Data-Driven - Methods - For - Predictive - Maintenance - of - Industrial - Equipment - A - Survey

IEEE SYSTEMS JOURNAL, VOL. 13, NO.
3, SEPTEMBER 2019 2213
Data-Driven Methods for Predictive Maintenance of

Industrial Equipment: A Survey
Weiting Zhang , Dong Yang , Member, IEEE, and Hongchao Wang , Member, IEEE
Abstract—With the tremendous revival of artificial intelligence, process [2]. Then, Germany presented “Industry 4.0” [3] at the
predictive maintenance (PdM) based on data-driven methods has Hannover Messe, in April 2013, and the core idea of its initiative
become the most effective solution to address smart manufacturing is to build smart factories, smart production, and smart logis-
and industrial big data, especially for performing health perception
(e.g., fault diagnosis and remaining life assessment). Moreover, tics. Moreover, the government of China proposed the “Made
because the existing PdM research is still in primary experimental in China 2025” strategy [4], which is particularly a significant
stage, most works are conducted utilizing several open-datasets, strategic initiative in the new round of technological revolution.
and the combination with specific applications such as rotating Its core meaning is to realize the transformation and upgrading
machinery is especially rare. Hence, in this paper, we focus on data- of manufacturing, that is, to achieve intelligence.
driven methods for PdM, present a comprehensive survey on its
applications, and attempt to provide graduate students, companies, With the arrival of “Industry 4.0,” the prognostic and health
and institutions with the preliminary understanding of the existing management (PHM) concept has become the inevitable ten-
works recently published. Specifically, we first briefly introduce the dency in the context of smart manufacturing and industrial
PdM approach, illustrate our PdM scheme for automatic washing big data [5], and it provides a reliable solution for managing
equipment, and demonstrate the challenges encountered when we the health status of equipment. In addition, the PdM strategy
conduct a PdM research. Second, we classify the specific industrial
applications based on six algorithms of machine learning and deep for industrial equipment can accurately perceive performance
learning (DL), and compare five performance metrics for each degradation since it was designed to achieve near-zero failures,
classification. Furthermore, the accuracy (a metric to evaluate the near-zero hidden dangers, near-zero accidents, and near-zero
algorithm performance) of these PdM applications is analyzed in pollution throughout the entire manufacturing process [6]. In
detail. There are some important conclusions: 1) the data used in particular, the prediction of machine health can, not only signif-
the summarized literature are mostly from public datasets, such as
case western reserve university (CWRU)/intelligent maintenance icantly reduce the unexploited downtime and expensive labor
systems (IMS); and 2) in recent years, researchers seem to focus costs but also ensure safe operation and optimize the mainte-
more on DL algorithms for PdM research. Finally, we summarize nance plan.
the common features regarding our surveyed PdM applications According to the review of current research results [6], the
and discuss several potential directions.
PdM methods are mainly divided into the following three cat-
Index Terms—Artificial intelligence (AI), deep learning (DL), egories: 1) model-based prognosis; 2) knowledge-based prog-
fault diagnosis, machine learning (ML), predictive maintenance nosis; and 3) data-driven prognosis. Notably, the data-driven
(PdM), remaining life assessment. PdM technology has attracted wide attention. Specifically, in
[7], the PdM methods of mechanical equipment were compre-
I. INTRODUCTION hensively studied and sorted out from the aspect of data acquisi-
ITH the requirement of economic considerations and tion, data processing, and decision support. A data-driven intel-
W the advancement of technology, many countries, includ-
ing the United States, Germany, and China, have successively
ligent PdM system was proposed in [8] for reaching zero-defect
manufacturing.
put forward plans to revitalize manufacturing [1]. Specifically, In recent years, industrial wireless sensor networks (IWSNs)
the United States deployed the “smart manufacturing” strategy, [9], [10] and industrial cyber physical systems [11] have be-
which is implemented by introducing the cyber physical system come an emerging data acquisition technology in the complex
(CPS, a system that links the cyber world of computing and industrial environment, and mechanical data can be collected
communication with the physical world) to the manufacturing using various types of sensors with high reliability and in a
real-time manner [12]. Obviously, because of the continuous
improvement of data acquisition ability [13], as well as the
Manuscript received September 17, 2018; revised January 28, 2019; accepted exponential growth of data volume [14], data-driven methods
March 11, 2019. Date of publication May 6, 2019; date of current version August for health monitoring have achieved great success and received
23, 2019. This work was supported in part by the National Key Research and widespread attention regarding the PdM of industrial equip-
Development Program of China under Grant 2018YFB1702001, in part by the
National Natural Science Foundation of China under Grant 61771040, and in ment. Therefore, since artificial intelligence (AI) algorithms
part by the Fundamental Research Funds for the Central Universities under have achieved important progress during the past five years,
Grant 2018YJS013. (Corresponding author: Dong Yang.) we considered the relevance of performing a summary study
The authors are with the Beijing Jiaotong University, Beijing 100044, China
(e-mail:,[email protected]; [email protected]; [email protected]). on guiding the selection of algorithms in specific applications,
Digital Object Identifier 10.1109/JSYST.2019.2905565 particularly the corresponding significance of AI to PdM.
1937-9234 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Badan Riset Dan Inovasi Nasional. Downloaded on September 13,2023 at 06:05:09 UTC from IEEE Xplore. Restrictions apply.
2214 IEEE SYSTEMS JOURNAL, VOL. 13, NO. 3, SEPTEMBER 2019
ral network (NN) models) avoids the above complex feature

engineering and can be learned using an end-to-end learning
manner, which is implemented by adding deep layers between
the raw data and the prediction result. Thus, the deep models
can be deemed as a “black box,” which output the prediction
result from the input directly, and it is the essential difference
between ML and DL. For all of these considerations, both ML
and DL have been, especially, widely used in the application of
PdM.
Accordingly, we concentrate on data-driven methods in the
PdM and present a related comprehensive survey on their ap-
plications from the aspects of traditional ML and DL. Our
objective is principally attempting to provide newly admitted
graduate students and newly established companies and institu-
tions with a preliminary understanding of the data-driven PdM,
Fig 1. Flowchart of the data-driven method for predictive maintenance (PdM). as well as the research status, methods, and effect of existing
works published in the past five years. We note that there are
several studies in the literature [19]–[21] related to this paper.
However, in contrast to the methods we are concerned about,
fault diagnosis techniques were reviewed in [19] from signal-
and model-based perspectives. Moreover, Zhao et al. [20] and
Liu et al. [21] not only reviewed the specific applications but
also summarized the algorithm itself. Specifically, Zhao et al.
[20] mainly focused on a deep NN (DNN), while Liu et al.
[21] paid more attention to traditional ML algorithms. Nev-
Fig. 2. Flow of ML and DL. ertheless, in this paper, we conduct a comprehensive review
on PdM applications from the aspects of ML and DL. In ad-
dition, we focus on those works with the highest and lowest
The increased availability of industrial data has paved the accuracy in each algorithm and present our insights about the
way for the development and deployment of the data-driven situations.
PdM, which utilizes cutting-edge computational methods to The contributions of this paper are as follows. First, a compre-
provide valuable information regarding the status of equipment hensive review of the PdM is conducted from four aspects, in-
acquired from the growing volume of operational data [15]. cluding the definition of the PdM, significance of the data-driven
As shown in Fig. 1, the data-driven PdM system consists of PdM, specific implementation methods of the data-driven PdM,
two phases: first, a learning process (i.e., model training) is challenges in the implementation process. Second, the PdM ap-
needed on the basis of historical raw sensor signals; second, plications of the six algorithms are compared from ML and DL
the trained model is applied to predict targets and make de- perspectives, respectively, to provide graduate students, com-
cisions. In general, each phase consists of the following three panies, and institutions with the preliminary understanding of
subprocesses: the existing works recently published. Third, on the basis of
1) data acquisition and preprocessing, which can be single- the above comparison, accuracy, which is the most widely used
sensory or multisensory; evaluation metric, is analyzed in detail, and the corresponding
2) feature engineering, which contains feature extraction, conclusions are drawn. In addition, some potential research di-
concatenation, and selection; and rections are provided consequently.
3) model training and predicting, in which a well-trained The structure of this paper is shown in Fig. 3, and the rest of
model will be generated with the optimal parameters. this paper is organized as follows. Section II presents the back-
After that, the model can predict the real-time data flow. ground and recent research on the PdM and data-driven PdM,
Moreover, the data-driven PdM has been extensively applied to provides a specific case of the data-driven PdM, and summa-
industrial manufacturing using machine learning (ML) [16] and rizes some challenges in practice. Sections III and IV introduce
deep learning (DL) algorithms [17]. Specifically, as shown in several applications of the data-driven PdM from ML (e.g., LR,
Fig. 2, traditional ML algorithms, (e.g., logistic regression (LR), SVM, DT and RF) and DL (e.g., artificial NN (ANN), DNN,
support vector machine (SVM), decision tree (DT), and random and auto-encoder (AE)) perspectives, respectively. In specific,
forest (RF)) generally require collecting large amount of data Section III provides a review of several mainstream ML algo-
from the health conditions and various failure status scenarios rithms for PdM applications. Then, Section IV reviews the NN
for model training. Next, feature engineering is conducted from and its derived models in PdM applications. Furthermore, Sec-
the time, frequency, and time-frequency domain [18], [19], and tion V provides several summarizations and an analysis of the
the representation learning of the device health is performed accuracy. Finally, the future research directions and conclusion
using the extracted features. However, DL (i.e., various neu- are given in Section VI and Section VII.
ZHANG et al.: DATA-DRIVEN METHODS FOR PREDICTIVE MAINTENANCE OF INDUSTRIAL EQUIPMENT: A SURVEY 2215
Fig. 4. PdM of degraded equipment [5], [6].
equipment, which is achieved by comprehensively utilizing the

latest research results of modern information technology and AI
technology [6]. In addition, the PdM is an important part of the
PHM system.
According to [23], maintenance methods are mainly divided
into three categories, namely, run-to-failure, preventive main-
tenance, and PdM. Regarding these methods, the first one is
the simplest. That is, the new component runs from installation
until failure occurs and then stops the machine for maintenance.
However, because of the extra maintenance expenditures and
unexpected downtime, the cost and efficiency of this method
are also the most unreasonable. The second method refers to
the failures that can be prevented, achieved by regular replace-
ment of components, but leading to additional operating costs
Fig. 3. Condensed structure of this survey.
and an increased unexploited lifetime [24]. In addition, the third
method is generally indicated based on the assessment of the
TABLE I health status of key components, regardless of the maintenance
TERMINOLOGIES AND ABBREVIATIONS
status, and its fundamental purpose is to make predictability
achievable. That is, the incipient problems that may evolve into
catastrophic failures can be predicted accurately, and effective
measures can be applied to avoid these failures on the basis
of the prediction results. This approach can not only minimize
maintenance costs but can also extend the useful life of the
equipment [25].
It is assumed that the industrial equipment can be repaired
at an appropriate time (i.e., before the fault occurs), as this
approach will restore the apparatus to its original condition
after each maintenance is completed. As shown in Fig. 4, the
curve in the figure demonstrates the performance trend of a
monitoring object, such as a component, device, or system. The
blue curve indicates an ideal maintenance where the repaired
Table I lists the terminologies and abbreviations appeared RUL (green) is basically identical. However, the reddish brown
in the following sections, which will be omitted afterward. curve indicates a nonideal maintenance, either too early or too
late, and its RUL (red) is different from that of the former.
The dotted lines in black and red represent the performance
II. PREDICTIVE MAINTENANCE threshold of the object in each operating cycle. In addition, it
reflects that the PdM of industrial equipment is moving toward
A. Predictive Maintenance
an extremely precise, efficient, and intelligent direction, even
In the late 20th century, PHM was first put forward for the with slight performance degradation or safety risks, which can
development of military projects [22]. However, in recent years, result in serious consequences. The health status of a component,
the PHM system has become a surefire solution for manag- machine, or system can be obtained at any time, and the failure
ing the health status of equipment (e.g., fault diagnosis and can be predicted and prevented to achieve near-zero downtime
remaining useful life (RUL) estimation), especially industrial performance [26]. In summary, the PdM focuses on utilizing
acquisition; 3) feature engineering; and 4) modeling. It is impor-

tant to note that feature engineering can be saved when employ-
ing DL algorithms. In addition, limited by computing power,
storage capacity, and time delay, the training and deployment
of the model need to be completed in different platforms [30].
Specifically, model training is usually carried out on the cloud
or server side because it requires strong computing power and
storage space as support [31]. However, the trained-down model
does not require much computing power and storage space to
predict the real-time dataflow. Accordingly, to maximize the
utilization of transmission and computing resources, the model
will be deployed on the edge nodes, such as the edge server or
edge gateway.
In summary, a considerable number of experimental studies
have proven that it not only results in better effects but also sim-
plifies the construction process of the data-driven PdM system.
Therefore, this paper will review the specific PdM application
from the perspectives of ML and DL.
C. Specific Case Based on Data-driven PdM

As depicted in the previous sections, to overcome the limita-
tions of model-based methods [32], data-driven methods have
been increasingly applied to the PdM [33]. As shown in Fig. 6,
we divide the PdM system into five parts: 1) a physical layer; 2)
perception layer; 3) signal analysis layer; 4) performance predic-
tion layer; and 5) decision-making layer, which are specifically
utilized to perform the PdM of automatic washing equipment
(a machine used for the cleaning stage before the maintenance
of a high-speed rail). It can be seen that such a system can
properly implement a production and maintenance strategy, and
Fig. 5. Implementation process of the data-driven PdM. it will lead to a longer running time and the sufficient use for
key components [34], such as bearings [35], blades [36], [37],
and driving motors [38], [39]. In addition, the effective degra-
predictive information to properly schedule future maintenance dation assessment for components is beneficial to reduce the
operations [6]. Accordingly, deploying the PdM to ensure the production downtime and maintenance cost [40].
safety of operations is of great significance [27]. However, because of different conditions such as lubrication,
load, and speed [41], the failure modes of the components are
also different. Therefore, in actual applications, the fault type
B. Data-Driven PdM
and degree need to be evaluated using signal analysis to deter-
At present, the proportion of maintenance expenditure in the mine the best time for the implementation of the maintenance
operating costs of enterprises is increasing, which is caused by strategy [42]–[44]. According to [45]–[49], in our research, four
inefficient maintenance methods (insufficient maintenance or types of sensors (i.e., vibration, temperature, electrical signal,
excessive maintenance). As a result, it is necessary to formu- and rotating speed) are taken as input signals to represent the
late a reasonable maintenance strategy that not only reduces the equipment performance. As a complementary approach, Jin
waste caused by excessive maintenance but also ensures that et al. [45] and Jie et al. [46] considered that vibration moni-
the equipment is adequately maintained and remains in good toring can provide the best means for failure identification, and
working condition [6]. A new MATLAB toolbox DB-KIT was both studies employed vibration signal to perceive the equip-
proposed in [28], which contains a series of basic and advanced ment status. However, Mobley [47] deemed that signal analysis
algorithms and auxiliary tools for the data-driven PdM. Obvi- of vibration cannot provide all the necessary information for the
ously, with the rise of AI once again, data-driven methods for fault diagnosis of mechanical equipment. Consequently, electric
fault diagnosis and RUL prediction have been the popular issues signals (e.g., current and voltage) in [48] and [49] were applied
in PdM system research. to identify the defects in equipment operation.
As shown in Fig. 5, the implementation process of the data- To construct the PdM system of automatic washing equip-
driven PdM is described in detail, and it is basically consistent ment, a set of data acquisition system is developed, as shown
with the design methodology of [29]. Overall, the process can in the submodule of Fig. 6, including three subgraphs. The left
be divided into four stages: 1) operational assessment; 2) data figure demonstrates a data acquisition device, which is utilized
TABLE II
CHALLENGES IN THE CONSTRUCTION OF THE PDM SYSTEM
general, issues such as time and economic costs need to be

considered first in the implementation process. The balance be-
tween the cost of introducing a PdM system and the loss of
equipment failure should be a better tradeoff. More importantly,
considering safety issues, the specific implementation will be
subject to various restrictions, such as the location of sensor
installation.
Most notably, data-driven methods make data acquisition an
indispensable part for the complete PdM design [52], provid-
Fig. 6. Framework of the PdM. When using DL, the raw signals can be input ing valuable information for specific tasks. Specifically, for aca-
directly into the model; when using ML, feature extraction is required. demic research, many groups choose to purchase existing equip-
ment to collect data. However, for engineering applications, a
personalized data acquisition system needs to be established to
to gather, store, and forward the collected data. The other two meet specific requirements. It has been proven that data anal-
graphs contain four type sensors, which measure acceleration, ysis at different stages of the machine life cycle can not only
temperature, current, and rotational speed, respectively. It can be process data or information more effectively but also achieve
seen that data acquisition plays a crucial role in the realization transparency of the health status of industrial equipment.
process of the PdM. In addition, how to address the obtained data effectively
(i.e., whether to use feature engineering or not) has become
a major challenge for algorithm selection. Here, we make
D. Challenges in Practice
a demonstration of common statistical features in Table III.
As mentioned above, the advantages of the PdM are self- The performance difference of the algorithms can be re-
evident. In recent years, numerous mature methods have been flected in specific applications, especially when faced with
formed, such as the methods in [50] and [51]. However, as far as industrial sensor data. Accordingly, the PdM applications are
the current literature is concerned, a large part of the literature summarized from the perspectives of ML and DL in the
on the data-driven PdM are based on some publicly available following.
datasets. In practice, it needs a comprehensive analysis from the Moreover, model building is also not to be underestimated,
perspectives of operational assessment, data acquisition, feature and the selection, training, and optimization of the model all
engineering, and model building. As a result, there are still many deserve more attention. In practice, supervised learning has
challenges to overcome to reach a mature application level. achieved remarkable results in many fields, but it requires data
On the basis of the design experience, we summarize and to be correctly annotated, which is also an enormous challenge
present these challenges in Table II. for the PdM. With the quantitative tendency of industrial big
Operational assessment is extremely critical, and it de- data, unsupervised learning has naturally become a significant
termines the success or failure of the PdM in a sense. In promising research direction.
TABLE III
SUMMARY OF STATISTICAL FEATURES IN TIME DOMAIN AND FREQUENCY DOMAIN ACCORDING TO [53]–[55]
III. ML FOR PDM Then, the loss function of LR is defined as:

m

ML is well known as the core technology of AI, and algo-
rithms play a particularly important role in its gradual rise. At L (θ) = (hθ (xi ))y i (1 − hθ (xi ))1−y i (2)
i=1
the same time, with the massively available data used to solve
various problems, ML has been widely used in computer science where m refers to the sample number, and xi and yi indicate the
and other fields, such as the PdM of industrial equipment, which ith sample and the corresponding label.
is one of the potential application fields for data-driven methods For the convenience of calculation, there is the following
(e.g., LR, SVM, DT, and RF) [56], [57]. As a result, although likelihood function:
increasingly high-performance algorithms are continually de- m

veloped, generally, employing simple and efficient methods is l (θ) = lnL (θ) = [yi ln (hθ (xi ))
considered first. i=1
In this section, we will mainly summarize the applica- + (1 − yi ) ln (1 − hθ (xi ))] . (3)
tion of these algorithms in the PdM. The comparison and
analysis of the reviewed literature are also given in Table IV. The backpropagation and parameter updating can be defined
as follows:
m
∂y 1
A. LR for PdM J (θ) = [hθ (xi ) − yi ] xji (4)
∂θj m i=1
1) LR Model: LR is a well-known classification model in
m
ML with the lowest algorithm complexity. It belongs to super- 1
θj := θj − α [hθ (xi ) − yi ] xji (5)
vised learning; therefore, the collected data must have corre- m i=1
sponding labels to be fed into the model. In addition, the LR
model takes a linear combination of features as its input and where xji indicates the jth feature of the ith sample, and α
applies a nonlinear function to conduct mapping, so that each represents the hyperparameter of learning rate.
output will fall within the range of (0, 1) and a probabilistic in- 2) PdM Applications of LR: This section will review several
terpretation can be obtained. Consequently, when we acquire a PdM applications related to LR.
large number of labeled features and have critical requirements Li et al. [59] proposed a method that combines an LR model
on model complexity, we can consider using the LR model to with the acoustic emission and cutting force signals; both of
solve these problems [58]. them were utilized to monitor the wearing process of cutting
The prediction function of the LR model is as follows: tools and determine the best maintenance time. A simple ap-
1 plication was introduced in [60] for elevator door health moni-
hθ (x) = gθ θT x = (1) toring by utilizing the LR model. Specifically, wavelet package
1 + e−θ T x
energies were used as features, Fisher’s criteria were used to
where θT x = ni=0 θi xi = θ0 + θ1 x1 + · · · + θi xi , x the input select critical features, and then the maximum likelihood es-
feature of this model, θ the internal parameter, which consists timation was applied to determine the parameters of the LR
of weights and biases, and g (·) a nonlinear transformation of model. Finally, a 100% diagnostic result was obtained. A data-
x, named logistic function. driven modeling method was demonstrated in [15] for the health
TABLE IV
SUMMARY OF THE DATA-DRIVEN PDM METHOD USING THREE TYPES OF ML ALGORITHMS
condition monitoring of gas circulator units by using a combi- in [64]. A dynamic regression model was developed according
nation of LR and L2-SVM models. Pandya et al. [61] focused to [65] for the health prognosis of rolling bearings, and ex-
on fault diagnosis of rolling bearings by using multinomial LR, cellent prognostic accuracy was obtained. In addition, Phillips
and the two features of energy and kurtosis were extracted us- et al. [66] argued that the LR model can be easily provided
ing wavelet packet decomposition. In addition, by a comparison to industrial experts with interpretability and that its predic-
with an ANN and an SVM, the multi-LR model was considered tive performance outperformed that of the ANN and SVM in
a more effective classifier with accuracy of 100% and compu- terms of predicting whether mechanical equipment or compo-
tational time of only 0.33 s. Similarly, Caesarendra et al. [62] nents run properly. A novel prognostic method combination of
proposed a hybrid method to estimate the performance degrada- an LR with manifold regularization was proposed in [67] to
tion from incipient failures, which combines a relevance vector assess the tool health state. Moreover, it is worth mentioning
machine (RVM) with LR. However, when the selected kernel that a penalization regularization method, which has attained
width in the simulated and experimental data is improper, the excellent learning performance, was designed to select effective
overfitting phenomenon occurs. A vibration-based monitoring prognostic features.
approach for the online detection of punch failure in a mi-
cropiercing process was proposed in [63], which utilized the
B. SVM for PdM
statistical overlap factor to select features and obtained more
than 98.6% prediction accuracy on three datasets. To monitor 1) SVM Model: Typically, the SVM model is used to tackle
the latent degradation level and track the failure progress, a the tasks for binary classification. In the PdM of industrial equip-
model that combined a Kalman filter with LR was presented ment, SVMs have been widely applied for identifying a specific
status based on the acquired signal [68]. Moreover, because of indicators from a vibration signal; an SVM model was then
the diversity of fault types and the ability of mapping low di- used as an effective solution to assess the degradation states,
mension features to hyperplanes, the SVM model can be utilized and an SVR model was employed to estimate the RUL of de-
to solve multiclass tasks. graded bearings. In contrast, feature extraction was conducted
In summary, the primary purpose of SVM is to find a hy- in [73] from the time domain of vibration signals, and a multi-
perplane and divide data points correctly on both sides of the class SVM was deployed to detect bearing faults. A multimodal
hyperplane, and the optimization object is represented by deep support vector classification model was developed in [74]
to diagnose the failure of a gearbox, and the performance was
1
argmax min yi wT · xi + b proven to outperform that of the basic SVM. However, when
w ,b w
detecting the faults of a gearbox, the identification capacity of

s.t. yi wT · xi + b ≥ 1 (6) an SVM is enormously influenced by the kernel function and its
parameters. Therefore, a parameter optimization method based
where (xi , yi ) refers to the sample that contains the feature and on the artificial bee colony algorithm was proposed in [75] to
label. ensure accurate identification. An SVM model with four kernel
Therefore, the optimization problem can be transformed to functions was applied in [76] for fault recognition, and statisti-
the following: cal features were extracted from vibration signals under normal
N and faulty conditions of a rotating mechanical system. Addi-
1
max w2 + C ξi tionally, in order to better represent the machine status, a DT
w ,b 2
i=1 algorithm was utilized to select the essential features. A hybrid
SVM model optimized by the intercluster distance in the fea-
s.t. yi wT · xi + b ≥ 1 − ξi (7)
ture space was developed in [77] to extract the multiple temporal
where C is a parameter that controls the weight of the relaxation scale features from a vibration signal and was used to identify
factor, and 0 ≤ α ≤ C. the fault types and their severity with respect to motor bearings.
Then, the Lagrange multiplier is applied to the linear SVM The above works are all about the research of SVMs for fault
objective function and its constraints. The optimization problem classification and linear regression. However, SVMs can also
of nonlinear SVM is described as follows: be used for nonlinear regression. For instance, Nieto et al. [78]
N N N applied an SVM in a nonlinear model to predict the remaining
1
min αi αj yi yj Ø (xi ) Ø (xj ) − αi life of aircraft engines.
α 2 i=1 j =1 i=1
N
C. DT and RF for PdM
s.t. αi yi = 0 (8) 1) DT and RF Model: DT classifiers have achieved great
i−1 success in various fields such as character recognition, medi-
where Ø(·) is a spatial transformation. cal diagnosis, and speech recognition. Most significantly, a DT
2) PdM Applications of SVM: Several PdM applications us- model has the ability to decompose a complex decision-making
ing SVMs or their variants will be reviewed in this section. process into a collection of simpler decisions by recursively
Widodo et al. [69] focused on fault diagnosis of bearings by partitioning the covariate space into subspaces, thus providing
using a multiclass RVM and an SVM; the data were acquired a solution that is prone to interpretation [79]. Furthermore, RF
from the low-speed bearing test rig by acoustic emission sensors is an ensemble learning algorithm composed of multiple DT
and accelerometers. Component analysis was also introduced to classifiers, and the category of its output is determined jointly
extract useful features and reduce the dimension of the original by these individual trees [80]. The RF is provided with many
features. In addition, it was shown that an SVM outperformed significant advantages. For instance, it can handle high dimen-
an RVM for the diagnosis of low-speed bearings. A multiple sional data without feature selection; trees are independent of
classifier (MC) methodology based on an SVM was proposed each other during training process, and the implementation is
in [23] for the PdM of semiconductor manufacturing (e.g., re- relatively simple; in addition, the training speed is usually fast,
placing tungsten filaments used in ion implantation). According and at the same time the generalization ability is strong enough.
to the comparison result, the MC-PdM-SVM system was able To construct a DT, the first problem is to determine which fea-
to achieve better performance than that of the kNN and RBF ap- ture should be divided first, that is, which feature on the current
proaches in terms of unexpected breaks and maintenance costs. dataset plays a decisive role in classification. Using the infor-
In order to extract useful information from raw sensor signals, mation theory to measure information is an effective method.
an autoregressive moving average model with six data-driven al- Information entropy is actually the mathematical expectation of
gorithms was employed in [70] to formulate airline maintenance the amount of information in a random variable. To calculate
scheduling, and the SVM obtained the best overall results. Ac- entropy, we need to calculate the expected value of the informa-
cording to [71], four classifiers were compared, namely, SVM, tion contained in all possible values of all categories, which can
DT, RF, and kNN. Among them, the SVM classifier obtained be obtained by the following formula:
the best performance on four operating systems, and the ac- n

curacy of the SVM was 96.6%, 98.7%, 98%, and 96.6%. The H=− p (xi ) log2 p (xi ) (9)
Hilbert-Huang transform was utilized in [72] to extract health i=1
where p(xi ) is the probability of choosing xi , and n indicates A. ANN for PdM
the feature number. 1) ANN Model: Inspired by the biological neural network,
In addition, the purity of Gini is also an indicator to measure ANN is designed to address nonlinear problems. It is a mas-
the purity of information. The formula is as follows: sively parallel computing system consisting of an extremely
large number of simple processors with many interconnections.
n
Instead of following a set of rules prescribed by human experts,
Gini (D) = 1 − p2i (10)
ANNs appear to learn underlying rules from the given collection
i=1
of representative examples [91]. The ability of ANN models to
automatically learn from examples makes them attractive and
where D refers to the dataset.
extensively applied. Moreover, compared with traditional data-
2) PdM Applications of DT and RF: In this section, we will
driven methods, NNs have obvious advantages in addressing
review several PdM applications of DT and RF models.
fuzzy data, random data, and nonlinear data. They are especially
According to [81], a DT algorithm was deployed to perform
suitable for systems with a large scale, a complex structure, and
condition monitoring of a spur gear. The RMS value, sum, skew-
unclear information. Here, ANN refers to an NN with a single
ness, minimum value, and variance were selected from the entire
hidden layer. Since the DNN was also gradually derived from
feature set as health indicators, which minimized the involve-
it, we therefore summarize ANNs in the DL section.
ment and demand of human expertise. An improved diagnosis
An ANN model consists of two crucial elements: 1) linear
method was employed in [82] for a practical variable refrigerant
summation and 2) nonlinear activation, which can be described
flow system, which was implemented by combining a DT model
as:
with virtual sensor-based fault indicators, and it outperformed
the CART, RF, and generalized boosted regression. A parallel M

[l] [l] [l−1] [l]
RF algorithm was developed in [33] by using the MapReduce Xj = f Wij · Xi + Bj (11)
framework to predict tool wear in dry milling operations. An i=1
RF-based model was employed in [18] to carry out real-time
PdM for wind turbines. According to the obtained results, the where l = 1, 2, indicates the hidden layer and output layer,
depth of the trees had a significant effect on the diagnostic ac- respectively, i the index of input feature maps, and j the
curacy. In addition, Santos et al. [83] conducted a comparison index of the output feature maps. Furthermore, the “·” de-
about an intelligent diagnostic model of wind turbines, and a notes a multiplication operation that is applied to the ith fea-
[l−1]
cost-sensitive classifier using a rotation forest ensemble with ture map Xi from the (l − 1)-th layer, using the ith filter
[l]
the C4.4 algorithm was the best model in this research. An Wij . In addition, a nonlinear activation function f is applied
RF-based method was utilized in [84] to monitor the health to each element of all the feature maps. The rectified linear
condition of bioreactors from a penicillin production process. unit (ReLU) f (z) = max(0, z), the hyperbolic tangent func-
In terms of diagnostic accuracy and generalization ability, the tion f (z) = [exp(z) − exp(−z)]/[exp(z) + exp(−z)], and the
overall performance was certified to be better than that of the logistic function f (z) = 1/[1 + exp(−z)] are the favored op-
SVM and ANN. Moreover, RF is easy to employ and can gener- tions.
alize to unseen data. However, in order to obtain optimal perfor- 2) PdM Applications of ANN: Next, we will introduce some
mance, the SVM and ANN require extra time and computing re- PdM applications based on ANNs.
sources to tune various parameters. A deep RF fusion technique The simple yet widely used ANN model was adopted in
was utilized in [85] to improve the diagnostic performance of [45] for diesel engine fault classification. Since the data were
gearboxes. collected from the same machine, the selected features can ef-
fectively represent the fault information. Consequently, all the
training, validation, and test results in this paper had a diagnos-
IV. DL FOR PDM tic accuracy of 100%. A feedforward NN prediction model was
Because of the lack of expressive ability and the defect of built in [17] to ensure precision estimating on laser welding sta-
dimensionality [86], traditional ML algorithms require prior ex- tus. Additionally, an ANN-based model was presented in [92]
perience and expertise to perform fault representation [87]. The to improve RUL prediction of bearing failures. A hierarchical
theory of DL was first proposed by Hinton and Salakhutdinov NN structure was used in [93] to perform the fault classifica-
in 2006 [88]. In the past decade, DL, a method of extracting tion of bearings. It was composed of two steps, and each step
structured information from massive datasets using layered ML consisted of a two-layer NN to assure the optimum recognition
algorithms, has made almost unimaginable advances in AI and performance. Similarly, a two-stage learning approach was pro-
technology industries [20]. Moreover, it has become an indis- posed in [94] to construct a fault diagnosis framework. First,
pensable technology in the data-driven PdM. Compared with an unsupervised two-layer NN, called sparse filtering, was uti-
the traditional shallow model, DL can automatically extract fea- lized to learn representative features extracted from mechanical
tures from the original data and accurately identify the health vibration signals. Second, health status identification was per-
status [89], [90]. Accordingly, this paper will review the most formed using a two-layer network, called softmax regression.
classic and widely used DL models, including ANN, DNN, and An ANN-based model was adopted in [95] to perform condition
AE. In addition, the comparisons of these models are given in monitoring and fault diagnosis of rolling bearings, and features
Table V. were extracted on the basis of empirical mode decomposition
TABLE V
SUMMARY OF THE DATA-DRIVEN PDM METHOD USING THREE TYPES OF DL ALGORITHMS
energy entropy. Finally, the proposed method showed that an presented in [100] for fault diagnosis of rolling bearings and
ANN can be utilized as an effective tool for bearing degrada- planetary gearboxes; it has obtained superior diagnostic accu-
tion assessment without human intervention. An engine fault racy compared with other ANN-based models. In addition, a
diagnosis method using vibration signal was developed in [96] hierarchical DNN (i.e., three level) was proposed in [101]. The
based on multilayer feedforward ANNs. In order to simplify the layers of these models were 5, 4, and 3. In detail, the first hierar-
prognosis task, the data-driven prognostic method based on the chy was specially designed for fault identification, the second for
ANN combined with the Weibull distribution was developed for locating the fault source, and the third for recognizing the fault
achieving more accurate bearing RUL prediction [97]. severity. As stated in [102], the most recent statistical techniques
(i.e., DNN) were applied to a naval propulsion system and have
shown the best performance. In addition, the proposed method
B. DNN for PdM
resulted in an optimal PdM strategy with reduction in the op-
Literally, a DNN is a deeper neural network. Here, it refers erational and maintenance cost. A DNN-based model was also
to an NN with many hidden layers, and all of the layers are employed in [103] to perform health monitoring of bearings.
fully connected to each other. Thus, the constraint condition l in Notably, the proposed approach considered temporal coherence
formula (11) is also extended to l ∈ {1, 2, . . . m}. Most impor- with a former time-series signal. A discrete wavelet transform
tantly, the ability of hierarchical nonlinear mapping learning is was employed in [104] to provide discriminative features from
a replacement for manual feature acquisition with unsupervised three common condition monitoring signals including acoustic,
or semisupervised feature learning. Compared with traditional vibration, and torque for a diagnostic model. Then, the DNN
data-driven methods, DNNs can self-adaptively extract fault was utilized to recognize five types of faults of a gearbox and
features to effectively represent crucial information and realize obtained an excellent result.
intelligent diagnosis; they can also improve identification accu-
racy and are extremely effective in reducing defects in manual
C. AE for PdM
design features [98].
According to [99], a DNN was employed to develop the pre- In the PdM of industrial equipment, labeled raw sensor signals
diction model for monitoring the wind turbine gearbox health are generally difficult to obtain, requiring specific and detailed
and identifying impending failures. A five-layer DNN was experimental settings. Obviously, unsupervised feature learning
is especially suitable for handling unlabeled data; it can provide packet decomposition and Fisher’s criterion. Similarly,
a feasible solution for fault identification and RUL estimation. Pandya et al. [61] paid more attention to feature engineer-
In addition, the AE model belongs to the type of unsupervised ing, and there were five real wavelets and two wavelet
learning that only requires unlabeled measurement data, and it selection criteria utilized to perform fault diagnosis. Per-
has received extensive attention and application. Accordingly, haps, most importantly, the experimental data were ac-
this section will give a specific presentation about the AE model. quired from a test rig (e.g., a motor), which resulted in
As stated in [105], a multimodel combination method was ap- lower data acquisition cost. In addition, the experimental
plied to identify the fault condition of rolling bearings. Among data can be filtered from a large amount of collected data,
three models, the stacked-AE is slightly better than the deep which is more helpful for model training. Therefore, we
belief network (DBN), but their diagnostic accuracy is more consider it is reasonable to achieve 100% accuracy on the
than 99%. According to [106], a deep AE method was applied basis of such adequate data acquisition and feature engi-
to the fault diagnosis of a gearbox and an electrical locomo- neering. In contrast, the accuracy does not seem optimistic
tive bearing, and the corresponding diagnostic accuracy was in [65], at just 69%. However, it is worth mentioning that
94.05% and 87.8%, respectively. Because of the existence of unlike fault diagnosis (i.e., classification), the task faced
noise disturbance, the composition of a fault vibration signal in this paper is RUL estimation (i.e., regression), and the
is relatively complicated. Furthermore, the nonlinear and non- accuracy here is the average of 11 bearings. In addition,
stationary characteristics of a fault vibration signal in a rolling compared with other works, this paper introduced less re-
bearing system are considerations. As a result, an intelligent garding feature engineering, and perhaps it is the reason
method was proposed in [107] based on wavelet transform and for the low accuracy.
a stacked-AE, where the model was employed to extract features 2) SVM: In reality, the SVM is not only good at classifica-
of the denoised signal. Moreover, an enhancement feature fu- tion tasks but also performs well in fault diagnosis. Data-
sion method [108] based on deep AE was developed for rotating driven algorithms mainly depend on the ability to carry
machinery and was constructed with a denoising AE and a con- out fault representation via learning from sensor signals.
tractive AE for improving the feature learning ability. Similarly, Moreover, Saimurugan et al. [76] conducted a multiclass
a stacked denoising AE was deployed in [109], which was able bearing fault diagnosis task. The difference between the
to adaptively extract significant fault features and effectively two types of fault signals may be especially slight, which
identify health conditions with the best diagnostic performance. will lead to the misclassification between classes, just as
In addition, the sparse AE possesses the ability to learn effective a7, a9, a10, and a12 are misclassified in the experiments.
feature representation from raw signals in an unsupervised man- Accordingly, we consider that the diagnostic accuracy of
ner. Accordingly, a sparse AE-based architecture was deployed 99.33% obtained in this paper is already considerable. In
in [110] for motors and an air compressor, and the final diagnos- contrast, a six-class diagnosis task was conducted in [73].
tic accuracy achieved was 97.22%. A multisensor data fusion According to the three confusion matrixes given in this pa-
model was developed in [111], which used a two-layer sparse per, we calculate the accuracy of the proposed algorithm
AE and combined it with a DBN to assess health conditions under three different speed and load conditions, as only
of rotating machinery. An effective approach utilizing sparse 50.96%, 59.09%, 71.71%. In addition, the dataset scale
AE was developed in [112] for the fault diagnosis of induction and the data processing flow are not explained in detail in
motor. Moreover, an ensemble deep AE method was proposed this paper, so it is not easy for us to provide an in-depth
in [113]. evaluation of the results.
3) DT and RF: It is well known that the DT model inherits
V. ANALYSIS OF THE METRIC: ABOUT ACCURACY a well interpretable and suitable approach for processing
high dimensional data but is prone to overfitting. How-
A. Analysis of Accuracy ever, RF solves the problem of weak generalization abil-
Both the optimization and innovation of methods are aimed at ity for DT. As a result, Wu et al. [33] obtained excellent
improving the effectiveness of the industrial PdM. Specifically, performance in the industrial PdM. Specifically, the au-
the purpose of these applications is to improve the accuracy thors collected the signals from three sensors (i.e., cutting
of fault diagnosis or RUL estimation. Accordingly, we provide force, vibration, and acoustic emission), which formed a
several viewpoints regarding the metric of accuracy and com- seven-channel data acquisition system, and then four fea-
bine the statistics in Tables III and IV. Specifically, we focus on tures were extracted from each channel. In summary, a
those works with the highest and lowest accuracy in each type detailed feature engineering task has been conducted in
of algorithm and present our opinions as follows. this paper, which is in accordance with the characteris-
1) LR: It is well known that the complexity of the LR model tic of the proposed model. Therefore, we deem that the
is the lowest among various ML algorithms. However, it tool wear prediction accuracy of 99.2% achieved by using
still achieves 100% accuracy in some applications, such as 28-dimension features is an evidence-based result. In con-
in [60] and [61]. There were three sensor signals (i.e., dis- trast, Canizo et al. [18] only obtained accuracy of 82.04%.
placement, vibration, and current) in [60] used for the fault Nevertheless, it is worth noting that this paper was based
representation learning, and 19-dimension features were on another existing work, and although the obtained ac-
selected as input vectors of an LR model using wavelet curacy does not seem as good, it still achieved an increase
of 5.54%.
4) ANN, DNN, and AE: It can be seen that among the come from public data centers, such as case western re-
three models of DL, in 80% of the literature, they have serve university/intelligent maintenance systems, or a lab-
achieved more than 97% accuracy, and five papers have level experimental platform (i.e., test rig). Only a few
even reached 100%, including [45], [92], [98], [103], and papers utilize datasets that are collected from the actual
[108]. These papers deployed an NN to arm models with operating equipment.
the ability to represent the equipment health condition by 6) With the great improvement of computational power and
using hierarchical learning. A simple NN is constructed the rapid growth of data volume, AI algorithms and their
in [44] and [90] with a single-hidden layer, and the infor- variant models have increasingly demonstrated a superior
mation learned from sensor signals is transmitted forward advantage of performance. With the continuous innova-
through linear weighting and a nonlinear activation function of algorithms, researchers will continue to focus on
tion. Especially for [44], 20 features are extracted from the the data-driven methods in PdM applications.
time, frequency, and time-frequency domain using three 7) In terms of model performance, these algorithms are ap-
sensors (i.e., vibration, pressure, and speed), and the ap- plicable to most industrial applications. But, in the aspect
proach is sufficient to support the model training for a of revealing the essential reasons, the existing algorithms
four-class task. In particular, a large dataset was built for not only lack interpretability but also lack the ability to
model training, and the data they used were all from the explain specific phenomena.
same equipment. Moreover, more hidden layers were de-
ployed in [98] and [103] to extend the network depth,
VI. FUTURE RESEARCH DIRECTIONS
ensuring that the network has a powerful learning ability.
Especially for [103], a six-layer DNN was built to con- Although data-driven methods have achieved excellent per-
duct fault diagnosis. In addition, the temporal coherence, formance in PdM applications, there is still a large potential for
which makes full use of the time characteristic for the sen- improvement and optimization, especially for practical appli-
sor sequence, is taken into consideration. A four-layer AE cations. Accordingly, some research trends and potential direc-
model that belongs to unsupervised learning was deployed tions are given as follows.
in [108]. The enhanced feature learning ability enabled the 1) Data validity: As is known, the significance of data to
proposed model to more accurately distinguish different an algorithm’s performance is self-evident. However, a
conditions. In summary, DL has demonstrated its powerful large number of studies are currently only using public
capabilities in the industrial PdM. Specifically, building datasets provided by some platforms, but fewer datasets
a deeper model allows for a more strong representation, originate from the actual operating equipment. In addition,
which will be extremely helpful in improving prediction the construction of the data acquisition system will be
accuracy. expensive, and the sensor itself also has the potential to
In summary, for traditional ML algorithms, sufficient data fail. When the cost of the diagnosis is higher than that
acquisition and feature engineering are helpful in optimizing of the maintenance, the diagnosis will lose its original
algorithm performance, while for DL algorithms, the deeper meaning in a sense. For these reasons, reliable CPS and
network architecture and the higher dimensional feature vectors IoT (e.g., IWSNs) technologies are required to provide
are more significant in improving the task metric. low acquisition cost and high utilization value for a PdM
study. This approach facilitates researchers’ capability to
make full use of data from the actual operating equipment,
B. Several Conclusions
rather than from the experimental platform, and to better
In this section, some conclusions are drawn from Tables III solve the practical problems in industrial manufacturing.
and IV and are shown as follows. 2) Dataset construction and the unbalanced feature of
1) Both ML and DL can remarkably complete the PdM task. datasets: There are many challenges that need to be re-
The average prediction accuracy of the reviewed literature solved when handling mechanical data, such as the accu-
can reach 95.06%, and the highest accuracy is 100%. racy and the actual meaning of the label, as well as the
2) Signals that can be used for fault diagnosis include acous- importance of the label for the PdM. In most instances, the
tic emission, electrical signature parameters (current and observed data volume for normal operations far exceeds
voltage), temperature, pressure, rotation speed, and vi- the observed data volume for an anomalous operation,
bration. However, vibration signals are exploited most which will cause great difficulties in conducting model
frequently in terms of the literature surveyed herein. training. As a result, it is especially necessary to acquire
3) By using comparison, the monitored key components such more data under fault status scenarios, which will make
as bearings and cutting blades are broadly divided into the entire dataset more balanced.
rotating machinery and reciprocating machinery. 3) Unsupervised learning: As mentioned above, the current
4) In terms of the tasks completed, they are mainly divided achievements of AI are mainly focused on supervised
into two aspects: fault diagnosis and RUL prediction. In learning, which means that datasets need to be uniquely
some cases, it is also referred to the performance degra- annotated. The challenges of this task are described in the
dation monitoring of equipment. previous item. Accordingly, learning with unlabeled data,
5) For these applications, a noteworthy point is the source of namely unsupervised learning, is a significant research
the dataset. As seen from our summaries, most datasets direction in the future.
4) Model migration and generalization ability: For practical [5] J. Lee, “Industrial Big Data, China: Mechanical Industry Press, Beijing,
applications, the model migration ability is identically 2015.
[6] D. H. Zhou, M. Y. Chen, and Z. G. Xu, Reliability Prediction and Optimal
important. Specifically, a trained model will ensure more Maintenance Technology, Univ. Sci. Technol. China Press, Hefei City,
applicability when it can be applied to other equipment, Anhui Province, 2013.
which can not only save unnecessary time consumption [7] A. K. S. Jardine, D. Lin, and D. Banjevic, “A review on machinery di-
agnostics and prognostics implementing condition-based maintenance,”
but also simplify the PdM design process. Mech. Syst. Signal Process., vol. 20, no. 7, pp. 1483–1510, 2006.
5) Multisensory technique: The first wave of DL is using [8] K. Wang, “Intelligent predictive maintenance (IPdM) system—Industry
massive amounts of data to train very large neural net- 4.0 scenario,” WIT Trans. Eng. Sci., vol. 113, pp. 259–268, 2016.
[9] S. Cheng, K. Tom, L. Thomas, and M. Pecht, “A wireless sensor sys-
works, and many of the breakthroughs were based on this tem for prognostics and health management,” IEEE Sensors J., vol. 10,
manner. However, not all tasks are liable to obtain large no. 4, pp. 856–862, Apr. 2010.
volume of effective data, so it is urgent to develop robust [10] K. Krishnamoorthy, P. Ashok, and D. Tesar, “Simultaneous sensor and
process fault detection and isolation in multiple-input–multiple-output
algorithm on small scale dataset, and multisensory fusion systems,” IEEE Syst. J., vol. 9, no. 2, pp. 335–349, Jun. 2015.
technique provides possibilities. [11] Y. Jiang, S. Yin, and O. Kaynak, “Data-driven monitoring and safety
6) Safety: Here, safety refers to the loss caused by system control of industrial cyber-physical systems: Basics and beyond,” IEEE
Access, vol. 6, pp. 47374–47384, 2018.
misdiagnosis. It can be divided into three categories: pos- [12] D. Yang et al., “Assignment of segmented slots enabling reliable real-
itive misdiagnosis (i.e., no damage but damage is diag- time transmission in industrial wireless sensor networks,” IEEE Trans.
nosed), intermediate misdiagnosis (i.e., one fault type is Ind. Electron., vol. 62, no. 6, pp. 3966–3977, Jun. 2015.
[13] L. D. Xu, W. He, and S. Li, “Internet of Things in industries: A survey,”
diagnosed as another fault type), and negative misdiagno- IEEE Trans. Ind. Inform., vol. 10, no. 4, pp. 2233–2243, Nov. 2014.
sis (i.e., the damage exists, but no damage is diagnosed). [14] X. Li et al., “A review of industrial wireless networks in the context of
Obviously, any misdiagnosis will cause a certain degree industry 4.0,” Wireless Netw., vol. 23, no. 1, pp. 23–41, 2017.
[15] J. J. A. Costello, G. M. West, and S. D. J. Mcarthur, “Machine learning
of loss to the system and may even cause system break- model for event-based prognostics in gas circulator condition monitor-
down. Consequently, how to design a safe and reliable ing,” IEEE Trans. Rel., vol. 66, no. 4, pp. 1048–1057, Dec. 2017.
PdM system has become another major challenge. [16] Y. Jiang and S. Yin, “Recursive total principle component regression
based fault detection and its application to vehicular cyber-physical
systems,” IEEE Trans. Ind. Inform., vol. 14, no. 4, pp. 1415–1423,
Apr. 2018.
VII. CONCLUSION [17] D. You, X. Gao, and S. Katayama, “WPD-PCA-based laser welding
In this paper, we conduct a comprehensive survey of the process monitoring and defects diagnosis by using FNN and SVM,”
IEEE Trans. Ind. Electron., vol. 62, no. 1, pp. 628–636, Jan. 2015.
PdM of industrial equipment. First, we provide a systematic [18] M. Canizo, E. Onieva, A. Conde, S. Charramendieta, and S. Trujillo,
overview of the PdM, propose an industrial PdM scheme for “Real-time predictive maintenance for wind turbines using Big Data
automatic washing equipment, and demonstrate the challenges frameworks,” in Proc. IEEE Int. Conf. Prognostics Health Manage.,
2017, pp. 70–77.
faced when we conduct a PdM research study. Second, we re- [19] Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and
view the industrial applications of PdM in the recent five-year fault-tolerant techniques—Part I: Fault diagnosis with model-based and
timeframe both from ML and DL perspectives, and then five in- signal-based approaches,” IEEE Trans. Ind. Electron., vol. 62, no. 6,
pp. 3757–3767, Jun. 2015.
dicators (e.g., signal type, application scenario, target, accuracy, [20] R. Zhao et al., “Deep learning and its applications to machine health
and data source) are selected to summarize the application of monitoring,” Mech. Syst. Signal Process., vol. 115, pp. 213–237, 2019.
these algorithms. Third, we summarize some conclusions from [21] R. Liu et al., “Artificial intelligence for fault diagnosis of rotating ma-
chinery: A review,” Mech. Syst. Signal Process., vol. 108, pp. 33–47,
the entire PdM perspective, and the accuracy (a metric to esti- 2018.
mate the algorithm performance) of each algorithm is analyzed [22] G. Smith, J. B. Schroeder, S. Navarro, and D. Haldeman, “Development
in detail. Finally, we discuss several research trends and future of a prognostics and health management capability for the Joint Strike
Fighter,” in Proc. IEEE Autotestcon , 1997, pp. 676–682.
directions. [23] G. A. Susto, A. Schirru, S. Pampuri, S. McLoone, and A. Beghi,
In summary, although PdM-related works are emerging end- “Machine learning for predictive maintenance: A multiple classifier
lessly, the challenges that the PdM is facing are self-evident approach,” IEEE Trans. Ind. Inform., vol. 11, no. 3, pp. 812–820,
Jun. 2015.
concerns. Accordingly, we will pay more attention to address- [24] J. Wan et al., “A manufacturing big data solution for active preventive
ing these challenges in practical applications. We hope this paper maintenance,” IEEE Trans. Ind. Inform., vol. 13, no. 4, pp. 2039–2047,
can provide graduate students, companies, and institutions with Aug. 2017.
[25] W. Qiao and D. Lu, “A survey on wind turbine condition monitoring and
a preliminary understanding of PdM research. fault diagnosis—Part I: Components and subsystems,” IEEE Trans. Ind.
Electron., vol. 62, no. 10, pp. 6536–6545, Oct. 2015.
[26] J. Lee et al., “Prognostics and health management design for rotary
REFERENCES machinery systems—Reviews, methodology and applications,” Mech.
Syst. Signal Process., vol. 42, no. 1-2, pp. 314–334, 2014.
[1] H. S. Kang et al., “Smart manufacturing: Past research, present findings, [27] Y. Xu, Y. Sun, J. Wan, X. Liu, and Z. Song , “Industrial big data for fault
and future directions,” Int. J. Precis. Eng. Manuf.-Green Technol., vol. 3, diagnosis: Taxonomy, review, and applications,” IEEE Access, vol. 5,
no. 1, pp. 111–128, 2016. no. 99, pp. 17368–17380, 2017.
[2] J. Lee, B. Bagheri, and H. A. Kao, “A cyber-physical systems architec- [28] Y. Jiang and S. Yin, “Recent advances in key-performance-indicator
ture for industry 4.0-based manufacturing systems,” Manuf. Lett., vol. 3, oriented prognosis and diagnosis with a MATLAB toolbox: DB-KIT,”
pp. 18–23, 2015. IEEE Trans. Ind. Inform., vol. 15, no. 5, pp. 2849–2858, May 2019.
[3] H. Lasi et al., “Industry 4.0. business & information systems engineer- [29] C. R. Farrar and K. Worden, Structural Health Monitoring: A Machine
ing,” Bus. Inf. Syst. Eng., vol. 6, no. 4, pp. 239–242, 2014. Learning Perspective. Hoboken, NJ, USA: Wiley, 2012.
[4] L. Ling, “China’s manufacturing locus in 2025: With a comparison of [30] Z. Chang, L. Lei, Z. Y. Zhou, S. Mao, and T. Ristaniemi, “Learn to cache:
“made-in-china 2025” and “industry 4.0”,” Technol. Forecasting Social Machine learning for network edge caching in the Big Data era,” IEEE
Change, vol. 135, pp. 66–74, 2017. Wireless Commun., vol. 25, no. 3, pp. 28–35, Jun. 2018.
[31] H. Li, K. Ota, and M. Dong, “Learning IoT in edge: Deep learning for [55] Y. Wang, Z. Wei, and J. Yang, “Feature trend extraction and adaptive
the Internet of Things with edge computing,” IEEE Netw., vol. 32, no. 1, density peaks search for intelligent fault diagnosis of machines,” IEEE
pp. 96–101, Jan./Feb. 2018. Trans. Ind. Inform., vol. 15, no. 1, pp. 105–115, Jan. 2018.
[32] Z. Jiang, Z. Li, N. Wu, and M. Zhou, “A Petri net approach to fault [56] A. Manghai, R. Jegadeeshwaran, and V. Sugumaran, “Brake fault diag-
diagnosis and restoration for power transmission systems to avoid nosis through machine learning approaches—A review,” SDHM Struct.
the output interruption of substations,” IEEE Syst. J., vol. 12, no. 3, Durability Health Monit., vol. 12, no. 1, pp. 43–67, 2017
pp. 2566–2576, Sep. 2017. [57] T. Wuest et al., “Machine learning in manufacturing: Advantages, chal-
[33] D. Wu, C. Jennings, J. Terpenny, and S. Kumara, “Cloud-based machine lenges, and applications,” Prod. Manuf. Res. An Open Access J., vol. 4,
learning for predictive analytics: Tool wear prediction in milling,” in no. 1, pp. 23–45, 2016.
Proc. IEEE Int. Conf. Big Data, 2017, pp. 2062–2069. [58] J. Hosmer et al., Applied Logistic Regression. Hoboken, NJ, USA: Wiley,
[34] G. A. Susto et al., “A predictive maintenance system based on regu- 2013, vol. 398.
larization methods for ion-implantation,” Adv. Semicond. Manuf. Conf., [59] H. Li et al., “Cutting tool operational reliability prediction based on
vol. 497, pp. 175–180, 2012. acoustic emission and logistic regression model,” J. Intell. Manuf., vol.
[35] W. Zhou, T. G. Habetler, and R. G. Harley, “Bearing condition monitoring 26, no. 5, pp. 923–931, 2015.
methods for electric machines: A general review,” in Proc. IEEE Int. [60] J. Yan and J. Lee, “Degradation assessment and fault modes classification
Symp. Diagnostics Elect. Mach., Power Electron. Drives, 2007, pp. 3–6. using logistic regression,” J. Manuf. Sci. Eng., vol. 127, no. 4, pp. 912–
[36] V. T. Tran, F. Althobiani, and A. Ball, “An approach to fault diagno- 914, 2005.
sis of reciprocating compressor valves using Teager–Kaiser energy op- [61] D. H. Pandya, S. H. Upadhyay, and S. P. Harsha, “Fault diagnosis of
erator and deep belief networks,” Expert Syst. Appl., vol. 41, no. 9, rolling element bearing by using multinomial logistic regression and
pp. 4113–4122, 2014. wavelet packet transform,” Soft Comput., vol. 18, no. 2, pp. 255–266,
[37] M. J. Schulz and M. J. Sundaresan, “Smart sensor system for structural 2014.
condition monitoring of wind turbines,” in Proc. Energy Convers. Congr. [62] W. Caesarendra, A. Widodo, and B. S. Yang, “Application of relevance
Expo., 2006, pp. 2417–2421. vector machine and logistic regression for machine degradation assess-
[38] G. K. Singh and S. A. S. Al Kazzaz, “Development of an intelligent ment,” Mech. Syst. Signal Process., vol. 24, no. 4, pp. 1161–1171, 2010.
diagnostic system for induction machine health monitoring,” IEEE Syst. [63] T.-L. Wu, D. Y. Sari, B.-T. Lin, and C.-W. Chang, “Monitoring of punch
J., vol. 2, no. 2, pp. 273–288, Jun. 2008. failure in micro-piercing process based on vibratory signal and logistic
[39] S. Nandi, H. A. Toliyat, and X. Li, “Condition monitoring and fault regression,” Int. J. Adv. Manuf. Technol., vol. 93, no. 5–8, pp. 2447–2458,
diagnosis of electrical motors—A review,” IEEE Trans. Energy Convers., 2017.
vol. 20, no. 4, pp. 719–729, Dec. 2005. [64] E. Skordilis and R. Moghaddass, “A condition monitoring approach
[40] G. W. Cai, Big Data Analytics in Structural Health Monitoring, for real-time monitoring of degrading systems using Kalman filter and
Nashville, TN, USA: Vanderbilt Univ., 2017. logistic regression,” Int. J. Prod. Res., vol. 55, no. 19, pp. 1–18, 2017.
[41] S. Hong and Z. Zhou, “Application of Gaussian process regression for [65] W. Ahmad et al., “A reliable technique for remaining useful life estima-
bearing degradation assessment,” Inf. Sci. Service Sci. Data Mining, tion of rolling element bearings using dynamic regression models,” Rel.
vol. 42, pp. 644–648, 2013. Eng. Syst. Saf., 2018.
[42] W. Qiao and D. Lu, “A survey on wind turbine condition monitoring and [66] J. Phillips et al., “Classifying machinery condition using oil samples
fault diagnosis—Part II: Signals and signal processing methods,” IEEE and binary logistic regression,” Mech. Syst. Signal Process., vol. 60–61,
Trans. Ind. Electron., vol. 62, no. 10, pp. 6546–6557, Oct. 2015. pp. 316–325, 2015.
[43] Y. Qu et al., “A new acoustic emission sensor based gear fault detection [67] J. Yu, “Tool condition prognostics using logistic regression with pe-
approach,” Int. J. Prognostics Health Manage., vol. 4, pp. 32–45, 2013. nalization and manifold regularization,” Appl. Soft Comput., vol. 64,
[44] F. Immovilli, A. Bellini, R. Rubini, and C. Tassoni, “Diagnosis of bearing pp. 453–467, 2017.
faults in induction machines by vibration or current signals: A critical [68] A. Widodo and B. S. Yang, “Support vector machine in machine con-
comparison,” IEEE Trans. Ind. Appl., vol. 46, no. 4, pp. 1350–1359, dition monitoring and fault diagnosis,” Mech. Syst. Signal Process.,
Jul.–Aug. 2010. vol. 21, no. 6, pp. 2560–2574, 2007.
[45] C. Jin, W. Zhao, Z. Liu, J. Lee, and X. He, “A vibration-based approach [69] A. Widodo et al., “Fault diagnosis of low speed bearing based on rele-
for diesel engine fault diagnosis,” in Proc. Int. Conf. Prognostics Health vance vector machine and support vector machine,” Expert Syst. Appl.,
Manage., pp. 1–9, 2014. vol. 36, no. 3, pp. 7252–7261, 2009.
[46] B. Jie, L. Hou, and M. Yongguang, “Machine fault diagnosis using [70] M. Baptista et al., “Forecasting fault events for predictive maintenance
industrial wireless sensor networks and support vector machine,” in Proc. using data-driven techniques and ARMA modeling,” Comput. Ind. Eng.,
12th IEEE Int. Conf. Electron. Meas. Instrum., 2015, vol. 1, pp. 153–158. vol. 115, pp. 41–53, 2018.
[47] R. K. Mobley, An Introduction to Predictive Maintenance. New York, [71] U. Shafi et al., “Vehicle remote health monitoring and prognostic main-
NY, USA: Elsevier, 2002. tenance system,” J. Adv. Transp., no. 1, pp. 1–10, 2018.
[48] D. Lu, X. Gong, and W. Qiao, “Current-based diagnosis for gear tooth [72] A. Soualhi, K. Medjaher, and N. Zerhouni, “Bearing health monitoring
breaks in wind turbine gearboxes,” in Proc. Energy Convers. Congr. based on Hilbert–Huang transform, support vector machine, and regres-
Expo., 2012, vol. 11, pp. 3780–3786. sion,” IEEE Trans. Instrum. Meas., vol. 64, no. 1, pp. 52–62, Jan. 2015.
[49] P. Mendonca et al., “Development of a reduced-model laboratory for [73] O. R. Seryasat, M. A. Shoorehdeli, F. Honarvar, and A. Rahmani, “Multi-
testing predictive fault system in internal combustion engines,” in Proc. fault diagnosis of ball bearing based on features extracted from time-
IEEE 10th Int. Symp. Diagnostics Elect. Mach., Power Electron. Drives, domain and multi-class support vector machine (MSVM),” in Proc. IEEE
2015, pp. 428–434. Int. Conf. Syst. Man Cybern., 2010, pp. 4300–4303.
[50] C. Chen, B. Zhang, and G. Vachtsevanos, “Prediction of machine health [74] C. Li et al., “Multimodal deep support vector classification with
condition using neuro-fuzzy and Bayesian algorithms,” IEEE Trans. In- homologous features and its application to gearbox fault diagnosis,”
strum. Meas., vol. 61, no. 2, pp. 297–306, Feb. 2012. Neurocomputing, vol. 168, pp. 119–127, 2015.
[51] X. Liu, P. Song, C. Yang, C. Hao, and W. Peng, “Prognostics and health [75] D. Yang et al., “Gear fault diagnosis based on support vector machine
management of bearings based on logarithmic linear recursive least- optimized by artificial bee colony algorithm,” Mechanism Mach. Theory,
squares and recursive maximum likelihood estimation,” IEEE Trans. vol. 90, pp. 219–229, 2015.
Ind. Electron., vol. 65, no. 2, pp. 1549–1558, Feb. 2018. [76] M. Saimurugan et al., “Multi component fault diagnosis of rotational
[52] B. Cline et al., “Predictive maintenance applications for machine learn- mechanical system based on decision tree and support vector machine,”
ing,” in Proc. Rel. Maintainability Symp., 2017, pp. 1–7. Expert Syst. Appl., vol. 38, no. 4, pp. 3819–3826, 2011.
[53] R. Zhao, D. Wang, R. Yan, K. Mao, F. Shen, and J. Wang, “Machine [77] X. Zhang, Y. Liang, and J. Zhou, “A novel bearing fault diagnosis model
health monitoring using local feature-based gated recurrent unit net- integrated permutation entropy, ensemble empirical mode decomposition
works,” IEEE Trans. Ind. Electron., vol. 65, no. 2, pp. 1539–1548, and optimized SVM,” Measurement, vol. 69, pp. 164–179, 2015.
Feb. 2018. [78] P. J. G. Nieto et al., “Hybrid PSO–SVM-based method for forecasting
[54] Y. Qin, X. Wang, and J. Zou, “The optimized deep belief networks with of the remaining useful life for aircraft engines and evaluation of its
improved logistic sigmoid units and their application in fault diagnosis reliability,” Rel. Eng. Syst. Saf., vol. 138, pp. 219–231, 2015.
for planetary gearboxes of wind turbines,” IEEE Trans. Ind. Electron.,
vol. 66, no. 5, pp. 3814–3824, May 2019.
[79] S. R. Safavian and D. Landgrebe, “A survey of decision tree classi- [104] M. Heydarzadeh et al., “Gear fault diagnosis using discrete wavelet
fier methodology,” IEEE Trans. Syst., Man, Cybern., vol. 21, no. 3, transform and deep neural networks,” in Proc. Annu. Conf. Ind. Electron.
pp. 660–674, May/Jun. 1991. Soc., 2016, pp. 1494–1500.
[80] L. Breiman, “Random forests,” Mach. Learn., vol. 45, no. 1, pp. 5–32, [105] Z. Chen et al., “Deep neural networks-based rolling bearing fault diag-
2001. nosis,” Microelectron. Rel., vol. 75, pp. 327–333, 2017.
[81] A. Krishnakumari et al., “Fault diagnostics of spur gear using decision [106] H. Shao et al., “A novel deep autoencoder feature learning method for
tree and fuzzy classifier,” Int. J. Adv. Manuf. Technol., vol. 89, no. 9–12, rotating machinery fault diagnosis,” Mech. Syst. Signal Process., vol. 95,
pp. 1–8, 2017. pp. 187–204, 2017.
[82] G. Li et al., “An improved decision tree-based fault diagnosis method [107] T. Junbo et al., “Fault diagnosis method study in roller bearing based
for practical variable refrigerant flow system using virtual sensor-based on wavelet transform and stacked auto-encoder,” in Proc. Control Decis.
fault indicators,” Appl. Thermal Eng., vol. 129, pp. 1292–1303, 2018. Conf., 2015, pp. 4608–4613.
[83] P. Santos, J. Maudes, and A. Bustillo, “Identifying maximum imbalance [108] H. Shao et al., “An enhancement deep feature fusion method for rotating
in datasets for fault diagnosis of gearboxes,” J. Intell. Manuf., vol. 29, machinery fault diagnosis,” Knowl.-Based Syst., vol. 119, pp. 200–220,
no. 2, pp. 333–351, 2018. 2016.
[84] R. Shrivastava, H. Mahalingam, and N. N. Dutta, “Application and evalu- [109] C. Lu et al., “Fault diagnosis of rotary machinery components using a
ation of random forest classifier technique for fault detection in bioreactor stacked denoising autoencoder-based health state identification,” Signal
operation,” Chem. Eng. Commun., vol. 204, no. 5, pp. 591–598, 2017. Process., vol. 130, pp. 377–388, 2017.
[85] C. Li et al., “Gearbox fault diagnosis based on deep random forest fusion [110] N. K. Verma, V. K. Gupta, M. Sharma, and R. K. Sevakula, “Intelli-
of acoustic and vibratory signals,” Mech. Syst. Signal Process., vol. 76, gent condition based monitoring of rotating machines using sparse auto-
pp. 283–293, 2016. encoders,” in Proc. IEEE Conf. Prognostics Health Manage., 2013, vol.
[86] Y. Lecun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, 7789, pp. 1–7.
no. 7553, pp. 436–444, 2015. [111] Z. Chen and W. Li, “Multisensor feature fusion for bearing fault diagnosis
[87] F. Lv, C. Wen, Z. Bao, and M. Liu, “Fault diagnosis based on deep using sparse autoencoder and deep belief network,” IEEE Trans. Instrum.
learning,” in Proc. Amer. Control Conf., 2016, pp. 6851–6856. Meas., vol. 66, no. 7, pp. 1693–1702, Jul. 2017.
[88] G.E. Hinton and R. Salakhutdinov, “Reducing the dimensionality of data [112] W. Sun et al., “A sparse auto-encoder-based deep neural network ap-
with neural networks,” Science, vol. 313, no. 5786, pp. 504–507, 2006. proach for induction motor faults classification,” Measurement, vol. 89,
[89] G. Zhao et al., “Research advances in fault diagnosis and prognostic pp. 171–178, 2016.
based on deep learning,” in Proc. Prognostics Syst. Health Manage. [113] H. Shao et al., “A novel method for intelligent fault diagnosis of rolling
Conf. (PHM-Chengdu), 2016, pp. 1–6. bearings using ensemble deep auto-encoders,” Mech. Syst. Signal Pro-
[90] J. Schmidhuber, “Deep learning in neural networks: An overview,” Neu- cess., vol. 102, pp. 278–297, 2018.
ral Netw., vol. 61, pp. 85–117, 2015.
[91] A. K. Jain, J. Mao, and K. M. Mohiuddin, “Artificial neural networks: A
tutorial,” Computer, vol. 29, no. 3, pp. 31–44, 1996.
[92] A. K. Mahamad, S. Saon, and T. Hiyama, “Predicting remaining useful
life of rotating machinery based artificial neural network,” Comput. Math. Weiting Zhang received the M.S. degree from the
Appl., vol. 60, no. 4, pp. 1078–1087, 2010. Inner Mongolia University of Technology, Hohhot,
[93] M. D. Prieto, G. Cirrincione, A. G. Espinosa, J. A. Ortega, and China, in 2017. He is currently working toward the
H. Henao, “Bearing fault detection by a novel condition-monitoring Ph.D. degree in communication and information sys-
scheme based on statistical-time features and neural networks,” IEEE tems at Beijing Jiaotong University, Beijing, China.
Trans. Ind. Electron., vol. 60, no. 8, pp. 3398–3407, Aug. 2013. His specific areas of research interest mainly focus
[94] Y. Lei, F. Jia, J. Lin, S. Xing, and S. X. Ding, “An intelligent fault di- on deep learning, predictive maintenance, and wire-
agnosis method using unsupervised feature learning towards mechanical less sensor networks.
big data,” IEEE Trans. Ind. Electron., vol. 63, no. 5, pp. 3137–3147, May
2016.
[95] J. B. Ali et al., “Application of empirical mode decomposition and
artificial neural network for automatic bearing fault diagnosis based on
vibration signals,” Appl. Acoust., vol. 89, pp. 16–27, 2015. Dong Yang (M’11) received the B.S. degree from
[96] R. Ahmed, M. El Sayed, S. A. Gadsden, J. Tjong, and S. Habibi, “Au- Central South University, Changsha, China, in 2003
tomotive internal-combustion-engine fault detection and classification and the Ph.D. degree in communications and in-
using artificial neural network techniques,” IEEE Trans. Veh. Technol., formation science from Beijing Jiaotong University,
vol. 64, no. 1, pp. 21–33, Jan. 2015. Beijing, China, in 2009.
[97] J. B. Ali et al., “Accurate bearing remaining useful life prediction based From March 2009 to June 2010, he was a Post-
on Weibull distribution and artificial neural network,” Mech. Syst. Signal Doctoral Research Associate with Jonkoping Univer-
Process., vol. 56, pp. 150–172, 2015. sity, Jonkoping, Sweden. In August 2010, he joined
[98] H. Hu, B. Tang, X. Gong, W. Wei, and H. Wang, “Intelligent fault the School of Electronic and Information Engineer-
diagnosis of the high-speed train with big data based on deep neural ing, Beijing Jiaotong University. Currently, he is
networks,” IEEE Trans. Ind. Inform., vol. 13, no. 4, pp. 2106–2116, Aug. working as a Professor. His research interests are
2017. network technologies, including Internet architecture, industrial Internet, and
[99] L. Wang, Z. Zhang, H. Long, J. Xu, and R. Liu, “Wind turbine gear- wireless sensor networks.
box failure identification with deep neural networks,” IEEE Trans. Ind.
Inform., vol. 13, no. 3, pp. 1360–1368, Jun. 2017.
[100] F. Jia et al., “Deep neural networks: A promising tool for fault character-
istic mining and intelligent diagnosis of rotating machinery with massive
data,” Mech. Syst. Signal Process., vol. 72, pp. 303–315, 2016. Hongchao Wang (M’14) received the B.S. degree
[101] F. Zhou, Y. Gao, and C. Wen, “A novel multimode fault classification in communication engineering and the Ph.D. de-
method based on deep learning,” J. Control Sci. Eng., vol. 2017, no. 1, gree in communication and information systems from
pp. 1–14, 2017, Art. no. 3583610. Beijing Jiaotong University, Beijing, China, in 2005
[102] F. Cipollini et al., “Condition-based maintenance of naval propul- and 2012, respectively.
sion systems with supervised data analysis,” Ocean Eng., vol. 149, Currently he is working with the School of Elec-
pp. 268–278, 2018. tronic and Information Engineering, Beijing Jiaotong
[103] R. Zhang et al., “Fault diagnosis from raw sensor data using deep neural University. His research interests include Internet
networks considering temporal coherence,” Sensors, vol. 17, no. 3, 2017, architecture, network security, and wireless sensor
Art. no. E549. networks.

Data-Driven - Methods - For - Predictive - Maintenance - of - Industrial - Equipment - A - Survey

Uploaded by

Copyright:

Available Formats

Data-Driven - Methods - For - Predictive - Maintenance - of - Industrial - Equipment - A - Survey

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Data-Driven - Methods - For - Predictive - Maintenance - of - Industrial - Equipment - A - Survey

Uploaded by

Copyright:

Available Formats

IEEE SYSTEMS JOURNAL, VOL. 13, NO.

3, SEPTEMBER 2019 2213

Data-Driven Methods for Predictive Maintenance of

ral network (NN) models) avoids the above complex feature

Fig. 4. PdM of degraded equipment [5], [6].

equipment, which is achieved by comprehensively utilizing the

acquisition; 3) feature engineering; and 4) modeling. It is impor-

C. Specific Case Based on Data-driven PdM

general, issues such as time and economic costs need to be

III. ML FOR PDM Then, the loss function of LR is defined as:

You might also like