An Overview of The Application of Machin
An Overview of The Application of Machin
PETROVIETNAM JOURNAL
Volume 10/2021, p. 47 - 61
ISSN 2615-9902
Summary
With the rise of industrial artificial intelligence (AI), smart sensing, and the Internet of Things (IoT), companies are learning how to
use their data not only for analysing the past but also for predicting the future. Maintenance is a crucial area that can drive significant cost
savings and production value around the world.
Predictive maintenance (PdM) is a technique that collects, cleans, analyses, and utilises data from various manufacturing and sensing
sources like machines usage, operating conditions, and equipment feedback. It applies advanced algorithms to the data, automatically
compares the fed data and the information from previous cases to anticipate or predict equipment failure before it happens, thus helping
optimise equipment utilisation and maintenance strategies, improve performance and productivity, and extend equipment life. Robust
PdM tools enable organisations to leverage and maximise the value of their existing data to stay ahead of potential breakdowns or
disruptions in services, and address them proactively instead of reacting to issues as they arise. Therefore, it has attracted more and more
attention of specialists in recent years.
This paper provides a comprehensive review of the recent advancements of machine learning (ML) techniques widely applied to PdM
by classifying the research according to the ML algorithms, machinery and equipment used in data acquisition. Important contributions
of the researchers are highlighted, leading to some guidelines and foundation for further studies. Currently, BIENDONG POC is running
some pilot PdM projects for critical equipment in Hai Thach - Moc Tinh gas processing plant.
Key words: Machine learning, predictive maintenance.
failure prediction in real-time, therefore reducing the Artificial intelligence (AI) is an element of Industry 4.0,
uncertainty of maintenance activities. showing its application in the manufacturing industry as
a powerful tool for machine health diagnosis. A recent
With the Fourth Industrial Revolution (Industry 4.0), IT
survey indicated that PdM is expected to be one of the first
infrastructure has been continuously improved to support
fields where AI-based technologies will be successfully
smart sensing, Internet of Things (IoT), big data collection,
implemented [9]. Indeed, it is almost impossible for a
and analytics tools; companies now can get the most
human operator to interpret data in real-time, and the
out of their data. Data-driven decisions can be made by
conventional systems can neither spot anomalies in data
analysing historical events and trends in the past [4, 5]. In
nor predict a sensor reading in a specified time window,
the case of complex equipment with a large amount of
AI and machine learning (ML) techniques have therefore
data, processing and analysing become more and more
emerged as a promising tool in PdM applications for
challenging for humans to handle. There are disadvantages
intelligent manufacturing in Industry 4.0, which involves
of human dependent scenarios, where humans monitor
rethinking and optimising the entire maintenance
equipment and make decisions manually for the work
strategy as a whole.
needed on the equipment. First, specialists’ performance
is highly dependent on their expertise levels. In a study As illustrated in Figure 1 machine learning is a
by Smith-Bindman R. et al. [6], less experienced physicians subfield of AI and defined as an algorithm or programme
tend to make false diagnostic decisions at a rate of 50% capable of learning independently with minimal or
higher than more experienced ones. This issue becomes without assistance from humans. ML supports solving
worse with industries that lack specialists who are well many complex human problems, such as image
trained and experienced. In addition, the consistency processing, big data, robotics, and speech recognition. By
of diagnostic decisions among specialists cannot be utilising ML in processing, continuously monitoring, and
guaranteed. A study by Gulshan et al. [7] compared analysing equipment’s health, corporates can achieve
decisions from seven US-certified ophthalmologists greater operational and maintenance efficiency and
regarding diagnosing the severity levels of diabetic effective growth management by improving equipment
retinopathy based on photographs of the retinal fundus. performance. A case study by Roosefert et al. show an
It showed poor consistency among ophthalmologists, in impressive 84% reduction of breakdown time and 88%
which only 20% of the cases demonstrating a complete reduction of breakdown occurrences when applying
agreement among them. Last but not least, specialists intelligent ML-based PdM approach compared to the
who analyse a large number of data for a long time can conventional Industry 3.0 setup [10]. This will also
make more mistakes and work less efficiently because of increase the availability and reliability of equipment and
mental fatigue and cognitive overload [8]. reduce operational risks, therefore improving corporates’
competitive advantages. With recent advances in ML
techniques, numerous studies have explored the potential
of the ML approach and demonstrated promising results.
For example, Akram et al. [11] developed the convolution
ARTIFICIAL INTELLIGENCE
neural networks (CNNs) for fault detection of photovoltaic
cell defects and achieved 93.02% accuracy. In another
study, Ren et al. [12] employed region-based CNN
(R-CNN) to automate object detection training in real-
time. Cha et al. [13] applied R-CNN to detect five types
MACHINE LEARNING
of structural surface damages on bridges and achieved
Algorithms or programmes capable of learning
independently with minimal or no additional a mean average precision (mAP) of 87.8%. The problem
assistance from humans of concern today regarding ML in PdM applications is to
choose the most suitable, simple, and effective algorithm
for each specific problem. ML algorithms often require
large data acquisition about failure situations and
different operating states of the system or device to train
the model.
Figure 1. Machine learning is a subset of artificial intelligence.
Continuous improve-
ment
Proactive
Fix it before it breaks
Routine repairs
Fix it after it breaks Predictive
Planned/ Preventa-
Performance measures
tive
Reactive/
Do nothing Breakdown
Defect elimination
Design out
Value focused
Predict faults
Run to failure Plan, schedule, coordinate Cost focused
Worry later Reliability focused
Lower costs
Longer equipment life Information gathering
PdM approaches have been extensively applied in In data acquisition phase, the measured data, such as
industries for handling the health status of equipment vibration, current, temperature, or pressure signals, are
by detecting early signs of failure in advance, enabling collected from sensors to monitor the health condition
maintenance measures to be taken ahead of time, which of machinery. In the next step, health indicators (HI)
allows the saving of more costs because repair after failure are constructed to represent the health condition of
is always more expensive than maintenance in advance machinery using statistics, signal processing, or artificial
[17]. According to data from McKinsey [18], PdM tools can intelligence techniques. Then, the HI will be analysed and
reduce manufacturing machine downtime by 30% - 50% classified into two or more health stages (HS) depending
and increase machine life by 20% - 40%. Besides, PdM on HI degradation trends. In the final step, the remaining
initiatives enable organisations around the world to save useful life is predicted to determine how much time is left
USD 17B in 2018 globally [19]. until machinery reaches the end of its lifetime, based on
which optimal maintenance activities can be scheduled.
In PdM, the optimal time point for maintenance
actions is predicted by analysing the system’s health 3.1. Data acquisition
state and historical maintenance data to make a timely
repair to avoid costly repairs due to system breakdown Data acquisition is a prerequisite from which raw
and premature maintenance activities which may induce data can be processed and analysed later to solve a PdM
infant mortality. The PdM approach is used in industrial problem. Data acquisition involves designing and sizing
sectors where reliability is paramount, like nuclear power the appropriate system architecture to be installed on
plants, transportation systems, or emergency systems [3]. the equipment to capture and store different sensing
The most used monitoring and diagnostic techniques data. A data acquisition system consists of sensors, data
include vibration monitoring, thermography, tribology, transmission, and data storage devices. Depending on
and visual inspection. equipment types, different combinations of sensors
are used so that the captured data can fully reflect the
On the other hand, proactive maintenance emphasises degradation process of machinery. Some commonly used
keeping assets and equipment in top conditions by sensors are accelerometers, acoustic emission sensors,
identifying potential problems and addressing them as infrared thermometers, current, temperature, and
early as possible. That is one of the main reasons why pressure sensors, etc. The captured data are transmitted
proactive maintenance is considered different from other into a PC or portable devices through a data transmission
approaches that help reduce maintenance costs and device and stored in a memory location or historian
risk of property damage. Some examples of proactive system for further analysis. With a rapid development of
maintenance activities are: (i) applying anti-corrosion sensor and communication technologies, more advanced
coating to metal surfaces; (ii) performing an inspection data acquisition devices have been designed and applied
for cracks, leaks, and rust; (iii) lubricating the machine to modern industries.
to reduce wear and diffusion corrosion and (iv) applying
sealant for junction boxes to prevent water ingress. In One of the essential data acquisition techniques
proactive maintenance approach, possible causes of often discussed in ML algorithms is data pre-processing
equipment failure are analysed, and works are carried techniques. In this step, commonly used techniques for
out to eliminate those causes in very early stages (e.g., data pre-processing include data normalisation and noise
designing, installation, and commissioning stages), which removal, feature extraction, and feature selection. Feature
definitely distinguish proactive maintenance from other extraction is the process of mapping the original dataset
approaches. to new data space, including new attributes that contain
more explicit information for performing PdM. There are
3. ML approach in PdM two basic feature extraction methods commonly applied
in PdM: data-driven and statistical techniques.
In PdM approach, equipment health is continuously
monitored and analysed in real-time. Lei et al. [20] With the data-driven technique, these feature
divide the process of diagnosing heath of machinery extraction processes are performed automatically based
into four steps: (i) data acquisition; (ii) health indicators on specific data sets (data-driven) using a machine
construction; (iii) health stage division; and (iv) remaining learning model. The weakness of this approach is that
useful life prediction. a lot of data will be needed to train the model, and the
user may not be able to be interpreted or described the 3.2. Health indicator construction
extracted attributes. Several proven examples of machine
After extracting the features, the next step is to
learning tools used for feature extraction include CNN
determine the health indicator. Indeed, a health indicator
[21, 22], autoencoder [22 - 24], or principal component
represents the health state of machinery in real-time,
analysis [25 - 27].
taking into account different kinds of condition monitoring
Feature extraction based on statistical theory is a signals, such as vibration, current, and acoustic emission
classic and commonly used method in data studies. signals. HI construction plays a significant role in PdM. A
Statistical methods will usually be used first in surveying suitable HI can help simplify the prognostic modelling
and understanding system performance trends from and produce a more accurate prediction. In addition,
sensor data. Statistical methods can be performed in from the HI information, specific methods can be devised
three types [27]: time domain, frequency domain, and for estimating the remaining useful life of the device. HI
time-frequency domain. Statistical methods in the time can be derived from monitoring signals using statistical
domain include monotony, trend, consistency, similarity, methods or signal processing methods, such as RMS of
stability, or correlation. Many techniques can be used vibration signals. In fact, machines are often monitored
to extract these features, including peak value, the by multiple sensors. While analysing a high-dimensional
difference between the highest and the lowest values, dataset, different AI techniques can be employed to fuse
mean absolute value, root mean square (RMS), crest factor, those multi-sensor signals into an indicator representing
standard deviation, kurtosis, or distributed shape factor. the machinery's degradation trend.
The feature extraction of the time domain technique have
In a study by Lei et al. in 2018, there are two main
many advantages - it is fast, simple, and can be used for
techniques to develop the HI [30], including (i) physical
many different types of device failures. The disadvantage
HI (PHI) relating to physical faults of the machine and
of time-domain feature extraction is that it is susceptible
usually extracted from the sensor using statistical or
to noise and often requires data pre-processing before
data processing techniques; (ii) virtual HI (VHI) built by
extraction. The advantage of feature extraction in
combining many physical indicators or the different
the frequency domain technique is its efficiency in
sensors. VHI usually carries no physical implication
representing anomaly information and equipment
about the device but only information about the failure
errors, which are not visible in the time domain. The
tendency of the device in a virtual manner.
disadvantage of the frequency domain technique is the
need for specialised knowledge of device failures in the Principal component analysis (PCA) is one of the most
frequency domain, and this method cannot be universally popular techniques for HI construction. PCA is a linear
applicable to all types of devices. The time-frequency dimension reduction technique that uses an orthogonal
domain technique is a powerful tool that allows analysing transformation to convert a set of observations of possibly
and representing essential features of the data spectrum correlated or dependent variables into a set of linearly
over time. Time-frequency domain feature extraction uncorrelated variables called principal components.
method is generally not affected by noise. However, some Some of the significant studies are the construction of
of its disadvantages are high computational cost, and an HI using PCA to reduce the dimensions of the feature
the difficulty in determining the parameters of the data sets and subsequently calculate the deviations between
transformation process from the original time domain to unknown states and the healthy state [25], using PCA
the time-frequency domain. Statistical approaches have combined with isometric feature mapping to construct
been widely applied in predictive intelligence problems an HI for cutting tools [26], and fusing multiple features to
and achieved many positive results [28, 29]. calculate the T2 statistics as an HI of bearings [27].
After collecting these values, we can do one more Self-organising map (SOM) is another widely
step called feature fusion or feature aggregation. This used technique for HI construction. The SOM is a non-
step combines the features extracted from the above supervised learning neural network that produces a
statistical method to create more complex features with low-dimensional representation of a higher-dimensional
better data representation. For example, we can use other dataset by organising itself according to the nature
mathematical formulas [28] or genetic algorithms [29] to of the input data. The SOM technique was introduced
combine features. into HI construction by Qiu et al. [31] for the condition
monitoring of roller bearing. Some authors also used the and eight time-frequency features using the recurrent
SOM technique for HI construction. Hong et al. [32] used neural network.
wavelet packet decomposition (WPD) and empirical mode
Besides using mathematical functions to define HI,
decomposition (EMD) to extract entropy sequences from
another way is to compare the current state of the device
original vibration signals and use them as input vectors
to be monitored to that of a similar one. The selected
of the SOM network. The confidence value derived from
devices are also working in the same environment and
the SOM is then used as an HI of bearing. Lei et al. [33]
performing the same task. When the operating envelope
trained the SOM by the feature vector in the operational
of the two devices is the same, their HI will be the same.
stage. Then the feature vector under an unidentified
condition is compared with the weight vector of its best 3.3. Health stage division
matching unit (BMU) in the SOM. Liao et al. [34] used
the Restricted Boltzmann Machine algorithm to extract Staged machine operation is the next step after
features from the dataset and compare them with the determining HI. However, it is not always easy to separate
weight vectors of all the units in the baseline map. The the working states of the machine. In particular, if the
distance to the baseline map is calculated as minimum device continuously works with a linear HI from start to
quantisation error (MQE), used as the machine’s health finish, it is impossible to subdivide the device's different
value. And Huang et al. [35] trained the SOM network working stages. In that case, the working state of the
using 6 vibration features and used the MQE indicator device changes uniformly at a constant rate [30]. However,
derived from SOM as an HI. it does not mean that this is an unimportant step in the
PdM application. With many other types of mechanical
Mahalanobis distance (MD) is used in some
devices (such as journal bearings), the operating state of
publications as an HI construction technique. MD
the equipment can be divided into two or more stages.
generalises multivariate dataset by finding how many
With these devices, in the healthy stage, the HI readings
standard deviations away a point is from the mean of
are nearly constant, so there is no information about the
the multivariate distribution. Wang et al. [36] utilised
failure trend of the device. Forecasting the RUL at this stage
Mahalanobis distance to construct HI from 14 different
would be neither accurate nor necessary. RUL forecasting
statistics into a new feature to reflect the degradation
should only begin to be performed after the machine
of the bearing. Jin et al. [37] calculated the energies of
transitions from normal operation to deterioration due
wavelet coefficients and fused them by calculating their
to failure or a so-called unhealthy state. Then, the division
Mahalanobis distance with reference to the healthy
of working states makes sense and is used to determine
one. Kumar et al. [38] also constructed an HI based on
the device's initial degradations and provide a reasonable
Mahalanobis distance to monitor bearings.
transition time for the RUL estimation in PdM.
Researchers also explored different combinations of
HIs of machinery generally present varying
AI techniques for HI construction. Ocak et al. [39] used the
degradation trends with the development of fault severity.
wavelet packet decomposition (WPD), a time-frequency
Subsequently, the degradation process of machinery can
domain technique, in connection with the hidden Markov
be divided into different health stages (HS) according to
modelling (HMM) technique to develop a method for
the varying trends of HI. HS division is vital in machinery
real-time tracking of bearing health and prognostic. HHM
health diagnostics because it identifies the start time of
has also been successfully applied to many other fields,
the unhealthy stage, also known as the first predicting
such as tool wear condition monitoring [40] and bearing
time (FPT), and triggers the remaining useful life (RUL)
diagnosis [41]. Shen et al. [42] constructed an HI that
prediction [20].
reflects the running state of the rolling bearing in real-
time and effectively guarantees the operation reliability of In one-stage HS, the health of machinery shows
bearings using the fuzzy support vector data description gradual degradation, which can be described using a single
(FSVDD) technique. Liu et al. [43] proposed an HI for degradation model. In this case, the HS division process is
bearings through phase space reconstitution combined not required. In two-stage HS, the degradation trend shows
with approximate diagonalisation of eigen-matrices. Guo two distinct stages: the healthy stage, where no fault occurs,
et al. [44] constructed an HI that contains rich degradation and the unhealthy stage, where there is an accelerated
signatures of bearings from six related-similarity features degradation trend over time. In multiple-stage HS, the
by-step status but a set of probability distributions. variability, and measurement variability [59]. Therefore,
HMM is also applied in the studies of Giantomassi et al. the statistical model-based approaches effectively
[49], Ramasso et al. [50], and Sloukia et al. [51]. Some AI describe the uncertainty of degradation process and its
classifiers are applied to multi-stage HS division as well, influence on RUL prediction.
such as artificial neural networks [52 - 54] and support
AI approaches attempt to learn the machinery
vector machine [55, 56].
degradation patterns using AI techniques from available
3.4. Remaining useful life prediction observations instead of building physical models or
statistical models. They can deal with prognostic issues
Prediction of the remaining useful life (RUL) of of complex mechanical systems whose degradation
machinery is the most critical problem in the field of PdM. processes are challenging to be interrelated by physical
The RUL of machinery is defined as “the length from the models or statistical models. Therefore, they are attracting
current time to the end of the useful life” [57]. The primary more and more attention in the field of machinery
task of RUL prediction is to forecast the time left before prognostics. The results of AI approaches are hard to
the machinery loses its operation ability based on the explain because of the lack of transparency; thus, these
condition monitoring information. It is the last technical techniques are consistently named “black boxes” [20]. The
process as well as the last goal of machinery prognostics. commonly used AI techniques in machinery prognostics
Approaches for RUL prediction can be categorised into include artificial neural networks, neuro-fuzzy systems,
physical model-based approaches, statistical model- support vector machine, K-nearest neighbour, and
based approaches, and AI approaches. Gaussian process regression (GPR), etc.
Physical model-based is a classical approach that The artificial neural network is an ML reasoning
describes degradation processes of machinery through technique inspired by the working process of human
building mathematical models based on the failure brains. It is a collection of nodes, also known as artificial
mechanisms or the first principal damage [58]. The neurons, in a complex structure of the input, hidden,
parameters of the physical models reflect the material and output layers. In ANNs, different layers may perform
properties and stress levels, which are generally obtained different transformations on their inputs. Input layer
by experiments or finite element analysis. To provide an nodes pass information to hidden layer nodes by
accurate estimation of the RUL physical model-based activation functions. Subsequently, the hidden layers
approach, it is required that the model is developed with apply weighting functions until some threshold of the
a complete understanding of the failure mechanisms hidden layer is reached. Finally, the result is passed to the
and effective estimation of model parameters. However, output layer. Because ANNs can learn complex non-linear
it is difficult for some complex mechanical systems to relationships by training the multi-layer networks, they
understand the physics the damage, which restricts the have a good performance in dealing with complex systems
application of these approaches [20]. and become the most used AI techniques in machinery
Statistical model-based approaches, also known as RUL prediction. Two widely used ANNs are feed-forward
empirical model-based approaches, estimate the RUL of neural network and recurrent neural network, which
machinery by establishing statistical models based on are classified based on the architecture of layers in the
empirical knowledge [57]. The statistical model-based network. The difference is illustrated in Figure 4.
approach does not rely on physics to construct RUL The feed-forward neural network (FFNN) is a class
prediction models. Rather, the RUL is predicted by fitting of ANNs, where inputs are processed forward. Sbarufatti
available observations into random coefficient models or et al. [60] combined FFNNs with sequential Monte-Carlo
stochastic process models under a probabilistic method. sampling to predict the RUL of fatigue cracks. Pan et al. [61]
In the RUL problem, this method uses probabilistic and Xiao et al. [62] used an FFNN to conduct multi-step
statistical models to represent the relationship between ahead prediction for the bearing health states. Wang et al.
variables in the state of the device. Random variances are [63] used a three-layer FFNN to predict the future HIs and
generally introduced into model parameters to describe input the predicted HIs into a PH model to estimate the
the uncertainties caused by different kinds of variability hazard rate and survival probability. The results showed
sources, such as temporal variability, unit-to-unit that RUL can be predicted with high accuracy in the
incipient degradation phase and trace the degradation fuzzy system trained by algorithms derived from neural
well with a limited amount of data. networks. The basic idea behind this NF system is that it
combines the human-like reasoning style of fuzzy systems
On the other hand, the recurrent neural network (RNN)
with neural networks' learning and connectionist structure
is another class of ANNs that have many applications
[69]. The NF-based time-series forecasting approach was
in solving time-series or sequential data problems.
proposed by Jang et al. [70]. Wang et al. [71] adopted the
Unlike FFNN, which assumes inputs and outputs are
NF system to develop an online prognostic approach
independent, RNN has a recurrent connection in the
for different kinds of gear faults. Wang [72] further
hidden state that makes the output of RNN depend on
improve this approach from several aspects to enhance
the prior or previous elements within the sequence. RNN
its forecasting performance. NF systems take advantage
method is widely used in RUL prediction because of its
of both the expert knowledge and the intelligent ANNs,
ability to deal with detailed time-series data. As illustrated
thus being competitive candidates for machinery RUL
in many studies, this method showed good performance,
prediction [73 - 77]. However, they still need lots of high-
such as the recurrent radial basis function network to
quality training data.
predict the RUL of machinery by Zemouri [64]. Malhi et
al. developed a competitive learning-based approach to Support vector machine (SVM) is a kind of ML
revise the training technique of RNNs that helps improve technique based on the statistical learning theory
the long-term prediction accuracy [65]. In addition, Peng proposed by Vapnik [78]. Different kinds of SVM have
et al. improved the RNN model by replacing a hidden been applied to the RUL prediction of machinery, such as
layer using a large sparse reservoir to develop a new RUL the least square-SVM [79], one-class SVM [80], and multi-
prediction approach [66]. Liu et al. proposed an enhanced class SVM [81]. Widodo et al. [82] trained an SVM model
RNN for RUL prediction by improving RNNs’ memory using both the censored and the complete data and
property [67]. predicted the survival probability of machinery. Tran et al.
[83] integrated an SVM-based RUL prediction module into
The ANN networks still have limitations in terms of
an intelligent condition-based maintenance platform. In
their low transparency and the requirement of a large
addition, support vector regression (SVR) is the standard
dataset for high-quality training process, which are
application form of SVM in prognostics [83]. Benkedjouh
difficult to capture in industrial applications. In addition,
et al. [84] used the SVR to map the HIs into non-linear
their structures and parameters are generally initialised
regressions and then fitted the obtained regressions
randomly or specified manually, which reduces their
into power models for RUL prediction of machinery.
generalisation ability among different cases [20].
Liu et al. [85] developed a modified probabilistic SVR to
The neural fuzzy (NF) approach is based on the predict the degradation processes of nuclear power plant
(a) (b)
Figure 4. Comparison of RNN (a) and FFNN (b) [68].
components. Fumeo et al. [86] developed an online SVR on advanced technology or requiring intensive training
model for the RUL prediction of bearings by optimising to use. Recently, PdM applications often used AI and ML.
the trade-off between accuracy and the computing Due to the lack of specialised research facilities for PdM
efficiency. in Vietnam, there will be many difficulties and challenges
for the technology to be applied in the country. However,
Compared to ANNs, SVM are superior in dealing
the benefits it brings if successfully implemented are very
with the issues of small sample sizes. Thus, they may be
desirable.
more suitable for the issues of RUL prediction where
only limited measurements are available. However, the 4. Conclusion
performance of SVM is highly dependent on the selected
kernel functions and parameter optimisation process. PdM practices become more and more popular in
many industrial sectors because of their effectiveness
Gaussian process regression (GPR) is an ML technique
in reducing unnecessary maintenance operations
that implements Gaussian processes for regression
and improving machinery reliability. With the recent
purposes [87]. Gaussian processes are cumulative damage
advancement of Industry 4.0, ML has found wide
processes of random variables with joint multivariate
applications in PdM practice to assist humans in
Gaussian distributions. Many regression models have
processing, continuously monitoring and analysing
been published recently, such as introducing theoretical
equipment’s health. This paper has discussed the practice
details and the flexibility of the GPR in non-linear
of PdM and reviewed the applications of some state-of-
regression [88], predicting RUL by GPR [89], applying
the-art AI techniques that have been studied in this field.
GPR with three different covariance functions for RUL
Different AI techniques are structurally discussed in the
prediction of bearings [32] and predicting the degradation
four technical steps of the machinery health prognostics
trends of rolling element bearings using an integrated
process, including data acquisition, HI construction, HS
GPR model [90]. In contrast to the above AI techniques,
division, and RUL prediction.
GPR is highly adaptable and suitable for RUL prediction
problem of high-dimension and small-size datasets [91]. BIENDONG POC is aiming to deploy PdM solutions
The major drawback of GPR is that it generally requires for monitoring, maintaining, and tracking equipment's
heavy computational work. reliability and performance in a plant by ML modelling.
The monitoring and quick identification of failures help
3.5. Challenges for PdM using ML in Vietnam make decisions and plans for maintenance and enable
support for continuous performance, health surveillance,
In essence, each device has many different failure
and focus on performance degradation as a leading
modes and must be tested in laboratory to produce the
indicator of potential equipment problems. Therefore, it is
most complete and standard datasets. During normal
necessary to review research achievements and trends in
operating conditions, these figures are difficult to obtain.
PdM to assess the current status and orientate the focus of
There can only be specific failure modes. In addition, it is
research at BIENDONG POC.
necessary to aggregate data on equivalent machines. It
is challenging to compare and synthesise data in many 5. Acknowledgement
cases due to the unavailability of machines in most
companies. Where possible, oil and gas companies, This research work is described herein as part of the
especially in Vietnam, need to collaborate and share government research project number 077.2021.CNKK.
data about machinery to support the development and QG/HĐKHCN, order 196/QD-BCT of Vietnam Ministry of
research of PdM. Industry and Trade, Viet Nam.
and fault prognosis for bearings", IEEE Transactions on [46] Kamran Javed, Rafael Gouriveau, and
Instrumentation and Measurement, Vol. 65, pp. 2046 - 2054, Noureddine Zerhouni, "A new multivariate approach for
2016. DOI: 10.1109/TIM.2016.2570398. prognostics based on extreme learning machine and fuzzy
clustering", IEEE Transactions on Cybernetics, Vol. 45, No. 12,
[38] Kumar H S, P. Srinivasa Pai, N. S. Sriram, and
pp. 2626 - 2639, 2015. DOI: 10.1109/TCYB.2014.2378056.c
Vijay Gs, "Rolling element bearing fault diagnostics:
Development of health index", Proceedings of the Institution [47] Zhilian Liu, Ming J. Zuo, and Yong Qin,
of Mechanical Engineers, Part C: Journal of Mechanical "Remaining useful life prediction of rolling element
Engineering Science 1989 - 1996, Vol. 231, No. 21, pp. 3923 - bearings based on health state assessment", Proceedings
3939, 2017. DOI: 10.1177/0954406216656214. of the Institution of Mechanical Engineers, Part C: Journal of
Mechanical Engineering Science, Vol. 230, No. 2, pp. 314 -
[39] Hasan Ocak, Kenneth A. Loparo, Fred M.
330, 2016. DOI: 10.1177/0954406215590167.
Discenzo, "Online tracking of bearing wear using wavelet
packet decomposition and probabilistic modeling: A [48] Patricia Scanlon, Darren F. Kavanagh, and
method for bearing prognostics", Journal of Sound and Francis Boland, "Residual life prediction of rotating
Vibration, Vol. 302, No. 4 - 5, pp. 951 - 961, 2007. DOI: machines using acoustic noise signals", IEEE Transactions
10.1016/j.jsv.2007.01.001. on Instrumentation and Measurement, Vol. 62, pp. 95 - 108,
2012. DOI: 10.1109/TIM.2012.2212508.
[40] Huseyin M. Ertunc, Kenneth A. Loparo, Hasan
Ocak, "Tool wear condition monitoring in drilling [49] Andrea Giantomassi, Francesco Ferracuti,
operations using hidden Markov models (HMMs)", Alessandro Benini, Gianluca Ippoliti, Sauro Longhi, and
International Journal of Machine Tools and Manufacture, Antonio Petrucci, "Hidden Markov model for health
Vol. 41, No. 9, pp. 1363 - 1384, 2001. DOI: 10.1016/S0890- estimation and prognosis of turbofan engines", ASME 2011
6955(00)00112-7. International Design Engineering Technical Conferences and
Computers and Information in Engineering Conference, 28 -
[41] Hasan Ocak and Kenneth A. Loparo, "HMM-
31 August, 2011. DOI: 10.1115/DETC2011-48174.
based fault detection and diagnosis scheme for rolling
element bearings", Journal of Vibration and Acoustics, Vol. [50] E. Ramasso and T. Denoeux, "Making use of
127, No. 4, pp. 299 - 306, 2005. DOI: 10.1115/1.1924636. partial knowledge about hidden states in HMMs: An
approach based on belief functions", IEEE Transactions on
[42] Zhongjjie Shen, Zhengjja He, Xuefeng Chen, Fuzzy Systems, Vol. 22, pp. 395 - 405, 2014.
Chuang Sun, and Zhiwen Liu, "A monotonic degradation
assessment index of rolling bearings using fuzzy support [51] F. Sloukia, R. Bouarfa, H. Medromi, and M. Wahbi,
vector data description and running time", Sensors, Vol. 12, "Bearings prognostic using mixture of Gaussians hidden
No. 8, pp. 10109 - 10135, 2012. DOI: 10.3390/s120810109. Markov model and support vector machine", International
Journal of Network Security & Its Applications, Vol. 5, pp. 85
[43] Yan-Cheng Liu, Siyuan Liu, and Ning Wang, - 97, 2013.
"Fully-tuned fuzzy neural network based robust adaptive
[52] P. Tamilselvan, Yibin Wang, and P. Wang, "Deep
tracking control of unmanned underwater vehicle with
belief network based state classification for structural
thruster dynamics", Neurocomputing, Vol. 196, pp. 1 - 13,
health diagnosis", 2012 IEEE Aerospace Conference,
2016. DOI: 10.1016/j.neucom.2016.02.042.
Montana, 3 - 10 March, 2012.
[44] Guo Liang, Li Naipeng, Jia Feng, Lei Yaguo,
[53] Liang Guo, Hongli Gao, Haifeng Huang, Xiang
and Lin Jing, "A recurrent neural network based health
He, and ShiChao Li, "Multifeatures fusion and non-linear
indicator for remaining useful life prediction of bearings",
dimension reduction for intelligent bearing condition
Neurocomputing, Vol. 240, pp. 98 - 109, 2017. DOI: 10.106/j.
monitoring", Shock and Vibration, Vol. 2016, pp. 1 - 10,
neucom.2017.02.045.
2016. DOI: 10.1155/2016/4632562.
[45] Emmanuel Ramasso, Michèle Rombaut, and
[54] Jaouher Ben Ali, Brigitte Chebel-Morello, Lotfi
Noureddine Zerhouni, "Joint prediction of continuous and
Saidi, Simon Malinowski, and Farhat Fnaiech, "Accurate
discrete states in time-series based on belief functions",
bearing remaining useful life prediction based on Weibull
IEEE Transactions on Cybernetics, Vol. 43, No. 1, pp. 37 - 50,
distribution and artificial neural network", Mechanical
2013.
Systems and Signal Processing, Vol. 56 - 57, pp. 150 - 172, An approach combining a RRBF network and an
2015. autoRegressive model", IFAC Proceedings Volumes, Vol. 43,
No. 3, pp. 140 - 145, 2010.
[55] A. Soualhi, K. Medjaher, and N. Zerhouni,
"Bearing health monitoring based on Hilber-Huang [65] Arnaz Malhi, Ruqiang Yan, and Robert X. Gao,
transform, support vector machine, and regression", IEEE "Prognosis of defect propagation based on recurrent
Transactions on Instrumentation and Measurement, Vol. 64, neural networks", IEEE Transactions on Instrumentation and
pp. 52 - 62, 2015. Measurement, Vol. 60, pp. 703 - 711, 2011. DOI: 10.1109/
TIM.2010.2078296.
[56] Hack-Eun Kim, Andy C.C. Tan, Joseph Mathew,
and Byeong-Keun Choi, "Bearing fault prognosis based on [66] Yu Peng, Hong Wang, Jianmin Wang, Datong
health state probability estimation", Expert Systems with Liu, and Xiyuan Peng, "A modified echo state network
Application, Vol. 39, pp. 5200 - 5213, 2012. based remaining useful life estimation approach", 2012
IEEE Conference on Prognostics and Health Management,
[57] Xiao-Sheng Si, Wenbin Wang, Chang-Hua Hu,
2012.
Dong-Hua Zhou, "Remaining useful life estimation - A
review on the statistical data driven approaches", European [67] Datong Liu, Wei Xie, Haitao Liao, and Yu Peng,
Journal of Operational Research, Vol. 213, pp. 1 - 14, 2011. "An integrated probabilistic approach to Lithium-Ion
battery remaining useful life estimation", IEEE Transactions
[58] Y. Li, S. Billington, C. Zhang, T. Kurfess, S. Danyluk,
on Instrumentation and Measurement, Vol. 64, pp. 660 -
and S. Liang, "Adaptive prognostics for rolling element
670, 2015.
bearing condition", Mechanical Systems and Signal
Processing, Vol. 13, No. 1, pp. 103 - 113, 1999. [68] IBM Cloud Education, “Recurrent neural
networks”, 2020. [Online]. Available: https://www.ibm.
[59] Yaguo Lei, Naipe Li, and Jing Lin, "A new
com/cloud/learn/recurrent-neural-networks.
method based on stochastic process models for machine
remaining useful life prediction", IEEE Transactions on [69] Samarjit Kar, Sujit Das, and Pijush Ghosh,
Instrumentation and Measurement, Vol. 65, pp. 2671 - 2684, "Applications of neuro fuzzy systems: A brief review and
2016. future outline", Applied Soft Computing, Vol. 15, pp. 243 -
259, 2014.
[60] C. Sbarufatti, M. Corbetta, A. Manes, and M.
Giglio, "Sequential Monte-Carlo sampling based on a [70] Jyh-Shing Roger Jang, Chuen-Tsai Sun, and Eiji
committee of artificial neural networks for posterior state Mizutani, Neuro-fuzzy and soft computing: A computational
estimation and residual lifetime prediction", International approach to learning and machine intelligence. Prentice-
Journal of Fatigue, Vol. 83, pp. 10 - 23, 2016. Hall Inc., 1996.
[61] Yongping Pan, Meng Joo Er, Xiang Li, Haoyong [71] Wilson Q. Wang, M. Farid Golnaraghi, and Fathy
Yu, and Rafael Gouriveau, "Machine health condition Ismail, "Prognosis of machine health condition using
prediction via online dynamic fuzzy neural networks", neuro-fuzzy systems", Mechanical Systems and Signal
Engineering Applications of Artificial Intelligence, Vol. 35, Processing, Vol. 18, No. 4, pp. 813 - 831, 2004. DOI: 10.1016/
pp. 105 - 113, 2014. s0888-3270(03)00079-7.
[62] Lei Xiao, Xiaohui Chen, Xinghui Zhang, and Min [72] Wilson Wang, "An adaptive predictor for
Liu, "A novel approach for bearing remaining useful life dynamic system forecasting", Mechanical Systems and
estimation under neither failure nor suspension histories Signal Processing, Vol. 21, No. 2, pp. 809 - 823, 2007.
condition", Journal of Intelligent Manufacturing, Vol. 28, pp. DOI:10.1016/j.ymssp.2005.12.008.
1893 - 1914, 2017. DOI: 10.1007/s10845-015-1077-x.
[73] Fagang Zhao, Jin Chen, Lei Guo, and Xinglin
[63] Lu Wang, Li Wang, Xue-zhi Wang, "Reliability Li, "Neuro-fuzzy based condition prediction of bearing
estimation and remaining useful lifetime prediction for health", Journal of Vibration and Control, Vol. 15, No. 7, pp.
bearing based on proportional hazard model", Journal of 1079 - 1091, 2009. DOI:10.1177/1077546309102665.
Central South University, Vol. 22, pp. 4625 - 4633, 2015.
[74] Van Tung Tran, Bo-Suk Yang, and Andy Chit
[64] Ryad Zemouri and Rafael Gouriveau, "Towards Chiow Tan, "Multi-step ahead direct prediction for the
accurate and reproducible predictions for prognostic: machine condition prognosis using regression trees and
neuro-fuzzy systems", Expert Systems with Applications, [83] Van Tung Tran and Bo-Suk Yang, "An intelligent
Vol. 36, No. 5, pp. 9378 - 9387, 2009. DOI: 10.1016/j. condition-based maintenance platform for rotating
eswa.2009.01.007 machinery", Expert Systems with Applications, Vol. 39, No.
3, pp. 2977 - 2988, 2012. DOI:10.1016/j.eswa.2011.08.159.
[75] Chaochao Chen, Bin Zhang, George
Vachtsevanos, and Marcos E.Orchard, "Machine condition [84] T. Benkedjouh, K. Medjaher, N. Zerhouni, and S.
prediction based on adaptive neuro-fuzzy and high-order Rechak, "Health assessment and life prediction of cutting
particle filtering", IEEE Transactions on Industrial Electronics, tools based on support vector regression", Journal of
Vol. 58, No. 9, pp. 4353 - 4364, 2011. Intelligent Manufacturing, Vol. 26, pp. 213 - 223, 2015. DOI:
10.1007/s10845-013-0774-6.
[76] Daniel Zurita Millan, Jesus A.Carino, Miguel
Delgado-Prieto, and Juan Antonio Redondo, "Distributed [85] Jie Liu and Enrico Zio, "An adaptive online
neuro-fuzzy feature forecasting approach for condition learning approach for support vector regression: online-
monitoring", 2014 IEEE Emerging Technology and Factory SVR-FID", Mechanical Systems and Signal Processing, Vol.
Automation (ETFA), 2015. 76, pp. 796 - 809, 2016. DOI: 10.1016/j.ymssp.2016.02.056.
[77] Sajid Hussain and Hossam A.Gabbar, "Vibration [86] Emanuele Fumeo, Luca Oneto, Davide Anguita,
analysis and time series prediction for wind turbine "Condition based maintenance in railway transportation
gearbox prognostics", International Journal of Prognostics systems based on big data streaming analysis", Procedia
and Health Management, Vol. 4, No. 3, 2013. DOI: 10.36001/ Computer Science, Vol. 53, pp. 437 - 446, 2015. DOI:
ijphm.2013.v4i3.2144. 10.1016/j.procs.2015.07.321.
[78] Vladimir N.Vapnik, "An overview of statistical [87] W. J. Padgett and Meredith A. Tomlinson,
learning theory", IEEE transactions on neural networks, Vol. "Inference from accelerated degradation and failure data
10, No. 5, pp. 988 - 999, 1999. based on Gaussian process models", Lifetime Data Analysis,
Vol. 10, pp. 191 - 206, 2004.
[79] Shaojiang Dong and Tianhong Luo, "Bearing
degradation process prediction based on the PCA [88] Carl Edward Rasmussen, "Gaussian processes in
and optimized LS-SVM model", Measurement, Vol. machine learning", Advanced Lectures on Machine Learning,
46, No. 9, pp. 3143 - 3152, 2013. DOI: 10.1016/j. pp. 63 - 71, 2003. DOI: 10.1007/978-3-540-28650-9_4.
measurement.2013.06.038.
[89] Kai Goebel, Bhaskar Saha, and Abhinav Saxena,
[80] J. Carino, D. Zurita, Miguel Delgado Prieto, J. A. "A comparison of three data-driven techniques for
Redondo, and R. Romero-Troncoso, "Remaining useful life prognostics", 2008.
estimation of ball bearings by means of monotonic score
[90] S.A. Aye, P.S. Heyns, "An integrated Gaussian
calibration", IEEE International Conference on Industrial
process regression for prediction of remaining useful
Technology (ICIT), 17 - 19 March 2015. DOI: 10.1109/
life of slow speed bearings based on acoustic emission",
ICIT.2015.7125351.
Mechanical Systems and Signal Processing, Vol. 84, pp. 485 -
[81] James K. Kimotho, Christoph Sondermann- 498, 2017. DOI: 10.1016/j.ymssp.2016.07.039.
Woelke, Tobias Meyer, and Walter Sextro, "Machinery
[91] Man Shan Kan, Andy C.C. Tan, and Joseph
prognostic method based on multi-class support vector
Mathew, "A review on prognostic techniques for non-
machines and hybrid differential evolution particle swarm
stationary and non-linear rotating systems", Mechanical
optimization", Chemical Engineering Transactions, Vol. 33,
Systems and Signal Processing, Vol. 62 - 63, pp. 1 - 20, 2015.
pp. 619 - 624, 2013.
DOI: 10.1016/j.ymssp.2015.02.016
[82] Achmad Widodo and Bo-Suk Yang, "Machine
health prognostics using survival probability and support
vector machine", Expert Systems with Applications,
Vol. 38, No. 7, pp. 8430 - 8437, 2011. DOI: 10.1016/j.
eswa.2011.01.038.