form (FFT) is widely used to decompose physical signals into number of discrete frequen-
cies or spectrum of continuous frequencies. After performing FFT, the anomalies can be
detected by high magnitude values at the characteristic frequencies (Yu 2012). Besides, the
envelope analysis is widely used to extract useful diagnostic indicators (Cong et al. 2012;
Leite et  al. 2015). In addition, power spectrum density (PSD) evaluation, that describes
the spectral energy distribution into its frequency components, is also an effective method
proposed in numerous diagnostics and prognostics studies (Shukla et al. 2015; Picot et al.
2014).
The last one, which is the time–frequency domain, includes the most powerful tech-
niques for analyzing and characterizing the non-stationary signals and therefore can be
viable for noisy signals. However, it requires high computation time and also abundant
experiences to choose appropriate hyper-parameters and to get useful information from the
results. Specifically, the short time Fourier transform (STFT) first decomposes the signals
into a set of data within a fixed window length and secondly performs the FT on every data
window. Then the spectrum magnitude at FFT characteristic frequencies is used for fault
detection and diagnostics (Yazici and Kliman 1999). However, the selection of the fixed
window length before performing STFT can strongly affect the method’s performance.
Therefore, the wavelet transform based on a variable window length is recommended for
diagnostics and prognostics studies (Deekshit  Kompella et  al. 2018; Tobon-Mejia et  al.
2012). However, the performance of wavelet analysis method strictly depends on the choice
of the mother wavelet. It is therefore suitable to use Hilbert–Huang Transform (HHT) for
extracting diagnostic or prognostic indicators when we do not have prior information about
the signal shape (Refaat et al. 2013; Elbouchikhi et al. 2017).

13
A review of artificial intelligence methods for engineering… 3673

Table 1  Summary of high-frequency signal processing techniques to extract useful features (EEMD:

Ensemble Empirical Mode Decomposition)
Method Strength Weakness

Time domain Classical, fast, and simple. Can Not viable for noisy signals. May
  Overall form generally be applied to different require pre-treatment techniques
  Statistical value fault types of various systems before evaluation of indicators
Frequency domain Effective techniques for detect- Require information about fault
  Fast Fourier transform ing system anomalies when characteristic frequencies. Are
  Envelope analysis knowing the fault characteristic not broadly applicable
  Power spectrum density frequencies
Time–frequency domain Powerful techniques for analyzing High computation time. Require
  Short time Fourier transf. and characterizing the signal experience to choose proper
  Wavelet transform spectrum in time. Are viable for hyperparameters and extract use-
  Hilbert–Huang transform noisy signals ful information from the results

4.3 Health indicator construction

The construction of health indicators is generally based on the engineering of features

in the data. A brief review of feature engineering for predictive maintenance can be
consulted in Aremu et al. (2020). This process involves (1) selecting relevant features in
the data, and (2) transforming or combining them to generate new powerful indicators
which are then used for system health management.

4.3.1 Feature selection

Feature selection (FS) aims to identify a subset of features that allows for reducing the
negative effects of irrelevant features and therefore provides a better prediction result
(Chandrashekar and Sahin 2014). Based on their evaluation criteria, FS techniques can
be classified into three groups: (1) filter methods, (2) wrapper methods, and (3) embed-
ded methods (Khaire and Dhanalakshmi 2019).
In the first group, filter methods first rank features based on certain statistical meas-
ures and then eliminate inappropriate features. An analysis and comparison of multiple
features for fault detection and prognostics in ball bearings was presented in Nguyen
et  al. (2018). However, it is not trivial to identify the number of selected features  for
filter methods. Furthermore, in practical applications, important features might have
low ranks according to certain evaluation criteria but could be more informative when
combined with others for a specific learning purpose (Chandrashekar and Sahin 2014).
Therefore, the second group, wrapper methods, proposes to incorporate a specific learn-
ing algorithm in the process of feature subset selection (Kohavi and John 1997). In
Khlaief et  al. (2019), the authors proposed using a genetic algorithm (GA) to find the
optimal feature subset for bearing defect diagnostics. Although wrapper methods pro-
vide better results compared to filter approaches, they are very computationally expen-
sive to implement, especially for data with a large number of features (Tang et al. 2014).
Unlike wrapper approaches that only use learning algorithms for testing the quality of
selected feature subsets, embedded methods directly integrate the optimization of fea-
ture subsets during the learning process. This integration allows for avoiding training
of the model each time a new feature subset is examined. Performance of embedded

13
 3674 K. T. P. Nguyen et al.

methods was demonstrated through numerous diagnostics and prognostics studies

(Zhang et al. 2018).
Based on the strengths and weaknesses of feature selection methods presented in dif-
ferent review papers (Khaire and Dhanalakshmi 2019), a general guideline for select-
ing appropriate feature selection methods for PHM according to data characteristics is
derived and presented in Fig.  10. In particular, filter approaches are suitable for unla-
beled data or when there does not exist a correlation between the features and labeled
data. Otherwise, embedded approaches are the best candidate for the cases that require
high accuracy and inexpensive computation. However, they are not suitable for high-
dimensional data, which can be overcome by using wrapper methods based on heuristic
search algorithms.

4.3.2 Feature fusion

A brief review of health indicator construction techniques for PHM purposes can be
consulted in Nguyen and Medjaher (2021). These techniques can be classified into four
categories as summarized in Table 2.
The first group involves statistical projection methods which are more appropriate for
diagnostics (Yu 2012) than for prognostics (Mosallam et al. 2016). It aims to project obser-
vations from a high dimensional space to a lower dimensional space. However, new fea-
tures created by these statistical projection methods are not interpretable, which may lead
to a deal-breaker in some settings.
In the second group, mathematical model based techniques focus on developing math-
ematical formula that allow for capturing the relations between CM measurements of the
system and its degradation process. The HI is also developed based on the expert’s knowl-
edge about the physical behavior of the system (Pan et al. 2018) and about the relevant fea-
ture used for creating effective HIs (Atamuradov et al. 2018). However, the above studies
are based on assumptions about degradation forms over time or the expertise knowledge
about signal processing techniques, data analysis, and system behaviors. Then, they might
not be suitable for complex systems where an automatic process is preferable.
The third group, including deep learning (DL) methods, provides alternative solutions
for automatically extracting and constructing useful information without requiring much
expert knowledge in the case of abundant data. For example, health indicators for system
prognostics can be developed based on long short-term memory (LSTM) encoder–decoder
(Malhotra et  al. 2016), recurrent neural networks (RNN) encoder–decoder (Gugulothu
et  al. 2017), or convolution neural networks (CNN) (Han et  al. 2019). From these stud-
ies, we can see that the deep learning approaches can take advantage of abundant data to
automatically generate health indicators without using much expert knowledge about the
system. Nevertheless, deep features created by these models are difficult to understand and
cannot be directly interpreted as the physical characteristics of the system.
The final group that is based on genetic programing aims to automatically find the best
mathematical formulation that combines low-level features to form more abstract high-
level prognostic features (Firpi and Vachtsevanos 2008; Liao 2014). This method is flexible
in formatting mathematical functions and allows for easily integrating the expert’s knowl-
edge about the HI formulations by defining an appropriate initial population. Furthermore,
the created HI, which is an explicit mathematical function of low-level features, can be
interpreted for further studies (Nguyen and Medjaher 2021).

13
A review of artificial intelligence methods for engineering… 3675

NO
Labeled data?

YES

Correlation NO
between Independent NO
features and data features
labels?

YES
YES

NO Filter methods based

High number Filter methods based on Pearson
of features? mutual information redundancy, Fisher
between features ratio, intern-extern
distance ratio, etc.
NO
YES
High computational
capacity?
Wrapper methods
based on heuristic
search algorithms
YES
(GA, PSO, etc)

Wrapper methods
Embedded
based on sequential
methods
search algorithms

Fig. 10  A guideline for choosing feature selection methods

Based on analyzing the advantages and limitations of the above-mentioned approaches,

a general guideline for selecting HI construction methods is presented in Fig.  11. With-
out expert knowledge, approaches that are based on statistical projection are good candi-
dates for constructing HI from high-dimensional non-abundant data, whereas deep learning
based approaches are more suitable for the cases of high-dimensional and abundant data.
Besides, GP-based approaches are promising solutions, which can be extended to integrate
possibly available expert knowledge. Finally, mathematical model based approaches are
the best alternative solution for low-dimensional data with available expert knowledge
about the system and/or its degradation process.
Discussion on HI performance metrics
In addition to choosing the appropriate method for HI construction, evaluating the suit-
ability of the constructed HIs is essential to the success of a PHM solution. In general,
the HI metrics can be classified into three categories: (1) metrics for evaluating a single
HI pattern (e.g., monotonicity, robustness, trendability, and identifiability), (2) metrics for
evaluating consistency of multiple HIs (e.g., consistency), and (3) hybrid metrics. For a
brief review of those metrics, we recommend referencing (Lei et  al. 2018; Nguyen and
Medjaher 2021).

4.4 Anomaly detection and diagnostics

In the literature, fault detection and diagnostics topics are usually reviewed and discussed
together (Abid et al. 2021). Table 3 summarizes the recent reviews on AI-based FDD meth-
ods. Inspired by these reviews, we propose classifying the AI-based FDD methods into

13
 3676 K. T. P. Nguyen et al.

four groups as illustrated in Fig.  12): statistical learning, neural networks based, genetic
algorithms based, and hybrid AI methods.

4.4.1 Statistical learning methods

With the ability to overcome some uncertainties in the data due to the use of statistics, sta-
tistical learning methods are widely used in modern FDD systems.
Principle component analysis (PCA) is one of the most popular statistical algorithms
that are implemented for fault detection and diagnostics in different fields, such as build-
ing energy systems (Zhao et  al. 2019), chemical process systems (Md et  al. 2020), and
nuclear power plants (Das and Parhi 2009). It uses singular value decomposition (SVD)
to transform correlated variables into a set of orthogonal vectors that are ordered by the
amount of variance. Then, a fault can be detected when the squared prediction error (SPE)
value exceeds a specific threshold (Chen and Lan 2009) while fault diagnostics can be per-
formed by using Q-contribution plots (Xiao et al. 2006). In Wang et al. (2009), the authors
improve the PCA performance for data having multivariate outliers by using winsorization
procedure with an M-estimator based on the T2-distribution. Du et al. in Du et al. (2017)
propose using a proper PCA model based on the combination of two statistic values, T2
and SPE, to isolate the faults of monitoring-type and controlling-type sensors in vapor
compression systems. To overcome the linear properties of PCA, the kernel PCA-based
fault detection algorithm is developed in Mansouri et  al. (2016). In a recent study (Luo
et al. 2017), a novel PCA method, called sparse PCA or SpPCA, is proposed to character-
ize meaningful variable correlations by sequentially extracting a set of sparse loading vec-
tors from process data.
Partial least squares (PLS) based methods are multivariate statistical techniques that
allow for measuring covariation between variables and creating a new fewest dimension
set of variables having maximal covariances. Similar to PCA, the PLS-based methods use
Table 2  Summary of transformation techniques for constructing health indicators reported in the literature
Method Strength Weakness

Statistical projection (Yu 2012; Adequate for cases that have Manually extract features before
Benkedjouh et al. 2013; Zhang various sensor sources. Do not projection. Only interest in data
et al. 2018) require expert knowledge about variance or class separability. HIs
the system cannot be interpretable
Mathematical model based (Ata- Well capture the characteristics Require expertise knowledge about
muradov et al. 2018; Sun et al. of a specific system. HIs can be the system and its degradation
2019; Pan et al. 2018) interpretable trend. Manually extract and
combine features. Not widely
applicable to various systems,
especially complex ones
Deep learning based (Malhotra Automatic process. Do not HIs may not be interpretable. Hard
et al. 2016; Gugulothu et al. require expert knowledge about to integrate expert knowledge
2017; Han et al. 2019) the system. Widely applicable if available. High computation
to various systems, especially time. Require abundant data and
with an abundence of data experience for tuning DL hyper-
parameters
GP based (Firpi and Vachtseva- Applicable to various systems. May have high computation time.
nos 2008; Liao 2014; Nguyen Do not require expert knowl- Require experience to tune GP
and Medjaher 2021) edge but allow for integra- hyperparameters.
tion if available. HIs may be
interpretable

13
A review of artificial intelligence methods for engineering… 3677

NO
High
dimensionality?

YES

NO Ability of
Available expert
integrating expert
knowledge?
knowledge?

YES
YES
NO Abundant
data?
GP based Mathematical model
approaches based approaches
YES

Statistical projection Deep learning based

based approaches approaches

Fig. 11  A guideline for selecting health indicator construction methods

statistic indexes for fault detection in various industrial systems (Zhang et  al. 2012). In
the study (Lee et al. 2004), the authors proposed combining the signed digraph and PLS
to improve the performance of multi-fault diagnostics in the Tennessee Eastman Process
(TEP). Godoy et  al. presented an extension of PLS-decomposition that includes the dis-
tance to the regression models for detecting and diagnosing different anomalies in produc-
tion systems (Godoy et al. 2013). Another extension of PLS presented in Shan et al. (2015)
is based on the idea of using slice transformation to handle nonlinear calibration. Its perfor-
mance is evaluated on two benchmark datasets and highlighted to be more attractive than
other nonlinear PS models.
Independent component analysis (ICA) based methods belong to blind source sep-
aration techniques that allow for revealing hidden information in non-Gaussian data. In
general, the ICA-based methods use statistical indices for fault detection in industrial pro-
cesses (Jiang and Yan 2013). To overcome the nonlinear properties of continuous anneal-
ing process, Zhang et al. (2013) introduced kernel ICA model. The idea of using kernel is
also applied to the modified dynamic ICA-based FDD framework in Fan and Wang (2014).
In recent studies, the authors in Li et al. (2017) proposed combining variable division and
independent component analysis while in Zhu et  al. (2017), a combination multi-kernel
ICA and adaptive rank-order morphological filter was used for TEP fault detection and
diagnostics.
Fisher discriminant analysis (FDA) based methods belong to projection methods
that allow integrating the data label information in finding informative projection vectors.
These vectors aim to maximize the separation between classes and minimize the separa-
tion within each class. The FDA based methods are widely applied for fault detection and
diagnostics in different studies (He et  al. 2005; Cho 2007). Recently, the authors in Ge
et al. (2016) introduced kernel-driven semi-supervised fisher discriminant analysis model
for nonlinear fault classification when missing a part of fault labels. Feng et al. improved

13
 3678 K. T. P. Nguyen et al.

FDA model performance by using local and global manifold learning and kernel method to
overcome the disturbances, non-Gaussian and nonlinearity aspects of the TEP (Feng et al.
2016).
Linear discriminant analysis (LDA) based methods, a generalization of FDA, aim
to find new projection directions by linearly combining existing features. These directions
help reduce the data dimensionality by maximizing the ratio of between-class scatter and
within-class scatter. LDA-based methods are usually applied for fault classification in bear-
ings (Hamadache et al. 2019). Harmouche et al. proposed using LDA to improve the clas-
sification of faults in bearing balls based on their spectral features (Harmouche et al. 2014).
The study in Jin et al. (2014) introduces a trace ratio linear discriminant analysis to over-
come the challenges of high-dimensional non-Gaussian fault data, e.g. single-point and
generalized-roughness faults in motor bearing faults.
Support vector machines (SVM) are a discriminative classifier that aims to find a
separating hyperplane formally defined by a number of support vectors. In fault detection
applications, the hyperplane allows for determining whether the monitoring data belong to
the fault class or not. For instance, the authors in Kulkarni et al. (2005) introduced an SVM
model with knowledge incorporation to detect faults in TEP. Multi-classifier SVM methods
have also been developed for classifying different types of faults in transmission systems
(Parikh et al. 2010) and chiller systems (Li et al. 2016). In addition, different hybrid SVM
models are proposed to firstly detect whether a fault occurs or not, and finally to classify
its corresponding type (Liu et  al. 2016; Dehestani et  al. 2011). To overcome the nonlin-
earity of data, different kernel functions, such as Gaussian, radial basis function (RBF),
polynomial, and sigmoidal functions, are implemented to improve the SVM performance
in fault detection (Yin et al. 2014) and also in fault diagnostics (Han et al. 2011). Another
extension of kernel trick based SVM developed in de Souza et al. (2014), called support
vector clustering, allows supervised learning from high-dimensional data with unlabeled
training samples. Furthermore, authors in Zhao et al. (2013) proposed using support vector
regression (SVR) derived from SVM for the detection and diagnosis of centrifugal chiller
faults at low severity levels. Tran et al. in Tran et al. (2016) developed the above study by
combing the nonlinear least squares SVR and the exponentially weighted moving average
control charts for improving the performance of chiller FDD strategy.
kNN based methods use the minimum distance criterion and the majority of votes
amongst the k neighbor points to determine the class of the observed point. As kNN algo-
rithm is very simple and easy to interpret, it is widely used for FDD in different systems
(García-Pedrajas and Ortiz-Boyer 2009), such as semiconductor manufacturing process
(He and Wang 2007), bearings (Lei et al. 2009), and induction motors (Nguyen and Lee
2010). To overcome the nonlinear, multimode, and non-Gaussian properties of industrial
processes, Zhou et al. in Zhou et al. (2016) propose an isolation index provided by decom-
posing the kNN distance in the original measurement space without correlation for fault
isolation.
Decision trees (DT) based methods use a flow-chart-like tree structure for fault clas-
sification. In FDD applications, each internal node denotes an observation feature while
each terminal node represents a class label, i.e., fault type. In Sun et al. (2007), the authors
proposed using a DT model based on the features reduced by PCA for fault diagnosis of
rotating machinery. To overcome the characteristic of burst and extremely short duration of
wireless mesh networks faults, the authors in Li et al. (2010) introduced a rule post-prun-
ing method based on the C4.5 algorithm, which is a traditional DT. Recently, Madhusu-
dana et al. presented a new decision tree algorithm (J48) using discrete wavelet features for
fault diagnostics of face milling tool. Besides, to limit over-fitting as well as error due to

13
A review of artificial intelligence methods for engineering… 3679

Table 3  Recent reviews on AI-based fault detection and diagnostics methods

Paper Method Application area

Gururajapathy et al. (2017) AI techniques include ANN, Power distribution

SVM, fuzzy logic, GA, and
matching approach
Nasiri et al. (2017) Bayesian networks, ANN, genetic Mechanics
algorithms, fuzzy logic, case-
based reasoning, and hybrid
methods
Liu et al. (2018) kNN, naive Bayes, SVM, ANN, Rotating machinery
deep learning
Zhao et al. (2019) Classification-based, unsuper- Building energy
vised learning, regression-
based, and knowledge driven-
based methods
Md et al. (2020) Multivariate statistical analysis, Chemical process
machine learning approaches,
and data-driven hybridization
methods

bias, the random forest (RF) algorithm was developed by aggregating many decision trees.
Thanks to its ability to handle large data sets with multiple input features and as well as
missing data, RF is powerful for fault diagnostics in aircraft engines (Yan 2006), induction
motors (Yang et al. 2008), and also in chemical processes TEP (Aldrich and Auret 2010).
Bayesian networks based methods Bayesian networks (BN) are probabilistic graphi-
cal models that allow for integrating historical data and expert knowledge to overcome the
uncertain, incomplete, and even conflicting properties of the data. In FDD applications,
it graphically presents the causes and symptoms of the faults and quantitatively evaluates
their probabilities of occurrence. It has been widely applied to fault detection and diag-
nostics of various industrial systems, such as gas turbine engines (Bickford and Malloy
2002), automobiles (Schwall and Gerdes 2002), chemical processes (Verron et al. 2006),
gear train systems (Kang et al. 2007), assembly processes (Jin et al. 2012), and building air
systems (Xiao et al. 2014). For instance, Zhao et al. (2017) developed a generic BN-based
method that allows for identifying and sensitively analyzing most of the common faults in
air handling units.
A general guideline for selecting statistical learning methods
Inspired by the strengths and the weaknesses of individual statistical learning meth-
ods presented in different reviews (Nasiri et al. 2017; Gururajapathy et al. 2017; Md et al.
2020; Zhao et al. 2019), we propose a general guideline to select the appropriate statistical
learning method for an FDD application as illustrated in Fig. 13. In short, Bayesian net-
works based methods allow for combining data with domain knowledge and representing
the causal relationship between variables. For unlabeled and high-dimensional data, statis-
tical projection methods such as PCA, PLS, ICA, and their variants are the most commonly
used ones. Finally, for labeled data, classifiers such as SVM, FDA, and LDA effectively
characterize linear discriminant patterns in data while kNN, DT, RF, and kernel-based
methods are usually used as alternative solutions for the opposite case.

13
 3680 K. T. P. Nguyen et al.

AI based FDD
methods

Statistical learning Neural network Genetic algorithm

Hybrid AI methods
methods based methods based methods

Fuzzy- Series Embedded Parallel

PCA PLS ICA FDA LDA SVM KNN DT Bayesian ANN DL
Neuron combination combination combination

Fig. 12  Classification of AI-based fault detection and diagnostics methods

4.4.2 Neural networks based methods

Neural networks (NN) based methods are computing systems inspired by biological neural
networks. In general, NN-based FDD methods can be classified into three principal groups:
artificial neural networks (ANN), deep learning (DL) and neuro-fuzzy systems.
Artificial neural networks (ANN) which belong to the supervised multi-classification
category are graphical models with input, output, and hidden nodes. For FDD applications,
input nodes usually capture sensor measurements; hidden nodes often represent extracted
features; and output nodes generally characterize fault types. In an early study, the authors
in Ganguli et al. (1998) used ANN for detection and quantification of both single and mul-
tiple faults on a helicopter rotor system. The performance of ANN for rolling bearing ele-
ments FDD is proved through various studies when using different features extracted from
vibration and current signals, such as frequency features (Li et  al. 2000), time features
(Samanta and Al-balushi 2003), time–frequency features (Pandya et  al. 2012), and the
combination of features from all three domains (Hariharan and Srinivasan 2009). Besides,
ANN is also widely used for structural damage detection and diagnostics, for example loca-
tion and severity quantification of the cracks in beam-like structures (Das and Parhi 2009).
In addition, various studies have implemented ANN for the identification and location of
multiple faults in energy transmission and distribution systems (Javadian et al. 2009). In a
recent study (Wang et al. 2020), a noise-based wavelet packet analysis ANN method has
been developed for engine fault diagnostics.
Deep learning (DL) Fig. 14 presents the major groups of deep learning methods that
are often employed in automated fault detection and diagnostics systems.
Among the DL techniques, deep neural networks (DNN) are a powerful tools built on
stacking multiple NN layers in hierarchical architectures (Khamis et al. 2016). It allows for
progressively extracting useful features from the raw input for FDD without the expert and
the priori knowledge (Fink et al. 2020). In fact, the authors in Yan and Yu (2015) proved
that the deep features extracted by a DNN better capture the complex relations among all
sensor measurements and the underlying system behaviors than handcrafted features, and
allow for improving the accuracy and robustness of the anomaly detection for gas turbine
combustors. For highly noisy data, the auto-encoder–decoder (AED) technique, which is
developed based on feed-forward DNN, is a promising candidate to learn and extract fea-
tures for intelligent diagnostics in various studies (Jia et al. 2016; Lee et al. 2019).
Besides, convolutional neural networks (CNN) are proven to be a powerful technique for
feature learning from very high-dimensional data, especially for image or video data. Spe-
cifically, CNN is a feed-forward neural network of multiple hidden layers, including mul-
tiple convolutional layers, pooling layers, fully connected layers, and normalization layers.
These layers allow for looking for low-level features from raw data, and then building up

13
A review of artificial intelligence methods for engineering… 3681

to more abstract characteristics of system health states for fault detection and diagnostics.
For instance, in Faghih-Roohi et al. (2016) a deep CNN solution was introduced to analyze
a huge amount of image data obtained from hours of video recordings for the automated
detection of rail surface defects. In addition to image data, CNN is also used to capture
several different small fault patterns of the front-end speed controlled wind generator from
vibration data (Dong et al. 2016).
In addition to CNN, recurrent neural networks (RNN) are also widely used for FDD
thanks to its ability to hold past information through cyclic connections in the hidden
units. For example, the studies in Abed et al. (2015); Sharma et al. (2015) proposed using
dynamic RNN for bearing fault detection and diagnostics using discrete wavelet transfor-
mation of vibration and stator current signals. Furthermore, to overcome the vanishing gra-
dient problem during back-propagation in RNN models, long short-term memory (LSTM)
techniques have been developed and implemented in different FDD systems. In Park et al.
(2018), the authors showed the best performance of an LSTM-based fault detection model
for an industrial robot manipulator when comparing with different architectures of neu-
ral networks, random forest, and support vector machines. A recent study in Dong (2019)
implemented a novel deep learning (combining LSTM and CNN networks) and transfer
learning (TL) technique for developing a comprehensive scheme for actuator/sensor fault
and icing detection/identification in the aircraft.
Besides deep neural networks, deep belief networks (DBN) are also employed as a pow-
erful condition monitoring tool because they can benefit from unlabeled data in pre-train-
ing phase using restricted Boltzmann machines and require a little amount of labeled data
in the fine-tune phase using feed-forward networks. Hence, DBN are proposed for condi-
tion monitoring and fault diagnostics of rolling bearing elements (Shao et  al. 2015). In
Yin and Zhao (2016), the authors proposed an automated diagnosis network of vehicle on-
board equipment for high-speed trains via a DBN trained on the basis of Restricted Boltz-
mann Machine. Recently, a DBN model and its relevant parameters optimization selection
strategy for fault diagnostics of a variable refrigerant flow air-conditioning system were
investigated in Guo et al. (2018).
Neuro-fuzzy systems are the results of combining ANN and fuzzy logics. This combi-
nation overcomes the gray-box limitation of ANN, thanks to fuzzy knowledge reasoning,
allowing for adjusting the parameters of the fuzzy structure based on the learning ability of
ANN even without human expertise. Hence, neuro-fuzzy systems have been widely used
in fault detection and diagnostics since the 2000s (Viharos and Kis 2015). For example,
in Evsukoff and Gentil (2005) the authors presented an application of recurrent neuro-
fuzzy systems for fault detection and isolation in nuclear reactors. The study in Zhou et al.
(2009) introduced fuzzy models for quantifying the residuals of key chiller performance
indices under various fault conditions, and then used these indices to train artificial neural
network models for chiller fault diagnostics. Similarly, Kocyigit (2015) proposed using a
fuzzy inference system (FIS) for sensor fault detection, and then injecting these outputs to
ANN for classifying different fault types of a vapor compression refrigeration system. In
Ye et  al. (2006), an adaptive-network-based fuzzy inference system (ANFIS) was devel-
oped for mechanical fault diagnostics of induction motors with wavelet packet decomposi-
tion features. This model is one of the most popular neuro-fuzzy architectures and has been
widely implemented for FDD of various systems, such as induction motors (Soualhi et al.
2019), continuous stirred tank reactors (Sabura Banu and Uma 2011), and industrial steam
turbine (Salahshoor et al. 2011).
A general guideline for selecting NN-based methods in FDD systems

13
 3682 K. T. P. Nguyen et al.

Combination of
data with domain
NO knowledge

YES

Labeled data?
NO
Casual relationship
YES between variables?

NO NO YES
Linear discriminant Time-varying
pattern properties?

Time-varying
properties?
NO
Missing data YES Gaussian NO
and complex relation distributed latent
between YES variables
variable? NO
High dimensional YES
NO
data?
YES YES

YES
PCA, PLS for linear data ICA for linear data; Dynamic
Kernel SVM, Dynamic PCA, Bayesian
DT, Kernel PCA, Space PCA Kernel ICA for non- Bayesian
Kernel FDA, SVM FDA, LDA Dynamic ICA network
RF for non-linear data linear data network
KNN

Fig. 13  A guideline for selecting fault detection and diagnostics methods based on statistical learning

A general guideline (Fig. 15) is derived from the strengths and the weaknesses of the
NN-based methods mentioned above for selecting an appropriate technique according
application characteristics. In summary, NN-based methods are usually used when the
expert and prior knowledge are unavailable, i.e., particular characteristics of the data are
unknown. Among different NN methods, AED techniques are the most appropriate for
unlabeled and highly noisy data. Besides, the performance of deep neural networks (such
as CNN, RNN, and LSTM) is normally impressive for very large datasets. Contrarily, when
the data is not abundant, neuro-fuzzy systems can be considered as one of the best candi-
dates for FDD of highly nonlinear, complex, and dynamic systems.

4.4.3 Genetic algorithms based methods

Genetic algorithms (GAs) are a heuristic search method inspired by natural evolution. It is
usually applied for identification and location of structural or mechanical damages (Vakil-
Baghmisheh et  al. 2008; Buezas et  al. 2011) and faults in transmission lines in various
studies (Li et  al. 2012; Jamil et  al. 2015). For instance, in Chou and Ghaboussi (2001)
GA was used to minimize the unmeasured displacements or the measured and correspond-
ing computed displacement for the localization of damages in structural engineering. The
authors in Vakil-Baghmisheh et al. (2008) used both binary and continuous genetic algo-
rithms to find the optimal location and depth by minimizing the cost function obtained
through a crack analytical model. The performance of continuous genetic algorithms com-
pared to the binary counterparts was also highlighted in Bedekar et  al. (2011) for iden-
tification of faulty parts in power systems. The authors in Buezas et  al. (2011) extended
the simulation of a crack model from 2D to 3D and used GA as an optimization method
in crack detection of engineering parts such as wind turbine blades. Mehrjoo et al. (2013)
used GA to detect multiple cracks depth and location on beam-like structures. To overcome
the challenge of fault localization in a large-scale system, the study in Jin and Ju (2012)

13
A review of artificial intelligence methods for engineering… 3683

Fig. 14  Major deep learning techniques used for fault detection and diagnostics

proposed dividing the distribution system into a main branch and individual regions, then
GA was used to find faults for the main branch and independent regions.

4.4.4 Hybrid AI methods

Since each AI-based FDD method has its own strengths and weaknesses, the construc-
tion of hybrid frameworks would allow for overcoming the limitations, and inheriting the
advantages, of individual approaches. This section focuses only on hybrid methods that are
developed by combining different AI techniques to meet a particular situation unsolvable
by an individual tool. These combinations can be classified into three main groups: serial,
embedded, and parallel combinations as shown in Table 4.
Table 4 summarizes different studies developing hybrid AI methods for FDD. In the first
group, serial combinations, the output of a method is used as the input to another. The most
common combination principle in this group is integrating dimensional reduction or fea-
ture selection methods into classifiers to improve the accuracy of fault diagnostics (Zhang
2009; Jiang et al. 2015; Gharahbagheri et al. 2017; Wang et al. 2018). The second group
concerns embedding an optimization method into a classifier or clustering method to find
its best parameters. One of the most popular combinations is the use of a genetic algorithm
to optimize the structure of neural networks (Pan et al. 2010; Unal et al. 2014). Finally, the
third group, parallel combinations, aims to fuse the outputs from different classifiers or
clustering methods to enhance the performance of FDD systems. The decision fusion can
be performed by majority voting (Zhang 2006), weighted average (Lei et al. 2007, 2010),
Bayesian voting (Seng Ng and Srinivasan 2010), or using the Dempster–Shafer theory
(Zhang and Ge 2015)

4.4.4.1  Discussion on  FDD performance metrics A large numbers of FDD performance
metrics have been developed in the literature. They can be classified into two categories:
temporal and static (Frank et al. 2019). The first category evaluates the responsiveness of
FDD algorithms to a time-varying error signal, e.g., time to detect, time to diagnose, detec-
tion stability factor, and diagnostic stability factor. Those metrics are suitable for applica-
tions in which the rapid response is a key requirement. Static metrics in the second category
quantify the time-independent performance of an FDD algorithm. Most of them are bor-
rowed from the performance assessment of classification and clustering algorithms devel-
oped in the field of computer science. Specifically, one can cite popular metrics such as false
positive or negative rate, true positive or negative rate, accuracy, F1-score, and area under

13
 3684 K. T. P. Nguyen et al.

Without the
expert Unknown data Neural network
and the prior YES characteristics YES based methods
knowledge?

YES

Labeled data?
NO

YES

NO
Abundant data?
High noisy data?

YES

NO YES
Highly
non-linear, complex and
dynamic systems
Very-high- Sequential
dimensional data? data?

YES YES YES

Neuro-fuzzy systems, Auto-Encoder-

ANN CNN RNN, LSTM
e.g. ANFIS Decoder

Fig. 15  A guideline for selecting neural networks based fault detection and diagnostics methods

curve. In addition, one can cite some particular metrics that have been proposed for specific
FDD purposes such as Correct Diagnosis Rate, Misdiagnosis Rate, Misdiagnosis Cost, or
Unable to Diagnose Fraction (Frank et al. 2019).

4.5 Failure prognostics

In the literature, see Table 5, several review papers have proposed classifying prognostics
methods into three main groups: model-based, data-driven, and hybrid methods (Kan et al.
2015; Javed et al. 2017; Atamuradov et al. 2017; Lei et al. 2018). A few articles focus on
shallow learning and deep learning models (Ellefsen et al. 2019; Hamadache et al. 2019).
In general, there is a lack of critical overviews focusing only on AI-based prognostics,
which is a specific part of the data-driven group but is also a parent category containing
deep learning models. Consequently, inspired by the relevant work this section provides
a comprehensive review of AI-based methods and proposes classifying them into four
groups as illustrated in Fig. 16: statistical learning methods (Sect. 4.5.1), neural networks
based methods (Sect. 4.5.2), fuzzy systems (Sect. 4.5.3), and hybrid methods (Sect. 4.5.4).
Based on the strengths and limitations of each approach, a general guideline for selecting
the appropriate methods according to the application’s characteristics is derived and dis-
cussed at the end of Sect. 4.5.4.

4.5.1 Statistical learning methods

One of the most widely used statistical learning methods in prognostics is the Bayesian
group that allows for handling uncertainty when estimating the system health state or its

13
A review of artificial intelligence methods for engineering… 3685

RUL based on the system’s prior knowledge. The paper (Gebraeel et  al. 2005) applied
Bayesian methods using real-time condition monitoring information to update stochastic
parameters of degradation models and then evaluated the residual-life distribution of the
monitored system. A dynamic Bayesian network and its inference algorithms were devel-
oped in Dong and Yang (2008) to predict the RUL of the drill-bits of a vertical drilling
machine. Another variant of Bayesian networks is the Bayesian hierarchical model pro-
posed in Zaidan et al. (2015), which combines two sources of information: historical in-
service data and monitoring performance measurements for probabilistic estimation of
the RUL of civil aerospace gas turbine engines. In addition, the Bayesian framework was
proved as an effective method to handle numerous uncertainty sources including unit-to-
unit variation, time uncertainty, and stochastic correlation (Jin et al. 2013).
Besides Bayesian methods, the support vector machine method is also widely used
in engineering prognostics due to its ability to avoid over-fitting and decent performance
on large datasets (Costello et  al. 2017). Numerous variants of SVM have been proposed
to handle different nonstationary and nonlinear problems. For example, to deal with
structural changes in the data, a dynamic version of SVM was proposed in Cao and Gu
(2002) for incorporating domain knowledge and non-stationary time series. In Chen et al.
(2013), the authors proposed a multivariate SVM, which allows for considering the influ-
ences among various variables and effectively excavating the potential information of small
samples; whereas the authors in Benkedjouh et al. (2013) combined the isometric feature
mapping reduction technique and support vector regression to predict the RUL of rolling
bearings. However, the identification of the kernel function without a general theoretical
guidance can be considered as the main drawback hindering the wide adoption of SVM in
prognostics.

4.5.2 Neural networks based methods

According to the development of monitoring techniques, artificial neural networks

(ANN), that allow for finding complex relationships between input and output variables
without prior knowledge, are widely used in prognostics. The early developed model,
named feed-forward neural networks (FFNN), can derive the characteristics of the deg-
radation evolution or the information of RUL from the monitoring data. This ability was
proved in numerous studies (Mahamad et al. 2010; Xiao et al. 2017), which use vibration
signals to predict bearing failures. The paper (Gope et al. 2015) also proved the effective-
ness of ANN in predicting the crack initiation and growth direction under multiple crack
geometries using crack position parameters. In addition, a variant of ANN that allows for
dealing with the suspension data, in which equipment are taken out of service before they
fail, was also developed in Tian et al. (2010) to improve the accuracy of RUL prediction.
Besides FFNN models, radial basis function networks (RBFN) which allows for mod-
eling complex mappings thanks to its ability to deal with nonlinear properties can be a
promising tool for prognostics. Indeed, the paper (Riverol and Pilipovik 2014) showed its
performance for assessing the failure frequency of a milk pasteurization unit. However, it is
not trivial to determine the RBFN activation function parameters, and optimize the number
of hidden neurons and the receptor size. These challenges can be considered as the draw-
backs that limit the wide application of RBFN to prognostics studies.
Although both of the two above-mentioned networks can capture the complex nonlinear
relationship between inputs and outputs, they cannot model the sequentiality of the data.
Hence, the recurrent neural networks (RNN), that allows for storing the previous inputs

13
 3686

13
Table 4  Hybrid AI methods for fault detection and diagnostics of industrial systems grouped by how the combination is established
Hybridization Methods Description

Serial combination KPCA, KICA, and SVM (Zhang 2009) KPCA for the Gaussian part and KICA for the non-Gaussian part to improve performance of the
SVM classifier for fault diagnostics
CVA and FDA (Jiang et al. 2015) CVA for handling the serial correlations to improve the performance of FDA for fault diagnostics
KPCA and BN (Gharahbagheri et al. 2017) KPCA for sensor fault detection, BN for integrating process knowledge into fault diagnostics
RF, BBN, and ARIMA (Wang et al. 2018) RF for selection of significant predictor, BBN for dependency among the predictors, ARIMA for
temporal causation
Embedded param- ANN and GA (Saridakis et al. 2008) GA to produce values for the cracks attributes as input arguments to ANN for cracks identifica-
eter optimization tion
ANN and GA (Pan et al. 2010; Unal et al. 2014) GA to optimize the structure parameters of ANN for structural damage identification and bearing
fault diagnostics
SVM and GS (Gao and Hou 2016) GS to optimize hyper-parameters of SVM classifier for TEP fault diagnostics
Parallel combination Multiple NNs (Zhang 2006) Modified majority voting for combination of the NN outputs in on-line process fault diagnostics
Multiple ANFIS and GA (Lei et al. 2007) GAs for evaluating the weights of the combination of the ANFIS outputs
Multi-agent systems (Seng Ng and Srinivasan 2010) Voting-based fusion and Bayesian fusion of multi-agent outputs for chemical process FDD
MLP, RBF, kNN and GA (Lei et al. 2010) GAs for evaluating the weights of the combination of different classifier outputs (MLP, RBF,
kNN)
Decision fusion systems (Zhang and Ge 2015) Dempster–Shafer theory based fusion of different outputs of FDD methods for TEP
SVM, C4.5, and kNN (Askarian et al. 2016) SVM, C4.5, and kNN produce the highest robustness and accuracy for TEP diagnostics
Ensemble SVMs (Zheng et al. 2017) Ensemble SVMs based on features extracted using composite multi-scale fuzzy entropy for the
intelligent fault classification

CVA: Canonical Variate Analysis, KLNPDA: Kernel Local–Nonlocal Preserving Discriminant Analysis, PNN: Probabilistic Neural Network, MLP: Multi-layer Perceptron
Neural Network, RBF: Radial Basis Function, GS: Grid Search, BDE: Binary Differential Evolution, RF: Random Forest, BBN: Bayesian Belief Network
K. T. P. Nguyen et al.
A review of artificial intelligence methods for engineering… 3687

via self-loops and backward connections between all neurons in the network, was proven as
a powerful approach for long-term predicting of failure progression using time series data
(Malhi et al. 2011; Guo et al. 2017). However, the gradient vanishing and exploding issues
make the training of RNN difficult. Then, gated-recurrent unit (GRU), which is a vari-
ant of RNN, was developed to overcome these issues. It allows for keeping features from
previous activations for a long time and using multiple bounded nonlinearities to reduce
the likelihood of vanishing gradient for back-propagation procedure. For prognostics pur-
pose, the authors in Zhong et  al. (2018) proposed using GRU to address the time series
and nonlinear characteristics of the parameters of exhaust gas temperature while the paper
(Chen et al. 2019) proved the performance of GRU in terms of RUL prediction of turbofan
engines. Nevertheless, as GRU works with only a reset gate and an update gate, it does
not separate the input from the forget gate; therefore exposes the entire memory content to
other units in the network. Long short-term memory (LSTM), that has the ability to add
or remove information to the memory content via separated input, forget, and output gates,
is an alternative solution for effectively tracking long-term dependencies. LSTM with its
numerous variants are widely used in prognostics. In Wu et al. (2018), Nguyen and Med-
jaher (2019), Nguyen et al. (2022), the LSTM performance for RUL estimation was proven
through different case studies such as turbofan engine, milling machine, mechanical sys-
tems, pulverizer mill, and aero-propulsion engines. Then, its variant, named bi-directional
LSTM, was developed to learn the dependencies of time series data from past and future
states simultaneously. This model was successfully applied in prognostics to discover the
underlying patterns from operational condition information and sensor signals for accurate
prediction of system RUL (Huang et al. 2019). In recent studies, Miao et al. (2019) devel-
oped a dual-task deep long short-term memory network for joint learning of degradation
assessment and remaining useful life prediction. Besides, the study (Nguyen and Medjaher
2019) used LSTM for prediction of the failure probability in different time intervals instead
of a particular RUL to enable the predictive maintenance planning for aeroengines.
Besides RNN, convolutional neural networks, which allow for dealing with high
dimensional raw input data without prior knowledge thanks to their spatially shared weights
and spatial pooling, are widely used in prognostics. Unlike the normal CNN structure in
computer vision applications, the convolution and pooling filters of CNN when deployed in
prognostics field are applied along the temporal dimension over the multi-channel sensor
data (Sateesh Babu et al. 2016). They allow for learning higher-level abstract representa-
tion from raw signals to improve the accuracy of RUL prediction (Li et  al. 2018). In a
recent study (Wen et al. 2019), the authors apply the residual block to skip several convolu-
tional layers through shortcut connections, and use the k-fold ensemble method to enhance
the CNN performance. Furthermore, several studies have proposed combining CNN with
LSTM to benefit their own advantages in the simultaneous extraction of both spatial and
temporal features (Al-Dulaimi et al. 2019; Li et al. 2019) for a better RUL prediction of
turbofan engines.
Different from above models (ANN, RBFN, RNN, GRU, LSTM, CNN) that are trained
using an supervised learning algorithm, deep belief networks (DBN) which consist of sev-
eral layers of restricted Boltzmann machines (RBM) can be considered as a semi-super-
vised learning approach. Indeed, thanks to RBM, a DBN can learn without labels to proba-
bilistically reconstruct its input data (Liao et al. 2016). After this feature extraction step,
it can be further trained using supervised learning algorithm to perform classification or
regression tasks. For an illustration, the paper (Deutsch and He 2018) proposed using DBN
for automatically learning features from vibration raw data, and then predicting the bear-
ing RUL using feed-forward neural networks. Besides DBN, several unsupervised learning

13
 3688 K. T. P. Nguyen et al.

neural networks such as self-organizing map (SOM) and auto-encoder (AE) networks
are also combined with supervised methods to enhance the prognostic performance. On the
one hand, the SOM, which implements an orderly mapping for data dimensionality reduc-
tion, can facilitate the assessment of system degradation states without a priori knowledge
of abnormal patterns (Huang et al. 2007). On the other hand, the AE, which learns efficient
data coding in an unsupervised manner, allows for extracting features from higher dimen-
sions of data and representing them in lower dimensions. It was combined with LSTM in
Malhotra et al. (2016) to reconstruct the time-series index characterizing the healthy state
of a system for RUL estimation. The performance of this combination was also highlighted
through different applications such as renewable energy power (Gensler et al. 2016), turbo-
fan engines, and milling machines (Yu et al. 2019).

4.5.3 Fuzzy and neuro‑fuzzy systems

Fuzzy systems are based on computational intelligence that uses a collection of mem-
bership functions and rules to represent values, and to implement inference mechanism,
respectively. In Biagetti and Sciubba (2004), the authors presented a critical and analytical
description of a research project that aimed to implement an expert system capable of con-
dition monitoring, diagnostics and prognostics based on expert knowledge and fuzzy rules.
The study (Zio and Di Maio 2010) derived a prognostics procedure based on a fuzzy point-
wise similarity concept for an on-line estimation of the system RUL.
Furthermore, fuzzy logic was also combined with neural networks to develop various
effective tools for prognostics purposes, named neuro-fuzzy systems (NFS). Beneficing the
advantages of neural networks and fuzzy logic, the NFS has automatic learning ability, lin-
guistic modeling capacity, and parallel processing facility. Their performances were proved
for RUL estimation of different applications such as cutting tools (Chinnam and Baruah
2004), gears (Wang et  al. 2004), boiler tubes (Majidian and Saidi 2007), and bearings
(Zhao et al. 2009). In addition, the study (Chen et al. 2013) developed a prognostics pro-
cedure using a-priori knowledge-based adaptive neuro-fuzzy inference system (ANFIS)

Table 5  Recent review articles on AI-based fault prognostics methods

Paper Proposed classification Methods

Kan et al. (2015) Model-based, data-driven based, ANN, fuzzy logic, neuro-fuzzy
and combined models systems, and SVM
Atamuradov et al. (2017) Model-based, data-driven, and ANN, fuzzy logic, SOM, SVM, and
hybrid methods Bayesian networks
Javed et al. (2017) Physics-based, data-driven, and ANN, neuro-fuzzy, Bayesian,
hybrid prognostics instance-based learning, and
combined methods
Leite et al. (2018) Knowledge-based, ANN, life ANN, ANFIS, and NARX
expectancy, and physical
models
Lei et al. (2018) Physics-based, statistics-based, ANN, neuro-fuzzy systems, SVM/
AI approaches, and hybrid RVM, and GPR
approaches
Ellefsen et al. (2019) Data-driven, machine learning, Deep belief networks, auto-
model-based, and hybrid encoder, LSTM, and CNN
Hamadache et al. (2019) Shallow and deep learning Statistical learning, ANN, and
combined methods

13
A review of artificial intelligence methods for engineering… 3689

AI based prognostics
approaches

Statistical Neural network

Fuzzy logic Hybrid
learning based

Embedded
Bayesian Neuro-fuzzy Series
SVM Supervised Unsupervised parameter
approaches systems combinations
optimization

Parallel
ANN DL SOM AE combinations

Semi-supervised
DBN

Fig. 16  Classification of AI-based prognostics approaches

to automatically detect pitch faults in wind turbines, while in Silva et al. (2014), the authors
directly used the condition monitoring measures of the output voltage as the input of ANFIS
to predict the degradation of fuel cell. One of the common limitations of NFS is the require-
ment of prior knowledge of a system and empirical data to construct fuzzy rules.

4.5.4 Hybrid AI methods

In the literature, most of the review papers (Kan et al. 2015; Atamuradov et al. 2017; Javed
et  al. 2017; Ellefsen et  al. 2019) consider hybrid prognostics approaches as a combina-
tion between model-based methods and data-driven methods. However, since this work
focuses on AI approaches, only combinations of AI tools are investigated. Similar to FDD
approaches, these combinations can be classified into three groups: (1) series combina-
tions, (2) parallel combinations, and (3) embedded parameter optimization methods.
Table 6 summarizes existing hybrid AI-based prognostics approaches. In the first group,
series combinations, Bayesian approaches and their variants are usually used to update the
prior RUL results, which are obtained using machine learning tools, such as neural net-
works (Gebraeel and Lawley 2008), neuro-fuzzy systems (Chen et  al. 2012), SVR (Liao
and Köttig 2014), kNN (Mosallam et al. 2016), and deep belief networks (Deutsch et al.
2017). Besides, other studies investigated prognostics as a procedure consisting of several
steps: fault detection, HI construction or degradation modeling, and RUL prediction. Each
step in this procedure is performed by an appropriate method (Soualhi et al. 2014).
In the second group, the parallel combination approaches often train an ensemble of
models to predict RUL and then refine these results using an optimization algorithm (Du
et al. 2012) or the particle filtering method (Baraldi et al. 2013). Besides, the prognostics
procedure can also be performed by a synchronous state estimation and prediction (Javed
et al. 2015) or simultaneous long- and short-term RUL predictions (Wang et al. 2018).
Tuning hyperparameters is one of the challenges in using machine learning based RUL
prediction approaches and this is addressed in the last hybrid group. In other words, this
group concerns embedding an optimization method to find the best hyperparameters for the

13
 3690 K. T. P. Nguyen et al.

prediction models. For instance, particle swarm optimization is used in Jiang et al. (2016);
Lu et  al. (2016) for optimizing hyperparameters of a support vector machine in order to
improve RUL prediction results. Besides, evolutionary algorithms are also used in Listou
Ellefsen et al. (2019); Zhang et al. (2017) to tune numerous structural parameters of deep
learning models.

A general guideline for selecting appropriate AI-based prognostics methods

Based on the strengths and the weaknesses of the AI-based prognostics approaches,
Fig.  17 presents a general guideline to select the appropriate one for a particular issue.
Firstly, if the practitioner prefers using the available prior knowledge, e.g. prior RUL distri-
bution of the system, then the Bayesian approach and its variants are favorable. Secondly,
if one tries to develop a solution that aims to address the expert knowledge’s flexibility, the
fuzzy systems are good candidates. Furthermore, a combination between fuzzy logics and
neural networks, for example ANFIS, facilitates integrating historical monitoring data in
prognostics based on the expert knowledge. Thirdly, when the expert and the prior knowl-
edge are unavailable, the neural network and its variants are usually preferred for large
monitoring database, while the SVM approaches are alternative solutions for a moderate
amount of data. Among different NN methods, the auto encoder and self-organizing map
techniques are the most appropriate ones for unlabeled data, while the deep belief network
can be used for partially labeled data. Besides, the performance of deep neural networks
(CNN, RNN, GBR, or LSTM) is highlighted for very large datasets. In addition, CNN is
the best candidate for extracting useful features of very high dimensional data, while RNN
and its variants are favorable for highly nonlinear and complex time-series data. Finally,
the hybrid approaches, that inherit the advantages of their baseline methods, could be
developed to address the heterogeneity of data sources (the monitoring data, the expert
and prior knowledge) or to response to the particular requirements of the issue and special
properties of the data.

Discussion on prognostics performance metrics

Metrics developed for evaluating prognostics algorithms’ performance can be classi-

fied into three categories: (1) accuracy-based metrics, (2) precision-based metrics, and (3)
robustness-based metrics. Among them, accuracy- and precision-based metrics are widely
used in existing studies while the notion of robustness was mentioned in several cases but
has not been formally defined. For a brief review of these metrics, the reader can consult
the paper (Saxena et al. 2008).

4.6 Decision support

Decision support is a crucial topic that has been widely studied in the operations research
domain. However, despite the increasing interest of predictive maintenance in practice, it
is still an under-explored topic in the field of PHM. Consequently, this section is dedicated
to briefly reviewing the work in which AI is the core for making decision in PHM applica-
tions. Inspired by previous reviews (Diez-Olivan et  al. 2019; Carvalho et  al. 2019), this
topic can be classified into three principal groups based on the purpose: (1) maintenance
actions; (2) joint maintenance and production scheduling or inventory management; and
(3) autonomous health management including mission reconfiguration, life extension, and

13
A review of artificial intelligence methods for engineering… 3691

intelligent control of the operational integrity. Figure  18 summarizes different decision-

making support approaches, which are recorded through a brief review of related work
in the literature, from input data to the methods for performing a specific decision. This
diagram presents the main sources of information along with the main goals of popular AI
approaches for making decisions in PHM applications.
As discussed in Sect. 4.1, all four data sources (expert knowledge, condition monitor-
ing data, historical operating & maintenance data, and external information) are useful
for obtaining accurate RULs and FDD information which are key inputs to the decision
module (Fig.  18). Besides, condition monitoring data can be exploited to assess the cur-
rent system’s heath state through Markov deterioration processes, failure probability, and
constructed HI, or to update a historical failure rate. Additionally, from historical operat-
ing & maintenance data, we can easily evaluate the MTBF/MTTR values, failure probabil-
ity, and failure rate to have the prior knowledge about the system’s health characteristics.
According to the obtained useful information, an appropriate decision support method is
developed to attain a particular objective. Further details, such as which method is recom-
mended for which input, will be discussed in the sequel.
In the largest group, the supporting methods for maintenance decisions using diagnos-
tics, prognostics, and reliability information have been studied in numerous studies. How-
ever, this section covers only the work that involves developing AI techniques for decision
support. For instance, fuzzy inference systems were proposed for planning maintenance
activities considering RUL values (Asmai et  al. 2010), historical operating and mainte-
nance data (Ruiz-Sarmiento et  al. 2020), and directly using vibration and temperature
monitoring data (Baban et al. 2016). This method allows for developing flexible rule-based
maintenance models that are able to exploit expert knowledge to overcome uncertainties
in input data. Besides, methods based on reinforcement learning (Ling et  al. 2018; Roc-
chetta et al. 2019), which consider maintenance activities as part of a multistage decision
process instead of a single action, allow for improving the long-term reward brought by
an existing system. However, these methods are often applied to cases where system deg-
radation can be formulated by a Markov process (Nguyen et al. 2013, 2014). This model
could be derived based on sensor measurements when there exits a sufficient number of
recorded life trajectories (Cannarile et al. 2018). In addition to the mentioned methods, the
real options analysis proposed in Lei et al. (2015) allows for responding to the flexibility of
maintenance decisions when considering large-scale systems, such as offshore wind farms
with multiple fan turbines. Those approaches are usually based on RUL values or HIs of
the system.
As maintenance activities have a close relationship with production scheduling and
logistics management, the joint optimization of these decisions can bring more profits for
managers (Nguyen et al. 2017, 2019). However, most of the work reported in this second
group is based on the available reliability information such as the MTTR/MTBF (Ahmadi
et al. 2016; Wang and Liu 2015), failure rate (Liao et al. 2017), failure probability (Diaz
et al. 2017), and health index (Liao et al. 2017), rather than prognostics information. Due
to the complexity and diversity of decisions that need to be optimized, these studies tended
to use some variant of evolutionary optimization algorithms for solving the joint main-
tenance and production scheduling problem taking into account the external information
about operating conditions. Moreover, the authors in Nguyen and Medjaher (2019) have
developed a dynamic predictive maintenance framework including spare part management.
This is a global process going from deep-learning-based prognostics to decision making
using an option analysis approach.

13
 3692 K. T. P. Nguyen et al.

In addition to using in maintenance decisions, fuzzy inference systems, which leverage

expert knowledge, also contribute to the development and implementation of autonomous
health management (AHM) systems. To implement an AHM system, we can first monitor
the system’s state in real time, then project it into the future to predict the RUL at each time
instance. This prognostic information is finally used for mission re-planning and intelligent
control of the operational integrity to ensure that the system continues to operate within an
acceptable and stable regime. This methodology has been applied to autonomous vehicles,
especially to predicting the battery RUL and re-planning their mission (Tang et al. 2011).
In Skima et al. (2019), the authors proposed a distributed approach based on the Dijkstra
algorithm to optimize the use of the conveying surface of micro-electro-mechanical sys-
tems and extend as long as possible the MEMES valves RULs.

5 Conclusions and perspectives

For the prognostics and health management community, artificial intelligence has opened
various growth opportunities. As seen in this review, although PHM applications unre-
lated to AI were born and developed earlier, since 2004 AI solutions have attracted more
and more attention, catching up with the growth of non-AI approaches and asserting their
superior position in recent years. Although the tremendous growth trend of AI-based PHM
applications has been confirmed in a variety of engineering disciplines through literature
surveys, in practice several non-AI-related methods are still preferable in certain industrial
applications. Especially in a well-established process or in use cases where unscheduled
downtimes are not a big deal, investing on an AI-based PHM solution may not be profit-
able compared to using a more traditional non-AI approach such as corrective or periodic
maintenance.
In addition to the broad quantitative survey, this paper has also presented a comprehen-
sive qualitative review of AI-based methods for implementing a complete PHM solution,
from data preparation and health indicator construction to decision support. One can notice
a large number of developed methods due to the diversity of the investigated issues. How-
ever, none of the existing approaches is superior than the others. Indeed, the superiority of
an approach is only case-specific and cannot be generalized for all systems. A proper selec-
tion of existing methods is essential to achieve maximum performance in the health man-
agement of the studied system. Thus the guidelines for choosing the appropriate method
to accomplish PHM tasks were also provided in this paper, aiming to help maintenance
practitioners efficiently implement autonomous health management solutions in practice.
Inspecting the growth trend of the current state of the art, our analysis not only provides
a better overview of current AI development in the PHM domain but also enlightens future
research directions. Specifically, critical challenges of PHM solutions for real-world prob-
lems are identified as follows:

• The challenge of the diversity, dissimilarity, and complexity of real-world prob-

lems. Although a wide range of AI methods is developed in the PHM field, most
of them may perform well in laboratory settings, yet have not been proven to work
effectively in the real world. In practice, operating conditions of industrial processes
are diverse and may be affected by many factors, such as manufacturing planning,
raw material quality, and operator skills. Consequently, an effective training phase
of AI algorithms requires a large amount of data samples which are difficult to get

13
 Table 6  Hybrid AI methods for engineering prognostics grouped by how the combination is established
Hybridization Methods Description

Serial combination NN and Bayesian (Gebraeel and Lawley 2008) NN for prediction of prior RUL distribution that is then updated by Bayesian approach
NFS and PF (Chen et al. 2012) NFS for prediction of the machine fault state evolution and PF for update the prior RUL distribution.
AC, HMM, and ANFIS (Soualhi et al. 2014) AC for fault detection, HMM for modeling degradation process and ANFIS for prediction of the
remaining time
SVR and PF (Liao and Köttig 2014) SVR to build the measurement model, similarity-based prediction fed into the PF for long-term
prediction
UL, kNN, and Bayesian (Mosallam et al. 2016) UL to construct HI, kNN for similarity-based prediction of HI evolution, and Bayesian filter for RUL
distribution prediction
DBN and PF (Deutsch et al. 2017) Integrating a deep belief network and a particle filter for RUL prediction.
Parallel combination NNs and PSO (Du et al. 2012) PSO to select an optimal subset of RUL predictions performed based on a bagged ensemble of NNs.
NNs and PF (Baraldi et al. 2013) Bagged ensemble of NNs is embedded in PF as an empirical measurement model to predict RUL
A review of artificial intelligence methods for engineering…

SWELM and SMEFC (Javed et al. 2015) Integrating two new algorithms, i.e. SWELM and SMEFC, for simultaneous discrete state estimation
and RUL prediction
LSTM, GBR, and NN (Wang et al. 2018) LSTM for long-term prediction of degradation trend, GBR for short-term prediction of RUL, and NN
to obtain hybrid results
Embedded param- PSO and SVM + DBN (Jiang et al. 2016) PSO to optimize the regularization and kernel parameters of SVM, which is combined with a deep
eter optimization belief network for state evolution prediction
PSO and SVM + PCA (Lu et al. 2016) PSO to optimize the parameters of least squares SVM, which is used to predict the degradation trend
of HI fused by PCA.
EA and DBN (Zhang et al. 2017) Evolutionary algorithm to optimize the parameters of a deep belief networks ensemble
GA and DL (Listou Ellefsen et al. 2019) Genetic algorithm to tune hyper-parameters of a semi-supervised deep architecture structure that
consists of RBM, LSTM, and FNN

NN: Neural Networks, SL: Supervised Learning, PCA: Principal Component Analysis, NFS: Neuro-Fuzzy System, PF: Particle Filtering; AC: Ant Clustering; HMM: Hid-
den Markov Model; ANFIS: Adaptive Neuro-Fuzzy Inference System; SVR: Support Vector Regression; UL: Unsupervised Learning; GBR: Gradient Boosting Regression;
SWELM: Summation Wavelet-Extreme Learning Machine; SMEFC: Subtractive-Maximum Entropy Fuzzy Clustering; PSO: Particle Swarm Optimization
3693

13
 3694 K. T. P. Nguyen et al.

Hybrid
approaches

YES
Particular
requirement?

YES YES YES

Choice of AI based
Based on prior Based on expert Large monitoring
prognostics SVM
knowledge? knowledge? data?
approaches? NO NO
NO

YES
YES YES
Neural network based
methods

NO
Labeled data?

YES NO
Partially labeled data?

NO
Abundant data? YES
Fuzzy
systems High noisy data?
YES DBN NO

Expert knowledge? YES

Sufficient history
data?
NO Sequential
Very-high-
dimensional data? data?

YES
YES YES YES

Bayesian Neuro-fuzzy systems, RNN, GBR,

approaches ANFIS ANN CNN Auto-Encoder SOM
e.g. ANFIS LSTM

Fig. 17  A guideline for selecting appropriate AI-based prognostics approaches

in practice. Thus developing algorithms that allow for integrating various types of
expert knowledge into the training is highly recommended.
• The challenge of a fully digitalized monitoring system. Nowadays, due to the
fierce global competition, companies must continuously improve their efficiency and
reduce maintenance and operation costs. One of the levers to reach those objectives
is to develop a fully digitalized traceable production line from input materials to
manufacturing configurations to product quality assurance. Deviations in the final
products not attributed to raw materials nor production recipes would imply issues
in the operation of sensors or control systems or thermo-electro-mechanical equip-
ment. In this situation, potential failures or degradations (or the need for a local-sys-
tem recalibration) can be figured out by using data collected from the whole end-to-
end production line rather than from data acquisition of a local equipment/module.
• The challenge of large-scale problems in Industry 4.0. In the context of Industry
4.0, the studied systems become large-scale and highly complex. It is therefore nec-
essary to investigate PHM issues not only at the system level but also at a company
level and toward large-scale ecosystems. Consequently, AI-based PHM methods
should take into account multi-level decision making. For instance, at the system
level, accurate machine health prognostics is necessary for predictive maintenance
decisions; while supplementary information, such as degradation states of other
machines, spare part inventory, production schedules, and repairman resources, is
essential for an efficient maintenance policy at the company level.
• The challenge of big and heterogeneous data. The performance of AI-based PHM
methods usually depends on the quantity and quality of the collected data. In recent
years, thanks to the great development of sensing technologies as well as the emer-
gence of IoT, the amount of data that can be generated and recorded has increased

13
A review of artificial intelligence methods for engineering… 3695

Fig. 18  Mapping between data sources, approaches, and main objectives for the decision-making support
phase

exponentially. However, it will be useless if these big and heterogeneous data cannot
be accessed and exploited quickly enough for early fault detection, diagnostics, and
prognostics to prevent potential damages in a timely manner.
• The challenge of real-time high-performance computing. Training an AI model
usually requires a lot of computational resources. For instance, deep learning, which
is essentially the optimization of millions of model parameters , needs a massive
amount of data. Consequently, this results in extensive consumption of high-perfor-
mance computing resources and long training times.

To address the above-mentioned challenges, some future research directions in develop-

ing AI-enabled PHM (AI–PHM) solutions are identified and summarized as follows:

• Future AI–PHM researches could focus on exploiting the adaptation and self-
learning ability of online PHM methods to handle unforeseen variabilities, and
additionally to deal with unspecified anomalies or new faults occurring during the
online operation. They should enable the autonomous detection of environment
changes and facilitate the transferability as well as adaptation of the trained models
to new operating conditions through an automatic update of the model’s hyperpa-
rameters.
• Future AI–PHM studies may prioritize the development of a more proactive physics-
aware machine learning framework to reduce data dependency, enhance the robust-
ness, and improve the interpretability of ML models. The developed framework should
facilitate the integration of physical properties of the considered system into the design
of new ML models, rather than just modifying existing ML architectures.
• Future AI–PHM solutions need to incorporate accessing a variety of information
flows across multiple organizations for the multi-location maintenance optimi-
zation. The developed solutions must be scalable, flexible, and easily integrated into
the available information technology infrastructure. To do this, they require inclusion
of advanced technologies, such as AI with automatic self-learning abilities, IoT for

13
 3696 K. T. P. Nguyen et al.

accessing multi-level databases, and cloud and cognitive computing for enhancing the
computational capability.
• Future AI–PHM research directions should encourage the development of powerful
approaches dedicated to multimodal data. For instance, a milling process of a multi-
axis robot can be monitored through camera, force, torque, vibration, and current sen-
sors. Measurements from the different sensors can reinforce the information inferred
from the numerical modality, while images and audios recorded by cameras can serve
as the complementary data to monitor the entire robot behaviors in the milling pro-
cess. Then the work of joining those multimodal data into a uniform representational
space becomes a crucial challenge that needs to be addressed. Besides, it is necessary
to develop efficient algorithms for learning the direct relationships between events and
indicators from each of the modalities, so that a future event can be first predicted based
on the earliest indicator, and then verified by indicators from the other modalities.
• Future AI–PHM solutions may include high-performance computing techniques via
the use of parallel programming for running advanced PHM applications quickly,
efficiently, and reliably. In addition, cloud computing is also a powerful environment
to implement high-computational AI-based PHM techniques. This enables sending
monitoring data to the cloud, where high-performing servers can process them simul-
taneously, extract features, assess the system’s current heath state, and predict its future
evolution. However, the continuous transmission of multiple large data flows to the
cloud would significantly increase the bandwidth cost on the one hand, and may cause
network congestion which results in unpredictable processing delays on the other hand.
In this respect, future research directions may encourage the development of AI-based
PHM frameworks in which monitoring data can be partly processed on edge com-
puting units rather than entirely collectively on cloud servers.

In addition to providing health indicators, PHM systems can be also used to help operation
managers optimize the lifecycle of their engineering systems. Future PHM will no longer
be an individual addition to supporting maintenance, but rather be integrated into the entire
system’s lifecycle from the beginning during the conceptualization and design phases.
Indeed, PHM information and feedback can be exploited to well design an engineering
system and make it easy to monitor and repair when failed. Also, a well-designed system
has better maintainability and therefore better availability, leading to more productivity and
less maintenance costs. To do this, new AI–PHM concepts should prioritize the develop-
ment of efficient methods allowing for dealing with uncertainties in the assessment, quan-
tification, and formalization of the operator’s knowledge and feedback in order to extract
useful information for further exploitation.
Besides, future AI–PHM solutions can be used to optimize the utilization of engineer-
ing systems taking into account multiple requirements from various stakeholders such as
logistics managers, maintenance technicians, and operators. We will therefore need effi-
cient methods that allow for formalization and integration of such requirements into the
decision-making process. Additionally, sequential-decision optimization methods targeted
at multiple criteria (such as reduction of operating cost, improvement of mission comple-
tion rate, and satisfaction of safety constraints) should also be prioritized going forward.

13
A review of artificial intelligence methods for engineering… 3697

Annex

Search string used for this review:

TS==(
((prognostic OR prognosis) AND (remaining useful life OR degradation OR predictive
maintenance))
OR
((machinery OR machine) AND (prognostic OR prognosis OR fault detection OR fault
diagnostic
OR fault diagnosis OR predictive maintenance OR condition based maintenance))
)
AND
SJ==(“ENGINEERING” OR “COMPUTER SCIENCE” OR “AUTOMATION CON-
TROL SYSTEMS” OR “TELECOMMUNICATIONS” OR “PHYSICS” OR “MECHAN-
ICS” OR “ACOUSTICS” OR “MATHEMATICS” OR “OPERATIONS RESEARCH
MANAGEMENT SCIENCE” OR “SCIENCE TECHNOLOGY OTHER TOPICS” OR
“ENERGY FUELS” OR “INSTRUMENTS INSTRUMENTATION” OR “CHEMIS-
TRY” OR “MATERIALS SCIENCE” OR “ROBOTICS” OR “TRANSPORTATION” OR
“THERMODYNAMICS” OR “MINING MINERAL PROCESSING” OR “IMAGING
SCIENCE PHOTOGRAPHIC TECHNOLOGY” OR “ENVIRONMENTAL SCIENCES
ECOLOGY” OR “ELECTROCHEMISTRY” OR “CONSTRUCTION BUILDING TECH-
NOLOGY” OR “NUCLEAR SCIENCE TECHNOLOGY” OR “INFORMATION SCI-
ENCE LIBRARY SCIENCE” OR “REMOTE SENSING” OR “AGRICULTURE” OR
“WATER RESOURCES”)
AND
LA==(“ENGLISH”)
To create the list of unsuitable categories According to the WC (Web of Science cat-
egory) column and the ID (keyword plus) column, we extract the unique categories/key-
words from their correspondent column and sort them with their occurrence frequency.
Then, by investigating the obtained results, we identify the unsuitable categories/keywords
according to our knowledge and record them in the particular lists.

• Examples of unsuitable categories: ‘audiology & speech-language pathology’,

‘ecology’, ‘agriculture, dairy & animal science’, ‘endocrinology & metabolism’,
‘oncology’, ‘genetics & heredity’, ‘biodiversity conservation’, ‘psychiatry’, ‘car-
diac & cardiovascular systems’, ‘geography, physical’, ‘business, finance’, ‘neu-
roimaging’, ‘transplantation’, ‘toxicology’, ‘geochemistry & geophysics’, ‘urban
studies’, ‘evolutionary biology’, ‘microscopy’, ‘critical care medicine’, ‘regional
& urban planning’, ‘microbiology’, ‘physiology’, ‘cell & tissue engineering’, ‘psy-
chology’, ‘forestry’, ‘surgery’, ‘dentistry, oral surgery & medicine’, ‘social sci-
ences, ‘psychology, experimental’, ‘political science’, ‘medicine, general &
internal’, ‘public & occupational health’, ‘information science & library science’,
‘pharmacology & pharmacy’, ‘agricultural economics & policy’, ‘social issues’,
‘medical laboratory technology’, ‘cell biology’, ‘rehabilitation’, ‘behavioral sci-
ences’, ‘nursing’, ‘orthopedics’, ‘communication’, ‘horticulture’, ‘immunology’,
‘limnology’, ‘health care sciences & services’.
• Examples of inappropriate keywords: ‘cancer’, ‘cells’, ‘gene-expression’,
‘carcinoma’, ‘apoptosis’, ‘biomarkers’, ‘bispectrum’, ‘brain’, ‘in-vitro’, ‘in-vivo’,

13
 3698 K. T. P. Nguyen et al.

‘adenocarcinoma’, ‘angiogenesis’, ‘hepatocellular-carcinoma’, ‘lung-cancer’,

‘squamous-cell carcinoma’, ‘colorectal-cancer’, ‘ovarian-cancer’, ‘tumor’, ‘tum-
origenesis’, ‘cancer-cells’, ‘prostate-cancer’, ‘lesions’, ‘dementia’, ‘alzheimers-
disease’, ‘adhesion’, ‘cell lung-cancer’, ‘colon-cancer’, ‘adults’, ‘methylation’,
‘prevalence’, ‘myocardial-infarction’, ‘estrogen’, ‘immunohistochemistry’, ‘micro-
rnas’, ‘gliomas’, ‘breast-cancer diagnosis’, ‘high-grade gliomas’,’drug-resistance’,
‘blood-pressure’, ‘pathology’, ‘sepsis’, ‘malignant gliomas’, ‘microrna-related
genes’, ‘mesenchymal transition’, ‘culture’, ‘monoclonal-antibodies’, ‘ubiquitina-
tion’, ‘heart-rate-variability’, ‘genome’, ‘uio sequences’, ‘polymorphisms’, ‘bispec-
tra’, ‘bioinformatics’, ‘chirplet transform’, ‘biogenesis’, ‘protein’, ‘metastases’,
‘chromatin’, ‘gene-expression’, ‘skeletal-muscle’, ‘lung’, ‘mice’, ‘e-cadherin’,’beta-
catenin’, ‘antibodies’.

To create the list of AI-related keywords We identify the appropriate keyworks from
the corresponding list (extracted from ID column) according to our knowledge and
record them in a particular list.

• Examples of AI-related keywords: ‘machine learning’, ‘supervised learning’,

‘unsupervised learning’, ‘semi supervised learning’, ‘reinforcement learning’,
‘evolutionary’, ‘genetic algorithm’, ‘genetic programming’, ‘fuzzy inference sys-
tem’, ‘clustering’, ‘classification’, ‘particle swarm’, ‘deep learning’,‘recurrent net-
work’, ‘convolutional network’, ‘recurrent-network’, ‘convolutional-network’,‘deep
belief’, ‘bayesian network’, ‘auto encoder’, ‘neural network’, ‘neuro fuzzy’ ‘svm’,
‘self organizing map’, ‘self-organizing map’ ‘decision tree’, ‘anfis’, ‘artificial intel-
ligence’, ‘ant colony optimization’, ‘gradient boosting regression’ ‘statistical
learning’, ‘partial least square’, ‘isomap’, ‘radial basis function’, ‘organizing map’,
‘neighbour nearest’, ‘pca’, ‘knn’, ‘lle’, ‘ lda ’, ‘ica’, ‘fda’, ‘ann’, ‘cnn’, ‘lstm’, ‘rnn’, ‘rbm’,
‘gru’, ‘narx’, ‘som’, ‘neural-network’, ‘extreme learning-machine’, ‘pattern-recog-
nition’, ‘fisher discriminant’, ‘particle swarm optimization’, ‘rbfn’, ‘classifier’, ‘vec-
tor machine’, ‘autoencoder’, ‘random forest’, ‘artificial-intelligence’, ‘learning-
method’, ‘restricted boltzmann’, ‘pattern recognition’, ‘extreme learning’, ‘t-sne’,
‘umap’,‘regression’.

Author Contributions  Khanh Nguyen: conceptualization, methodology, software, validation, formal analy-
sis, investigation, writing of original draft, visualization, review and editing; Kamal Medjaher: conceptu-
alization, methodology, writing of original draft, visualization, review and editing; Do Tran: conceptualiza-
tion, data curation, software, writing of original draft, review and editing.

Funding  Not applicable

Availability of data and materials The extracted and curated data used in this paper are available upon
request to the corresponding author.

Code availability  The Python scripts used to generate analytical results are available upon request to the
corresponding author.

Declarations 
Competing interests  The authors declare that they have no competing interests.

13
A review of artificial intelligence methods for engineering… 3699

References
Abed W, Sharma S, Sutton R, Motwani A (2015) A Robust bearing fault detection and diagnosis technique
for brushless DC motors under non-stationary operating conditions. J Control Automat Electric Syst
26(3):241–254
Abid A, Khan MT, Iqbal J (2021) A review on fault detection and diagnosis techniques: basics and beyond.
Artif Intell Rev 54(5):3639–3664. https://​doi.​org/​10.​1007/​s10462-​020-​09934-2
Ahmadi E, Zandieh M, Farrokh M, Emami SM (2016) A multi objective optimization approach for flexible
job shop scheduling problem under random machine breakdown by evolutionary algorithms. Comput
Oper Res 73:56–66
Aldrich C, Auret L (2010) Fault detection and diagnosis with random forest feature extraction and variable
importance methods. IFAC Proc Vol 43(9):79–86
Al-Dulaimi A, Zabihi S, Asif A, Mohammadi A (2019) A multimodal and hybrid deep neural network
model for Remaining Useful Life estimation. Comput Ind 108:186–196
An D, Kim NH, Choi J-H (2015) Practical options for selecting data-driven or physics-based prognostics
algorithms with reviews. Reliab Eng Syst Saf 133:223–236
Aremu OO, Cody RA, Hyland-Wood D, McAree PR (2020) A relative entropy based feature selection
framework for asset data in predictive maintenance. Comput Ind Eng
Askarian M, Escudero G, Graells M, Zarghami R, Jalali-Farahani F, Mostoufi N (2016) Fault diagnosis
of chemical processes with incomplete observations: a comparative study. Comput Chem Eng
84:104–116
Asmai SA, Hussin B, Yusof MM (2010) A framework of an intelligent maintenance prognosis tool. In: 2010
second international conference on computer research and development, pp 241–245
Atamuradov V, Medjaher K, Dersin P, Lamoureux B, Zerhouni N (2017) Prognostics and health manage-
ment for maintenance practitioners—review, implementation and tools evaluation. Int J Prognost
Health Manag 8:1–31
Atamuradov V, Medjaher K, Camci F, Dersin P, Zerhouni N (2018) Degradation-level assessment and
online prognostics for sliding chair failure on point machines. IFAC-PapersOnLine 51(24):208–213
Baban CF, Baban M, Suteu MD (2016) Using a fuzzy logic approach for the predictive maintenance of tex-
tile machines. J Intell Fuzzy Syst 30(2):999–1006
Baraldi P, Compare M, Sauco S, Zio E (2013) Ensemble neural network-based particle filtering for prognos-
tics. Mech Syst Signal Process 41(1):288–300
Bedekar PP, Bhide SR, Kale VS (2011) Fault section estimation in power system using Hebb’s rule and con-
tinuous genetic algorithm. Int J Electric Power Energy Syst 33(3):457–465
Benkedjouh T, Medjaher K, Zerhouni N, Rechak S (2013) Remaining useful life estimation based on non-
linear feature reduction and support vector regression. Eng Appl Artif Intell 26(7):1751–1760
Biagetti T, Sciubba E (2004) Automatic diagnostics and prognostics of energy conversion processes via
knowledge-based systems. Energy 29(12):2553–2572
Bickford R, Malloy D (2002) Development of a real-time turbine engine diagnostic system. In: The 38th
AIAA/ASME/SAE/ASEE joint propulsion conference & exhibit. American Institute of Aeronautics
and Astronautics
Buezas FS, Rosales MB, Filipich CP (2011) Damage detection with genetic algorithms taking into account
a crack contact model. Eng Fract Mech 78(4):695–712
Cannarile F, Compare M, Baraldi P, Di Maio F, Zio E (2018) Homogeneous continuous-time, finite-state
hidden semi-Markov modeling for enhancing empirical classification system diagnostics of industrial
components. Machines 6(3):34
Cao L, Gu Q (2002) Dynamic support vector machines for non-stationary time series forecasting. Intel
Data Anal 6(1):67–83
Carvalho TP, Soares FAAMN, Vita R, Francisco RdP, Basto JP, Alcalá SGS (2019) A systematic litera-
ture review of machine learning methods applied to predictive maintenance. Comput Ind Eng
Cerrada M, Sánchez RV, Cabrera D, Zurita G, Li C (2015) Multi-stage feature selection by using genetic
algorithms for fault diagnosis in gearboxes based on vibration signal. Sensors 15(9):23903–23926
Chandrashekar G, Sahin F (2014) A survey on feature selection methods. Comput Electr Eng
40(1):16–28
Chen Y, Lan L (2009) A fault detection technique for air-source heat pump water chiller/heaters. Energy
Build 41(8):881–887
Chen C, Zhang B, Vachtsevanos G (2012) Prediction of machine health condition using neuro-fuzzy and
Bayesian algorithms. IEEE Trans Instrum Meas 61(2):297–306

13
 3700 K. T. P. Nguyen et al.

Chen X, Shen Z, He Z, Sun C, Liu Z (2013) Remaining life prognostics of rolling bearing based on
relative features and multivariable support vector machine. Proc Inst Mech Eng C J Mech Eng Sci
227(12):2849–2860
Chen B, Matthews PC, Tavner PJ (2013) Wind turbine pitch faults prognosis using a-priori knowledge-
based ANFIS. Expert Syst Appl 40(17):6863–6876
Chen J, Jing H, Chang Y, Liu Q (2019) Gated recurrent unit based recurrent neural network for remain-
ing useful life prediction of nonlinear deterioration process. Reliab Eng Syst Saf 185:372–382
Chiang LH, Kotanchek ME, Kordon AK (2004) Fault diagnosis based on Fisher discriminant analysis
and support vector machines. Comput Chem Eng 28(8):1389–1401
Chinnam RB, Baruah P (2004) A neuro-fuzzy approach for estimating mean residual life in condition-
based maintenance systems. Int J Mater Prod Technol 20(1–3):166–179
Cho H-W (2007) Identification of contributing variables using kernel-based discriminant modeling and
reconstruction. Expert Syst Appl 33(2):274–285
Chou J-H, Ghaboussi J (2001) Genetic algorithm in structural damage detection. Comput Struct
79(14):1335–1353
Cong F, Chen J, Dong G (2012) Spectral kurtosis based on AR model for fault diagnosis and condition
monitoring of rolling bearing. J Mech Sci Technol 26(2):301–306
Costello JJA, West GM, McArthur SDJ (2017) Machine learning model for event-based prognostics in
gas circulator condition monitoring. IEEE Trans Reliab 66(4):1048–1057
Das HC, Parhi DR (2009) Application of Neural Network for fault diagnosis of cracked cantilever beam.
In: 2009 World Congress on nature biologically inspired computing (NaBIC), pp. 1303–1308
de Souza DL, Granzotto MH, de Almeida GM, Oliveira-Lopes LC (2014). Fault detection and diagno-
sis using support vector machines—a SVC and SVR comparison. https://​doi.​org/​10.​5923/j.​safety.​
20140​301.​03
Deekshit  Kompella KC, Venu Gopala  Rao M, Srinivasa  Rao R (2018) Bearing fault detection in a 3
phase induction motor using stator current frequency spectral subtraction with various wavelet
decomposition techniques. Ain Shams Eng J 9(4):2427–2439
Dehestani D, Eftekhari F, Guo Y, Ling S, Su S, Nguyen H (2011) Online support vector machine applica-
tion for model based fault detection and isolation of HVAC system. Int J Mach Learn Comput 66–72
Deutsch J, He D (2018) Using deep learning-based approach to predict remaining useful life of rotating
components. IEEE Trans Syst Man Cybernet Systms 48(1):11–20
Deutsch J, He M, He D (2017) Remaining useful life prediction of hybrid ceramic bearings using an
integrated deep learning and particle filter approach. Appl Sci 7(7):649
Diaz JE, Handl J, Xu D-L (2017) Evolutionary robust optimization in production planning—interactions
between number of objectives, sample size and choice of robustness measure. Comput Oper Res
79:266–278
Diez-Olivan A, Del Ser J, Galar D, Sierra B (2019) Data fusion and machine learning for industrial
prognosis: trends and perspectives towards Industry 4.0. Inf Fusion 50:92–111
Dong H-Y, Yang L-X, Hong-Wei L (2016) Small Fault diagnosis of front-end speed controlled wind
generator based on deep learning
Dong Y (2019) Implementing deep learning for comprehensive aircraft icing and actuator/sensor fault
detection/identification. Eng Appl Artif Intell 83:28–44
Dong M, Yang Z-b (2008) Dynamic Bayesian network based prognosis in machining processes. J Shang-
hai Jiaotong Univ (Science) 13(3):318–322
Du S, Lv J, Xi L (2012) Degradation process prediction for rotational machinery based on hybrid intel-
ligent model. Robot Comput Integr Manufact 28(2):190–207
Du Z, Chen L, Jin X (2017) Data-driven based reliability evaluation for measurements of sensors in a
vapor compression system. Energy 122:237–248
Elbouchikhi E, Choqueuse V, Amirat Y, Benbouzid MEH, Turri S (2017) An efficient Hilbert-Huang
transform-based bearing faults detection in induction machines. IEEE Trans Energy Convers
32(2):401–413
Ellefsen AL, Æsøy V, Ushakov S, Zhang H (2019) A comprehensive survey of prognostics and health
management based on deep learning for autonomous ships. IEEE Trans Reliab 68(2):720–740
Evsukoff A, Gentil S (2005) Recurrent neuro-fuzzy system for fault detection and isolation in nuclear
reactors. Adv Eng Inform 19(1):55–66
Faghih-Roohi S, Hajizadeh S, Núñez A, Babuska R, De  Schutter B (2016) Deep convolutional neural
networks for detection of rail surface defects. In: 2016 international joint conference on neural
networks (IJCNN), pp 2584–2589. https://​doi.​org/​10.​1109/​IJCNN.​2016.​77275​22

13
A review of artificial intelligence methods for engineering… 3701

Fan J, Wang Y (2014) Fault detection and diagnosis of non-linear non-Gaussian dynamic processes
using kernel dynamic independent component analysis. Inf Sci 259:369–379. https://​doi.​org/​10.​
1016/j.​ins.​2013.​06.​021
Feng J, Wang J, Zhang H, Han Z (2016) Fault diagnosis method of joint Fisher discriminant analysis
based on the local and global manifold learning and its Kernel version. IEEE Trans Autom Sci
Eng 13(1):122–133. https://​doi.​org/​10.​1109/​TASE.​2015.​24178​82
Fink O, Wang Q, Svensén M, Dersin P, Lee W-J, Ducoffe M (2020) Potential, challenges and future
directions for deep learning in prognostics and health management applications. Eng Appl Artif
Intell 92:103678
Firpi H, Vachtsevanos G (2008) Genetically programmed-based artificial features extraction applied to
fault detection. Eng Appl Artif Intell 21(4):558–568
Frank SM, Lin G, Jin X, Singla R, Farthing A, Zhang L, Granderson J (March 2019) Metrics and meth-
ods to assess building fault detection and diagnosis tools. Technical report NREL/TP–5500-72801,
1503166. https://​doi.​org/​10.​2172/​15031​66
Ganguli R, Chopra I, Haas DJ (1998) Helicopter rotor system fault detection using physics-based model
and neural networks. AIAA J 36(6):1078–1086. https://​doi.​org/​10.​2514/2.​483
Gao X, Hou J (2016) An improved SVM integrated GS-PCA fault diagnosis approach of Tennessee East-
man process. Neurocomputing 174:906–911. https://​doi.​org/​10.​1016/j.​neucom.​2015.​10.​018
García-Pedrajas N, Ortiz-Boyer D (2009) Boosting k-nearest neighbor classifier by means of input space
projection. Expert Syst Appl 36(7):10570–10582. https://​doi.​org/​10.​1016/j.​eswa.​2009.​02.​065
Ge Z, Zhong S, Zhang Y (2016) Semisupervised Kernel learning for FDA model and its application for
fault classification in industrial processes. IEEE Trans Industr Inf 12(4):1403–1411. https://​doi.​
org/​10.​1109/​TII.​2016.​25716​80
Gebraeel NZ, Lawley MA (2008) A neural network degradation model for computing and updating
residual life distributions. IEEE Trans Autom Sci Eng 5(1):154–163. https://​doi.​org/​10.​1109/​
TASE.​2007.​910302
Gebraeel NZ, Lawley MA, Li R, Ryan JK (2005) Residual-life distributions from component degrada-
tion signals: a Bayesian approach. IIE Trans 37(6):543–557. https://​doi.​org/​10.​1080/​07408​17059​
09290​18
Gensler A, Henze J, Sick B, Raabe N (2016) Deep Learning for solar power forecasting—an approach
using AutoEncoder and LSTM Neural Networks. In: 2016 ieee international conference on sys-
tems, man, and cybernetics (SMC), pp 002858–002865. https://​doi.​org/​10.​1109/​SMC.​2016.​78446​
73
Gharahbagheri H, Imtiaz S, Khan F (2017). Root cause diagnosis of process fault using KPCA and
Bayesian network. https://​doi.​org/​10.​1021/​acs.​iecr.​6b019​16
Godoy JL, Vega JR, Marchetti JL (2013) A fault detection and diagnosis technique for multivariate pro-
cesses using a PLS-decomposition of the measurement space. Chemom Intell Lab Syst 128:25–36.
https://​doi.​org/​10.​1016/j.​chemo​lab.​2013.​07.​006
Gope D, Gope PC, Thakur A, Yadav A (2015) Application of artificial neural network for predicting
crack growth direction in multiple cracks geometry. Appl Soft Comput 30:514–528. https://​doi.​
org/​10.​1016/j.​asoc.​2015.​02.​003
Gouriveau R, Medjaher K, Zerhouni N (2016) From prognostics and health systems management to pre-
dictive maintenance 1: monitoring and prognostics. https://​doi.​org/​10.​1002/​97811​19371​052
Gugulothu N, TV V, Malhotra P, Vig L, Agarwal P, Shroff G (2017) Predicting remaining useful life
using time series embeddings based on recurrent neural networks. arXiv:​1709.​01073 [cs]
Guo L, Li N, Jia F, Lei Y, Lin J (2017) A recurrent neural network based health indicator for remaining
useful life prediction of bearings. Neurocomputing 240:98–109. https://​doi.​org/​10.​1016/j.​neucom.​
2017.​02.​045
Guo L, Lei Y, Li N, Yan T, Li N (2018) Machinery health indicator construction based on convolutional
neural networks considering trend burr. Neurocomputing 292:142–150
Gururajapathy SS, Mokhlis H, Illias HA (2017) Fault location and detection techniques in power distribu-
tion systems with distributed generation: a review. Renew Sustain Energy Rev 74:949–958. https://​
doi.​org/​10.​1016/j.​rser.​2017.​03.​021
Hamadache M, Jung JH, Park J, Youn BD (2019) A comprehensive review of artificial intelligence-based
approaches for rolling element bearing PHM: shallow and deep learning. JMST Adv 1:125–151.
https://​doi.​org/​10.​1007/​s42791-​019-​0016-y
Han H, Gu B, Hong Y, Kang J (2011) Automated FDD of multiple-simultaneous faults (MSF) and the appli-
cation to building chillers. Energy Build 43(9):2524–2532. https://​doi.​org/​10.​1016/j.​enbui​ld.​2011.​06.​
011

13
 3702 K. T. P. Nguyen et al.

Han T, Liu C, Yang W, Jiang D (2019) Learning transferable features in deep convolutional neural networks
for diagnosing unseen machine conditions. ISA Trans 93:341–353
Hariharan V, Srinivasan P (2009) New approach of classification of rolling element bearing fault using arti-
ficial neural network. J Mech Eng 40(2):119–130. https://​doi.​org/​10.​3329/​jme.​v40i2.​5353
Harmouche J, Delpha C, Diallo D (2014) Linear discriminant analysis for the discrimination of faults in
bearing balls by using spectral features. In: 2014 First International conference on green energy ICGE
2014, Sfax, Tunisia, pp 182–187. https://​doi.​org/​10.​1109/​ICGE.​2014.​68354​19
Hasani RM, Wang G, Grosu R (2017) An Automated Auto-encoder correlation-based health-monitoring
and prognostic method for machine bearings. arXiv:​1703.​06272 [cs, stat]
He QP, Wang J (2007) Fault detection using the k-nearest neighbor rule for semiconductor manufacturing
processes. IEEE Trans Semicond Manuf 20(4):345–354. https://​doi.​org/​10.​1109/​TSM.​2007.​907607
He QP, Qin SJ, Wang J (2005) A new fault diagnosis method using fault directions in Fisher discriminant
analysis. AIChE J 51(2):555–571. https://​doi.​org/​10.​1002/​aic.​10325
Heng A, Zhang S, Tan ACC, Mathew J (2009) Rotating machinery prognostics: state of the art, challenges
and opportunities. Mech Syst Signal Process 23(3):724–739. https://​doi.​org/​10.​1016/j.​ymssp.​2008.​
06.​009
Huang R, Xi L, Li X, Richard Liu C, Qiu H, Lee J (2007) Residual life predictions for ball bearings based
on self-organizing map and back propagation neural network methods. Mech Syst Signal Process
21(1):193–207. https://​doi.​org/​10.​1016/j.​ymssp.​2005.​11.​008
Huang C-G, Huang H-Z, Li Y-F (2019) A bidirectional LSTM prognostics method under multiple opera-
tional conditions. IEEE Trans Ind Electron 66(11):8792–8802. https://​doi.​org/​10.​1109/​TIE.​2019.​
28914​63
Jamil M, Sharma SK, Chaturvedi D (2015) Fault classification of three-phase transmission network using
genetic algorithm. Int J Eng Appl Sci 2(7):84–88
Jardine AKS, Lin D, Banjevic D (2006) A review on machinery diagnostics and prognostics implementing
condition-based maintenance. Mech Syst Signal Process 20(7):1483–1510
Javadian SAM, Nasrabadi AM, Haghifam M-R, Rezvantalab J (2009) Determining fault’s type and accurate
location in distribution systems with DG using MLP Neural networks. In: 2009 international confer-
ence on clean electrical power, pp 284–289. https://​doi.​org/​10.​1109/​ICCEP.​2009.​52120​44
Javed K, Gouriveau R, Zerhouni N (2015) A new multivariate approach for prognostics based on extreme
learning machine and fuzzy clustering. IEEE Trans Cybern 45(12):2626–2639. https://​doi.​org/​10.​
1109/​TCYB.​2014.​23780​56
Javed K, Gouriveau R, Zerhouni N (2017) State of the art and taxonomy of prognostics approaches, trends
of prognostics applications and open issues towards maturity at different technology readiness levels.
Mech Syst Signal Process 94:214–236. https://​doi.​org/​10.​1016/j.​ymssp.​2017.​01.​050
Jia F, Lei Y, Lin J, Zhou X, Lu N (2016) Deep neural networks: a promising tool for fault characteristic
mining and intelligent diagnosis of rotating machinery with massive data. Mech Syst Signal Process
72–73:303–315. https://​doi.​org/​10.​1016/j.​ymssp.​2015.​10.​025
Jiang Q, Yan X (2013) Non-Gaussian chemical process monitoring with adaptively weighted independ-
ent component analysis and its applications. J Process Control 23(9):1320–1331. https://​doi.​org/​10.​
1016/j.​jproc​ont.​2013.​09.​008
Jiang B, Zhu X, Huang D, Paulson JA, Braatz RD (2015) A combined canonical variate analysis and Fisher
discriminant analysis (CVA–FDA) approach for fault diagnosis. Comput Chem Eng 77:1–9. https://​
doi.​org/​10.​1016/j.​compc​hemeng.​2015.​03.​001
Jiang P, Chen C, Liu X (2016) Time series prediction for evolutions of complex systems: a deep learning
approach. In: 2016 IEEE International conference on control and robotics engineering (ICCRE), Sin-
gapore, pp 1–6. https://​doi.​org/​10.​1109/​ICCRE.​2016.​74761​50
Jin S, Liu Y, Lin Z (2012) A Bayesian network approach for fixture fault diagnosis in launch of the assembly
process. Int J Prod Res 50(23):6655–6666. https://​doi.​org/​10.​1080/​00207​543.​2011.​611543
Jin G, Matthews DE, Zhou Z (2013) A Bayesian framework for on-line degradation assessment and residual
life prediction of secondary batteries in spacecraft. Reliab Eng Syst Saf 113(C):7–20
Jin X, Zhao M, Chow TWS, Pecht M (2014) Motor bearing fault diagnosis using trace ratio linear discrimi-
nant analysis. IEEE Trans Ind Electron 61(5):2441–2451. https://​doi.​org/​10.​1109/​TIE.​2013.​22734​71
Jin Q, Ju R (2012) fault location for distribution network based on genetic algorithm and stage treatment.
In: 2012 Spring congress on engineering and technology, pp 1–4. https://​doi.​org/​10.​1109/​SCET.​2012.​
63420​90
Kan MS, Tan ACC, Mathew J (2015) A review on prognostic techniques for non-stationary and non-linear
rotating systems. Mech Syst Signal Process 62–63:1–20. https://​doi.​org/​10.​1016/j.​ymssp.​2015.​02.​016
Kang Y, Wang C-C, Chang Y-P (2007) Gear fault diagnosis in time domains by using Bayesian networks.
In: Castillo O, Melin P, Ross OM, Sepúlveda Cruz R, Pedrycz W, Kacprzyk J (eds) Theoretical

13
A review of artificial intelligence methods for engineering… 3703

advances and applications of fuzzy logic and soft computing. Advances in soft computing. Springer,
Berlin, Heidelberg, pp 741–751. https://​doi.​org/​10.​1007/​978-3-​540-​72434-6_​75
Katipamula S, Brambley MR (2005) Review article: methods for fault detection, diagnostics, and prognos-
tics for building systems—a review, part I. HVAC&R Res 11(1):3–25. https://​doi.​org/​10.​1080/​10789​
669.​2005.​10391​123
Khaire UM, Dhanalakshmi R (2019) Stability of feature selection algorithm: A review. J King Saud Univ
xxx:1–14
Khamis M, Gomaa W, Galal B (2016) Deep learning is competing random forest in computational docking.
arXiv:​1608.​06665 [cs, q-bio]
Khan S, Yairi T (2018) A review on the application of deep learning in system health management. Mech
Syst Signal Process 107:241–265. https://​doi.​org/​10.​1016/j.​ymssp.​2017.​11.​024
Khlaief A, Nguyen K, Medjaher K, Picot A, Maussion P, Tobon D, Chauchat B, Cheron R (2019) Feature
engineering for ball bearing combined-fault detection and diagnostic. In: 2019 IEEE 12th interna-
tional symposium on diagnostics for electrical machines, power electronics and drives (SDEMPED),
pp. 384–390. https://​doi.​org/​10.​1109/​DEMPED.​2019.​88648​99
Kocyigit N (2015) Fault and sensor error diagnostic strategies for a vapor compression refrigeration system
by using fuzzy inference systems and artificial neural network. Int J Refrig 50:69–79. https://​doi.​org/​
10.​1016/j.​ijref​rig.​2014.​10.​017
Kohavi R, John GH (1997) Wrappers for feature subset selection. Artif Intell 97(1):273–324
Korytkowski P, Malachowski B, Wisniewski T (2009) Genetic algorithm for optimization of inspection sta-
tions allocation in multi-product manufacturing systems. In: 2009 international conference on com-
puters & industrial engineering, Troyes, France, pp 560–565. https://​doi.​org/​10.​1109/​ICCIE.​2009.​
52237​68
Kuhnle A, Jakubik J, Lanza G (2019) Reinforcement learning for opportunistic maintenance optimization.
Prod Eng Res Devel 13(1):33–41. https://​doi.​org/​10.​1007/​s11740-​018-​0855-7
Kulkarni A, Jayaraman VK, Kulkarni BD (2005) Knowledge incorporated support vector machines to
detect faults in Tennessee Eastman Process. Comput Chem Eng 29(10):2128–2133. https://​doi.​org/​10.​
1016/j.​compc​hemeng.​2005.​06.​006
Lee G, Han C, Yoon ES (2004) Multiple-fault diagnosis of the Tennessee Eastman process based on sys-
tem decomposition and dynamic PLS. Ind Eng Chem Res 43(25):8037–8048. https://​doi.​org/​10.​1021/​
ie049​624u
Lee J, Wu F, Zhao W, Ghaffari M, Liao L, Siegel D (2014) Prognostics and health management design
for rotary machinery systems—reviews, methodology and applications. Mech Syst Signal Process
42(1):314–334. https://​doi.​org/​10.​1016/j.​ymssp.​2013.​06.​004
Lee S, Kwak M, Tsui K-L, Kim SB (2019) Process monitoring using variational autoencoder for high-
dimensional nonlinear processes. Eng Appl Artif Intell 83:13–27. https://​doi.​org/​10.​1016/j.​engap​pai.​
2019.​04.​013
Lei Y, He Z, Zi Y, Hu Q (2007) Fault diagnosis of rotating machinery based on multiple ANFIS combina-
tion with GAs. Mech Syst Signal Process 21(5):2280–2294. https://​doi.​org/​10.​1016/j.​ymssp.​2006.​11.​
003
Lei Y, He Z, Zi Y (2009) A Combination of WKNN to fault diagnosis of rolling element bearings. J Vibr
Acoust. https://​doi.​org/​10.​1115/1.​40004​78
Lei Y, Zuo MJ, He Z, Zi Y (2010) A multidimensional hybrid intelligent method for gear fault diagnosis.
Expert Syst Appl 37(2):1419–1430. https://​doi.​org/​10.​1016/j.​eswa.​2009.​06.​060
Lei X, Sandborn P, Bakhshi R, Kashani-Pour A, Goudarzi N (2015) PHM based predictive maintenance
optimization for offshore wind farms. In: 2015 IEEE conference on prognostics and health manage-
ment (PHM), pp 1–8. https://​doi.​org/​10.​1109/​ICPHM.​2015.​72450​27
Lei Y, Li N, Guo L, Li N, Yan T, Lin J (2018) Machinery health prognostics: a systematic review from data
acquisition to rul prediction. Mech Syst Signal Process 104:799–834. https://​doi.​org/​10.​1016/j.​ymssp.​
2017.​11.​016
Leite VCMN, Borges da Silva JG, Veloso GFC, Borges da Silva LE, Lambert-Torres G, Bonaldi EL, de
Lacerda de Oliveira LE (2015) Detection of localized bearing faults in induction machines by spectral
kurtosis and envelope analysis of stator current. IEEE Trans Ind Electron 62(3):1855–1865
Leite GdNP, Araújo AM, Rosas PAC (2018) Prognostic techniques applied to maintenance of wind turbines:
a concise and specific review. Renew Sustain Energy Rev 81:1917–1925. https://​doi.​org/​10.​1016/j.​
rser.​2017.​06.​002
Li B, Chow M-Y, Tipsuwan Y, Hung JC (2000) Neural-network-based motor rolling bearing fault diagnosis.
IEEE Trans Ind Electron 47(5):1060–1069. https://​doi.​org/​10.​1109/​41.​873214

13
 3704 K. T. P. Nguyen et al.

Li Y, Zhang S, Li H, Zhai Y, Zhang W, Nie Y (2012) A fault location method based on genetic algorithm
for high-voltage direct current transmission line. Eur Trans Electric Power 22(6):866–878. https://​doi.​
org/​10.​1002/​etep.​1659
Li G, Hu Y, Chen H, Li H, Hu M, Guo Y, Shi S, Hu W (2016) A sensor fault detection and diagnosis strat-
egy for screw chiller system using support vector data description-based D-statistic and DV-contribu-
tion plots. Energy Build 133:230–245. https://​doi.​org/​10.​1016/j.​enbui​ld.​2016.​09.​037
Li S, Zhou X, Pan F, Shi H, Li K, Wang Z (2017) Correlated and weakly correlated fault detection based on
variable division and ICA. Comput. Ind Eng 112:320–335. https://​doi.​org/​10.​1016/j.​cie.​2017.​08.​035
Li X, Ding Q, Sun J-Q (2018) Remaining useful life estimation in prognostics using deep convolution neural
networks. Reliab Eng Syst Saf 172:1–11. https://​doi.​org/​10.​1016/j.​ress.​2017.​11.​021
Li J, Li X, He D (2019) A directed acyclic graph network combined with CNN and LSTM for remaining
useful life prediction. IEEE Access 7:75464–75475. https://​doi.​org/​10.​1109/​ACCESS.​2019.​29195​66
Liao L (2014) Discovering prognostic features using genetic programming in remaining useful life predic-
tion. IEEE Trans Ind Electron 61(5):2464–2472
Liao L, Köttig F (2014) Review of hybrid prognostics approaches for remaining useful life prediction of
engineered systems, and an application to battery life prediction. IEEE Trans Reliab 63(1):191–207.
https://​doi.​org/​10.​1109/​TR.​2014.​22991​52
Liao L, Jin W, Pavel R (2016) enhanced restricted boltzmann machine with prognosability regularization for
prognostics and health assessment. IEEE Trans Ind Electron 63(11):7076–7083. https://​doi.​org/​10.​
1109/​TIE.​2016.​25864​42
Liao W, Chen M, Yang X (2017) Joint optimization of preventive maintenance and production scheduling
for parallel machines system. J Intel Fuzzy Syst 32(1):913–923. https://​doi.​org/​10.​3233/​JIFS-​161385
Li W, Li M, Fan R, Li L (2010) A fault diagnosis method based on decision tree for wireless mesh network.
In: 2010 IEEE 12th international conference on communication technology, pp 231–234. https://​doi.​
org/​10.​1109/​ICCT.​2010.​56892​72
Ling Z, Wang X, Qu F (2018) Reinforcement learning-based maintenance scheduling for resource con-
strained flow line system. In: 2018 IEEE 4th international conference on control science and systems
engineering (ICCSSE), pp 364–369. https://​doi.​org/​10.​1109/​CCSSE.​2018.​87248​07
Listou Ellefsen A, Bjørlykhaug E, Æsøy V, Ushakov S, Zhang H (2019) Remaining useful life predic-
tions for turbofan engine degradation using semi-supervised deep architecture. Reliab Eng Syst Saf
183:240–251. https://​doi.​org/​10.​1016/j.​ress.​2018.​11.​027
Liu R, Yang B, Zhang X, Wang S, Chen X (2016) Time-frequency atoms-driven support vector machine
method for bearings incipient fault diagnosis. Mech Syst Signal Process 75:345–370. https://​doi.​org/​
10.​1016/j.​ymssp.​2015.​12.​020
Liu R, Yang B, Zio E, Chen X (2018) Artificial intelligence for fault diagnosis of rotating machinery: a
review. Mech Syst Signal Process 108:33–47. https://​doi.​org/​10.​1016/j.​ymssp.​2018.​02.​016
Lu C, Chen J, Hong R, Feng Y, Li Y (2016) Degradation trend estimation of slewing bearing based on
LSSVM model. Mech Syst Signal Process 76–77:353–366. https://​doi.​org/​10.​1016/j.​ymssp.​2016.​02.​
031
Luo L, Bao S, Mao J, Tang D (2017) Fault detection and diagnosis based on sparse PCA and two-level con-
tribution plots. Ind Eng Chem Res 56(1):225–240. https://​doi.​org/​10.​1021/​acs.​iecr.​6b015​00
Mahamad AK, Hiyama T (2008) Development of artificial neural network based fault diagnosis of induction
motor bearing. In: 2008 IEEE 2nd international power and energy conference, Johor Bahru, Malaysia,
pp. 1387–1392
Mahamad AK, Saon S, Hiyama T (2010) Predicting remaining useful life of rotating machinery based artifi-
cial neural network. Comput Math Appl 60(4):1078–1087. https://​doi.​org/​10.​1016/j.​camwa.​2010.​03.​
065
Mahdaoui R, Mouss LH (2012) A TSK-type recurrent neuro-fuzzy systems for fault prognosis. J Softw Eng
Appl 5(7):477–482. https://​doi.​org/​10.​4236/​jsea.​2012.​57055
Majidian A, Saidi MH (2007) Comparison of Fuzzy logic and Neural Network in life prediction of boiler
tubes. Int J Fatigue 29(3):489–498. https://​doi.​org/​10.​1016/j.​ijfat​igue.​2006.​05.​001
Malhi A, Yan R, Gao RX (2011) Prognosis of defect propagation based on recurrent neural networks. IEEE
Trans Instrum Meas 60(3):703–711. https://​doi.​org/​10.​1109/​TIM.​2010.​20782​96
Malhotra P, TV V, Ramakrishnan A, Anand G, Vig L, Agarwal P, Shroff G (2016) Multi-Sensor Prognostics
using an Unsupervised Health Index based on LSTM encoder-decoder. arXiv:​1608.​06154 [cs]
Mansouri M, Nounou M, Nounou H, Karim N (2016) Kernel PCA-based GLRT for nonlinear fault detec-
tion of chemical processes. J Loss Prev Process Ind 40:334–347. https://​doi.​org/​10.​1016/j.​jlp.​2016.​
01.​011

13
A review of artificial intelligence methods for engineering… 3705

Marseguerra M, Zio E, Podofillini L (2002) Condition-based maintenance optimization by means of genetic

algorithms and Monte Carlo simulation. Reliab Eng Syst Saf 77(2):151–165. https://​doi.​org/​10.​1016/​
S0951-​8320(02)​00043-1
Md NN, Che HCR, Hussain MA (2020) A review of data-driven fault detection and diagnosis methods:
applications in chemical process systems. Rev Chem Eng 36(4):513–553. https://​doi.​org/​10.​1515/​
revce-​2017-​0069
Mehrjoo M, Khaji N, Ghafory-Ashtiany M (2013) Application of genetic algorithm in crack detection
of beam-like structures using a new cracked Euler-Bernoulli beam element. Appl Soft Comput
13(2):867–880. https://​doi.​org/​10.​1016/j.​asoc.​2012.​09.​014
Miao H, Li B, Sun C, Liu J (2019) Joint learning of degradation assessment and RUL prediction for aero-
engines via dual-task deep LSTM networks. IEEE Trans Ind Inf 15(9):5023–5032. https://​doi.​org/​10.​
1109/​TII.​2019.​29002​95
Mosallam A, Medjaher K, Zerhouni N (2016) Data-driven prognostic method based on Bayesian approaches
for direct remaining useful life prediction. J Intell Manuf 27(5):1037–1048
Nasiri S, Khosravani MR, Weinberg K (2017) Fracture mechanics and mechanical fault detection by artifi-
cial intelligence methods: a review. Eng Fail Anal 81:270–293. https://​doi.​org/​10.​1016/j.​engfa​ilanal.​
2017.​07.​011
Nguyen KTP, Amor K, Medjaher K, Picot A, Maussion P, Tobon D, Chauchat B, Cheron R (2018) Analysis
and comparison of multiple features for fault detection and prognostic in ball bearings. In: PHM Soci-
ety European Conference, vol 4. Utrecht, The Netherlands, pp 1–6
Nguyen N-T, Lee H-H (2010) Induction motor fault diagnosis based on the k-NN and optimal feature selec-
tion. Int J Electron 97(9):1071–1081. https://​doi.​org/​10.​1080/​00207​217.​2010.​482023
Nguyen KTP, Medjaher K (2019) A new dynamic predictive maintenance framework using deep learning
for failure prognostics. Reliab Eng Syst Saf 188:251–262. https://​doi.​org/​10.​1016/j.​ress.​2019.​03.​018
Nguyen KTP, Medjaher K (2021) An automated health indicator construction methodology for prognostics
based on multi-criteria optimization. ISA Trans 113:81–96. https://​doi.​org/​10.​1016/j.​isatra.​2020.​03.​
017
Nguyen TPK, Yeung TG, Castanier B (2013) Optimal maintenance and replacement decisions under tech-
nological change with consideration of spare parts inventories. Int J Prod Econ 143(2):472–477.
https://​doi.​org/​10.​1016/j.​ijpe.​2012.​12.​003
Nguyen TPK, Castanier B, Yeung TG (2014) Maintaining a system subject to uncertain technological evo-
lution. Reliab Eng Syst Saf 128:56–65. https://​doi.​org/​10.​1016/j.​ress.​2014.​04.​004
Nguyen KTP, Yeung T, Castanier B (2017) Acquisition of new technology information for maintenance and
replacement policies. Int J Prod Res 55(8):2212–2231. https://​doi.​org/​10.​1080/​00207​543.​2016.​12290​
69
Nguyen KTP, Do P, Huynh KT, Bérenguer C, Grall A (2019) Joint optimization of monitoring quality and
replacement decisions in condition-based maintenance. Reliab Eng Syst Saf 189:177–195. https://​doi.​
org/​10.​1016/j.​ress.​2019.​04.​034
Nguyen KTP, Medjaher K, Gogu C (2022) Probabilistic deep learning methodology for uncertainty quantifi-
cation of remaining useful lifetime of multi-component systems. Reliab Eng Syst Saf 222. https://​doi.​
org/​10.​1016/j.​ress.​2022.​108383
Pan D-G, Lei S-S, Wu S-C (2010) Two-stage damage detection method using the artificial neural networks
and genetic algorithms. In: Zhu R, Zhang Y, Liu B, Liu C (eds) Information computing and applica-
tions. Lecture notes in computer science. Springer, Berlin, Heidelberg, pp 325–332. https://​doi.​org/​
10.​1007/​978-3-​642-​16167-4_​42
Pan H, Lü Z, Wang H, Wei H, Chen L (2018) Novel battery state-of-health online estimation method using
multiple health indicators and an extreme learning machine. Energy 160:466–477
Pandya DH, Upadhyay SH, Harsha SP (2012) ANN based fault diagnosos of rolling element bearing using
time-frequency domain feature
Parikh UB, Das B, Maheshwari R (2010) Fault classification technique for series compensated transmission
line using support vector machine. Int J Electr Power Energy Syst 32(6):629–636. https://​doi.​org/​10.​
1016/j.​ijepes.​2009.​11.​020
Park D, Kim S, An Y, Jung J-Y (2018) LiReD: a light-weight real-time Fault detection system for edge
computing using LSTM recurrent neural networks. Sensors 18(7):2110. https://​doi.​org/​10.​3390/​
s1807​2110
Picot A, Obeid Z, Régnier J, Poignant S, Darnis O, Maussion P (2014) Statistic-based spectral indicator
for bearing fault detection in permanent-magnet synchronous machines using the stator current.
Mech Syst Signal Process 46(2):424–441

13
 3706 K. T. P. Nguyen et al.

Refaat SS, Abu-Rub H, Saad MS, Aboul-Zahab EM, Iqbal A (2013) ANN-based for detection, diagnosis
the bearing fault for three phase induction motors using current signal. In: 2013 IEEE interna-
tional conference on industrial technology (ICIT), Cape Town, South Africa, pp 253–258
Riverol C, Pilipovik V (2014) Assessing the failure frequency of potential hazardous incidents using
radial basis function networks (RBFN). A milk pasteurization unit as study case. Food Control
35(1):18–21. https://​doi.​org/​10.​1016/j.​foodc​ont.​2013.​06.​041
Rocchetta R, Bellani L, Compare M, Zio E, Patelli E (2019) A reinforcement learning framework for
optimal operation and maintenance of power grids. Appl Energy 241:291–301. https://​doi.​org/​10.​
1016/j.​apene​rgy.​2019.​03.​027
Ruiz-Sarmiento J-R, Monroy J, Moreno F-A, Galindo C, Bonelo J-M, Gonzalez-Jimenez J (2020) A pre-
dictive model for the maintenance of industrial machinery in the context of industry 4.0. Eng Appl
Artif Intell 87:103289. https://​doi.​org/​10.​1016/j.​engap​pai.​2019.​103289
Sabura Banu U, Uma G (2011) ANFIS based sensor fault detection for continuous stirred tank reactor.
Appl Soft Comput 11(2):2618–2624. https://​doi.​org/​10.​1016/j.​asoc.​2010.​10.​009
Salahshoor K, Khoshro MS, Kordestani M (2011) Fault detection and diagnosis of an industrial steam
turbine using a distributed configuration of adaptive neuro-fuzzy inference systems. Simul Model
Pract Theory 19(5):1280–1293. https://​doi.​org/​10.​1016/j.​simpat.​2011.​01.​005
Samanta B, Al-balushi KR (2003) Artificial Neural Network based fault diagnostics of rolling element
bearings using time-domain features. Mech Syst Signal Process 17(2):317–328. https://​doi.​org/​10.​
1006/​mssp.​2001.​1462
Saridakis KM, Chasalevris AC, Papadopoulos CA, Dentsoras AJ (2008) Applying neural networks,
genetic algorithms and fuzzy logic for the identification of cracks in shafts by using coupled
response measurements. Comput Struct 86(11):1318–1338. https://​doi.​org/​10.​1016/j.​comps​truc.​
2007.​08.​004
Sateesh Babu G, Zhao P, Li X-L (2016) Deep convolutional neural network based regression approach
for estimation of remaining useful life. In: Navathe SB, Wu W, Shekhar S, Du X, Wang XS, Xiong
H (eds) Database systems for advanced applications. Lecture notes in computer science. Springer
International Publishing, Cham, pp 214–228. https://​doi.​org/​10.​1007/​978-3-​319-​32025-0_​14
Saufi SR, Ahmad ZAB, Leong MS, Lim MH (2019) Challenges and opportunities of deep learning mod-
els for machinery fault detection and diagnosis: a review. IEEE Access 7:122644–122662. https://​
doi.​org/​10.​1109/​ACCESS.​2019.​29382​27
Saxena A, Celaya J, Balaban E, Goebel K, Saha B, Saha S, Schwabacher M (2008) Metrics for evaluat-
ing performance of prognostic techniques. In: 2008 international conference on prognostics and
health management, pp 1–17. https://​doi.​org/​10.​1109/​PHM.​2008.​47114​36
Schwall ML, Gerdes JC (2002) A probabilistic approach to residual processing for vehicle fault detec-
tion. In: In Proceedings of the 2002 ACC, pp 2552–2557
Seng Ng Y, Srinivasan R (2010) Multi-agent based collaborative fault detection and identification in chemi-
cal processes. Eng Appl Artif Intell 23(6):934–949. https://​doi.​org/​10.​1016/j.​engap​pai.​2010.​01.​026
Shan P, Peng S, Tang L, Yang C, Zhao Y, Xie Q, Li C (2015) A nonlinear partial least squares with slice
transform based piecewise linear inner relation. Chemom Intell Lab Syst 143:97–110. https://​doi.​
org/​10.​1016/j.​chemo​lab.​2015.​02.​015
Shao H, Jiang H, Zhang X, Niu M (2015) Rolling bearing fault diagnosis using an optimization deep belief
network. Meas Sci Technol 26(11):115002. https://​doi.​org/​10.​1088/​0957-​0233/​26/​11/​115002
Sharma S, Abed W, Sutton R, Subudhi B (2015) Corrosion Fault diagnosis of rolling element bearing
under constant and variable load and speed conditions. IFAC-PapersOnLine 48(30):49–54. https://​
doi.​org/​10.​1016/j.​ifacol.​2015.​12.​352
Shukla S, Yadav RN, Sharma J, Khare S (2015) Analysis of statistical features for fault detection in
ball bearing. In: 2015 IEEE International conference on computational intelligence and computing
research (ICCIC), Madurai, India, pp 1–7
Sikorska JZ, Hodkiewicz M, Ma L (2011) Prognostic modelling options for remaining useful life estimation
by industry. Mech Syst Signal Process 25(5):1803–1836. https://​doi.​org/​10.​1016/j.​ymssp.​2010.​11.​018
Silva RE, Gouriveau R, Jemeï S, Hissel D, Boulon L, Agbossou K, Yousfi Steiner N (2014) Proton exchange
membrane fuel cell degradation prediction based on Adaptive Neuro-Fuzzy Inference Systems. Int J
Hydrogen Energy 39(21):11128–11144. https://​doi.​org/​10.​1016/j.​ijhyd​ene.​2014.​05.​005
Skima H, Varnier C, Dedu E, Medjaher K, Bourgeois J (2019) Post-prognostics decision making in
distributed MEMS-based systems. J Intell Manuf 30(3):1125–1136. https://​doi.​org/​10.​1007/​
s10845-​017-​1312-8
Skordilis E, Moghaddass R (2020) A deep reinforcement learning approach for real-time sensor-driven
decision making and predictive analytics. Comput Ind Eng 147:106600. https://​doi.​org/​10.​1016/j.​
cie.​2020.​106600

13
A review of artificial intelligence methods for engineering… 3707

Sorzano COS, Vargas J, Montano AP (2014) A survey of dimensionality reduction techniques. arXiv:​
1403.​2877 [cs, q-bio, stat]
Soualhi A, Razik H, Clerc G, Doan DD (2014) Prognosis of bearing failures using hidden Markov mod-
els and the adaptive neuro-fuzzy inference system. IEEE Trans Ind Electron 61(6):2864–2874.
https://​doi.​org/​10.​1109/​TIE.​2013.​22744​15
Soualhi M, Nguyen KTP, Soualhi A, Medjaher K, Hemsas KE (2019) Health monitoring of bearing and
gear faults by using a new health indicator extracted from current signals. Measurement 141:37–51
Subramaniyan M, Skoogh A, Muhammad AS, Bokrantz J, Johansson B, Roser C (2020) A data-driven
approach to diagnosing throughput bottlenecks from a maintenance perspective. Comput Ind Eng.
https://​doi.​org/​10.​1016/j.​cie.​2020.​106851
Sun W, Chen J, Li J (2007) Decision tree and PCA-based fault diagnosis of rotating machinery. Mech
Syst Signal Process 21:1300. https://​doi.​org/​10.​1016/j.​ymssp.​2006.​06.​010
Sun H-C, Huang Y-C, Huang C-M (2012) Fault diagnosis of power transformers using computational
intelligence: a review. Energy Procedia 14:1226–1231. https://​doi.​org/​10.​1016/j.​egypro.​2011.​12.​
1080
Sun J, Li C, Liu C, Gong Z, Wang R (2019) A data-driven health indicator extraction method for aircraft
air conditioning system health monitoring. Chin J Aeronaut 32(2):409–416
Tang L, Hettler E, Zhang B, DeCastro J (2011) A testbed for real-time autonomous vehicle PHM and
contingency management applications. In: Annual conference of the Prognostics and Health Man-
agement Society 2011, vol 2, pp 1–11
Tang J, Alelyani S, Liu H (2014) Feature selection for classification: a review. Data classification: algo-
rithms and applications, pp 37–64
Thieullen A, Ouladsine M, Pinaton J (2012) A survey of health indicators and data-driven prognosis in
semiconductor manufacturing process. IFAC Proc Vol 45(20):19–24
Tian Z, Wong L, Safaei N (2010) A neural network approach for remaining useful life prediction utiliz-
ing both failure and suspension histories. Mech Syst Signal Process 24(5):1542–1555. https://​doi.​
org/​10.​1016/j.​ymssp.​2009.​11.​005
Tobon-Mejia DA, Medjaher K, Zerhouni N, Tripot G (2012) A data-driven failure prognostics method
based on mixture of Gaussians Hidden Markov models. IEEE Trans Reliab 61(2):491–503
Tran DAT, Chen Y, Ao HL, Cam HNT (2016) An enhanced chiller FDD strategy based on the combina-
tion of the LSSVR-DE model and EWMA control charts. Int J Refrig 72:81–96. https://​doi.​org/​10.​
1016/j.​ijref​r ig.​2016.​07.​024
Unal M, Onat M, Demetgul M, Kucuk H (2014) Fault diagnosis of rolling bearings using a genetic
algorithm optimized neural network. Measurement 58:187–196. https://​doi.​org/​10.​1016/j.​measu​
rement.​2014.​08.​041
Uraikul V, Chan CW, Tontiwachwuthikul P (2007) Artificial intelligence for monitoring and supervisory
control of process systems. Eng Appl Artif Intell 20(2):115–131. https://​doi.​org/​10.​1016/j.​engap​
pai.​2006.​07.​002
Vakil-Baghmisheh M-T, Peimani M, Sadeghi MH, Ettefagh MM (2008) Crack detection in beam-like
structures using genetic algorithms. Appl Soft Comput 8(2):1150–1160. https://​doi.​org/​10.​1016/j.​
asoc.​2007.​10.​003
Van Weele S, Ramirez-Marquez JE (2012) Optimization of inspection for dual contraband using a
genetic algorithm. J Risk Reliab 226(5):508–525. https://​doi.​org/​10.​1177/​17480​06X12​458386
Verron S, Tiplica T, Kobi A (2006) Fault diagnosis with Bayesian Networks: application to the Tennes-
see Eastman Process. In: 2006 IEEE international conference on industrial technology, pp 98–103.
IEEE, Mumbai, India. https://​doi.​org/​10.​1109/​ICIT.​2006.​372301
Viharos ZJ, Kis KB (2015) Survey on Neuro-Fuzzy systems and their applications in technical diagnos-
tics and measurement. Measurement 67:126–136. https://​doi.​org/​10.​1016/j.​measu​rement.​2015.​02.​
001
Wang S, Liu M (2015) Multi-objective optimization of parallel machine scheduling integrated with
multi-resources preventive maintenance planning. J Manuf Syst 37:182–192. https://​doi.​org/​10.​
1016/j.​jmsy.​2015.​07.​002
Wang WQ, Golnaraghi MF, Ismail F (2004) Prognosis of machine health condition using neuro-fuzzy sys-
tems. Mech Syst Signal Process 18(4):813–831. https://​doi.​org/​10.​1016/​S0888-​3270(03)​00079-7
Wang N, Yuan Z, Wang D (2009) Improving process fault detection and diagnosis using robust PCA and
robust FDA. In: 2009 WRI World Congress on computer science and information engineering, vol 2,
pp 54–59. https://​doi.​org/​10.​1109/​CSIE.​2009.​348
Wang C-H, Cheng H-Y, Deng Y-T (2018) Using Bayesian belief network and time-series model to conduct
prescriptive and predictive analytics for computer industries. Comput Ind Eng 115:486–494. https://​
doi.​org/​10.​1016/j.​cie.​2017.​12.​003

13
 3708 K. T. P. Nguyen et al.

Wang S, Zhang X, Gao D, Chen B, Cheng Y, Yang Y, Yu W, Huang Z, Peng J (2018) A remaining useful
life prediction model based on hybrid long-short sequences for engines. In: 2018 21st international
conference on intelligent transportation systems (ITSC), pp 1757–1762. https://​doi.​org/​10.​1109/​
ITSC.​2018.​85696​68
Wang YS, Liu NN, Guo H, Wang XL (2020) An engine-fault-diagnosis system based on sound intensity
analysis and wavelet packet pre-processing neural network. Eng Appl Artif Intell 94:103765. https://​
doi.​org/​10.​1016/j.​engap​pai.​2020.​103765
Wen L, Dong Y, Gao L (2019) A new ensemble residual convolutional neural network for remaining useful
life estimation. Math Biosci Eng 16(2):862–880. https://​doi.​org/​10.​3934/​mbe.​20190​40
Wu Y, Yuan M, Dong S, Lin L, Liu Y (2018) Remaining useful life estimation of engineered systems using
vanilla LSTM neural networks. Neurocomputing 275:167–179. https://​doi.​org/​10.​1016/j.​neucom.​
2017.​05.​063
Xiao F, Wang S, Zhang J (2006) A diagnostic tool for online sensor health monitoring in air-conditioning
systems. Autom Constr 15(4):489–503. https://​doi.​org/​10.​1016/j.​autcon.​2005.​06.​001
Xiao F, Zhao Y, Wen J, Wang S (2014) Bayesian network based FDD strategy for variable air volume termi-
nals. Autom Constr 41:106–118. https://​doi.​org/​10.​1016/j.​autcon.​2013.​10.​019
Xiao L, Chen X, Zhang X, Liu M (2017) A novel approach for bearing remaining useful life estimation
under neither failure nor suspension histories condition. J Intell Manuf 28(8):1893–1914. https://​doi.​
org/​10.​1007/​s10845-​015-​1077-x
Yan W (2006) Application of random forest to aircraft engine fault diagnosis. In: The proceedings of the
multiconference on “computational engineering in systems applications”, vol 1, pp 468–475. https://​
doi.​org/​10.​1109/​CESA.​2006.​42816​98
Yan W, Yu L (2015) On accurate and reliable anomaly detection for gas turbine combustors: a deep learning
approach 8
Yan K, Zhong C, Ji Z, Huang J (2018) Semi-supervised learning for early detection and diagnosis of various
air handling unit faults. Energy Build 181:75–83. https://​doi.​org/​10.​1016/j.​enbui​ld.​2018.​10.​016
Yang B-S, Di X, Han T (2008) Random forests classifier for machine fault diagnosis. J Mech Sci Technol
22(9):1716–1725. https://​doi.​org/​10.​1007/​s12206-​008-​0603-6
Yazici B, Kliman GB (1999) An adaptive statistical time-frequency method for detection of broken bars and
bearing faults in motors using stator current. IEEE Trans Ind Appl 35(2):442–452
Ye Z, Sadeghian A, Wu B (2006) Mechanical fault diagnostics for induction motor with variable speed
drives using Adaptive Neuro-fuzzy Inference System. Electric Power Syst Res 76(9):742–752. https://​
doi.​org/​10.​1016/j.​epsr.​2005.​10.​011
Yin J, Zhao W (2016) Fault diagnosis network design for vehicle on-board equipments of high-speed rail-
way: a deep learning approach. Eng Appl Artif Intell 56:250–259. https://​doi.​org/​10.​1016/j.​engap​pai.​
2016.​10.​002
Yin S, Zhu X, Jing C (2014) Fault detection based on a robust one class support vector machine. Neurocom-
puting 145:263–268. https://​doi.​org/​10.​1016/j.​neucom.​2014.​05.​035
Yu J (2012) Local and nonlocal preserving projection for bearing defect classification and performance
assessment. IEEE Trans Ind Electron 59(5):2363–2376
Yu B, Zhao H, Xue D (2017) A multi-population co-evolutionary genetic programming approach for opti-
mal mass customisation production. Int J Prod Res 55(3):621–641. https://​doi.​org/​10.​1080/​00207​543.​
2016.​11945​38
Yu W, Kim IY, Mechefske C (2019) Remaining useful life estimation using a bidirectional recurrent neu-
ral network based autoencoder scheme. Mech Syst Signal Process 129:764–780. https://​doi.​org/​10.​
1016/j.​ymssp.​2019.​05.​005
Zaidan MA, Harrison RF, Mills AR, Fleming PJ (2015) Bayesian Hierarchical Models for aerospace gas tur-
bine engine prognostics. Expert Syst Appl 42(1):539–553. https://​doi.​org/​10.​1016/j.​eswa.​2014.​08.​007
Zhang J (2006) Improved on-line process fault diagnosis through information fusion in multiple neural net-
works. Comput Chem Eng 30(3):558–571. https://​doi.​org/​10.​1016/j.​compc​hemeng.​2005.​11.​002
Zhang Y (2009) Enhanced statistical analysis of nonlinear processes using KPCA, KICA and SVM. Chem
Eng Sci 64(5):801–811. https://​doi.​org/​10.​1016/j.​ces.​2008.​10.​012
Zhang F, Ge Z (2015) Decision fusion systems for fault detection and identification in industrial processes. J
Process Control 31:45–54. https://​doi.​org/​10.​1016/j.​jproc​ont.​2015.​04.​004
Zhang Y, Li S, Hu Z, Song C (2012) Dynamical process monitoring using dynamical hierarchical kernel
partial least squares. Chemom Intell Lab Syst 118:150–158. https://​doi.​org/​10.​1016/j.​chemo​lab.​2012.​
07.​004
Zhang Y, An J, Zhang H (2013) Monitoring of time-varying processes using kernel independent component
analysis. Chem Eng Sci 88:23–32. https://​doi.​org/​10.​1016/j.​ces.​2012.​11.​008

13
A review of artificial intelligence methods for engineering… 3709

Zhang C, Lim P, Qin AK, Tan KC (2017) Multiobjective deep belief networks ensemble for remaining use-
ful life estimation in prognostics. IEEE Trans Neural Netw Learn Syst 28(10):2306–2318. https://​doi.​
org/​10.​1109/​TNNLS.​2016.​25827​98
Zhang X, Zhang Q, Chen M, Sun Y, Qin X, Li H (2018) A two-stage feature selection and intelligent fault
diagnosis method for rotating machinery using hybrid filter and wrapper method. Neurocomputing
275(C):2426–2439. https://​doi.​org/​10.​1016/j.​neucom.​2017.​11.​016
Zhang Y, Ye D, Liu Y (2018) Robust locally linear embedding algorithm for machinery fault diagnosis.
Neurocomputing 273:323–332
Zhao F, Chen J, Guo L, Li X (2009) Neuro-fuzzy based condition prediction of bearing health. J Vib Con-
trol 15(7):1079–1091. https://​doi.​org/​10.​1177/​10775​46309​102665
Zhao Y, Wang S, Xiao F (2013) A statistical fault detection and diagnosis method for centrifugal chillers
based on exponentially-weighted moving average control charts and support vector regression. Appl
Therm Eng 51(1):560–572. https://​doi.​org/​10.​1016/j.​applt​herma​leng.​2012.​09.​030
Zhao Y, Wen J, Xiao F, Yang X, Wang S (2017) Diagnostic Bayesian networks for diagnosing air handling
units faults - part I: Faults in dampers, fans, filters and sensors. Appl Therm Eng 111:1272–1286.
https://​doi.​org/​10.​1016/j.​applt​herma​leng.​2015.​09.​121
Zhao Y, Li T, Zhang X, Zhang C (2019) Artificial intelligence-based fault detection and diagnosis meth-
ods for building energy systems: Advantages, challenges and the future. Renew Sustain Energy Rev
109:85–101. https://​doi.​org/​10.​1016/j.​rser.​2019.​04.​021
Zheng J, Pan H, Cheng J (2017) Rolling bearing fault detection and diagnosis based on composite multi-
scale fuzzy entropy and ensemble support vector machines. Mech Syst Signal Process 85:746–759.
https://​doi.​org/​10.​1016/j.​ymssp.​2016.​09.​010
Zhong S, Li Z, Lin L, Zhang Y (2018) Aero-engine exhaust gas temperature prognostic model based on
gated recurrent unit network. In: 2018 IEEE international conference on prognostics and health man-
agement (ICPHM), pp 1–5. https://​doi.​org/​10.​1109/​ICPHM.​2018.​84488​57
Zhou Q, Wang S, Xiao F (2009) A novel strategy for the fault detection and diagnosis of centrifugal chiller
systems. HVAC&R Res 15(1):57–75. https://​doi.​org/​10.​1080/​10789​669.​2009.​10390​825
Zhou Z, Wen C, Yang C (2016) Fault isolation based on k-nearest neighbor rule for industrial processes.
IEEE Trans Ind Electron 63:2578–2586. https://​doi.​org/​10.​1109/​tie.​2016.​25208​98
Zhu Q-X, Meng Q-Q, He Y-L (2017) Novel multidimensional feature pattern classification method and its
application to fault diagnosis. Ind Eng Chem Res 56(31):8906–8916. https://​doi.​org/​10.​1021/​acs.​iecr.​
7b000​27
Zio E, Di Maio F (2010) A data-driven fuzzy approach for predicting the remaining useful life in dynamic
failure scenarios of a nuclear system. Reliab Eng Syst Saf 95(1):49–57. https://​doi.​org/​10.​1016/j.​ress.​
2009.​08.​001

Publisher’s Note  Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor holds exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is
solely governed by the terms of such publishing agreement and applicable law.

13