Predictive Maintenance for Pump Systems and
Thermal Power Plants: State-of-the-Art Review,
Trends and Challenges
Jonas Fausing Olesen 1,2 and Hamid Reza Shaker 1,*
1Center for Energy Informatics, University of Southern Denmark, 5230 Odense, Denmark; email@example.com
2Ørsted, Markets & Bioenergy, Asset Risk Management, Kraftværksvej 53, 7000 Fredericia, Denmark
Received: 8 April 2020; Accepted: 21 April 2020; Published: 24 April 2020
Thermal power plants are an important asset in the current energy infrastructure,
delivering ancillary services, power, and heat to their respective consumers. Faults on critical
components, such as large pumping systems, can lead to material damage and opportunity losses.
Pumps plays an essential role in various industries and as such clever maintenance can ensure
cost reductions and high availability. Prognostics and Health Management, PHM, is the study
utilizing data to estimate the current and future conditions of a system. Within the ﬁeld of PHM,
Predictive Maintenance, PdM, has been gaining increased attention. Data-driven models can
be built to estimate the remaining-useful-lifetime of complex systems that would be difﬁcult to
identify by man. With the increased attention that the Predictive Maintenance ﬁeld is receiving,
review papers become increasingly important to understand what research has been conducted and
what challenges need to be addressed. This paper does so by initially conceptualising the PdM ﬁeld.
A structured overview of literature in regard to application within PdM is presented, before delving
into the domain of thermal power plants and pump systems. Finally, related challenges and trends
will be outlined. This paper ﬁnds that a large number of experimental data-driven models have
been successfully deployed, but the PdM ﬁeld would beneﬁt from more industrial case studies.
Furthermore, investigations into the scale-ability of models would beneﬁt industries that are looking
into large-scale implementations. Here, examining a method for automatic maintenance of the
developed model will be of interest. This paper can be used to understand the PdM ﬁeld as a broad
concept but does also provide a niche understanding of the domain in focus.
machine learning; predictive maintenance; remaining useful lifetime; state of the
Production facilities are built with increasing complexity to satisfy the increase in demand for
quality products and availability. To deliver on this, manufacturers need to understand the current state
of their assets to avoid unnecessary downtime [
]. The maintenance procedure can be compared to the
health care procedures conducted on man. If conducted efﬁciently, the system will be operating under
fewer restrictions and prove to have higher performance. This means cost reductions, resource-saving,
increased life expectancy, and reduced downtime. On the other hand, poor maintenance management
can result in a reduction in safety, reputation, and quality of product delivered [
]. Hence, the question
becomes when maintenance should be deployed. Traditionally maintenance had been done either
reactive or preventive [
]. The reactive maintenance approach allows components and machinery to
run till failure. The consequence is long downtime and thus a market opportunity loss. The preventive
maintenance approach deploys maintenance based on a time or cycle metric. This is seen in cars
Sensors 2020,20, 2425; doi:10.3390/s20082425 www.mdpi.com/journal/sensors
Sensors 2020,20, 2425 2 of 25
going x kilometers before going for service or jet engines doing x cycles. This signiﬁcantly reduces
downtime, as the metric is decided based on expert knowledge. The downside is that asset operators
do not know whether the maintenance is deployed efﬁciently [
]. If the engine could go longer
without maintenance a result would be poor resource management. To answer the when, a just-in-time
maintenance approach would be optimal .
1.1. The Role of Power Plants and the Need for Pumps
Combined heat and power plants, CHP, are an important asset in current energy infrastructures,
as they can provide power, grid stability, and heating simultaneously. With increasing amounts of
intermittent renewable capacity being installed both power grids and CHP operators are pressured.
Power grids face larger instability, while CHP operators are pressured by low electricity prices due to
cheap renewable energy [
]. Hence, power grid operators depend on the availability and quality of
the product provided by the CHPs, while CHP units strive to stay competitive. Though depending on
the energy market framework, the CHPs tend to be competing with each other and thus the plant with
higher efﬁciency or ﬂexibility tend to be activated. This means that asset managers at all time need to
know, which plants require maintenance to reduce downtime at high electricity price hours. If a critical
component fails at a bad time, it can cause severe damage and represent a signiﬁcant loss. Hence,
knowing when to conduct maintenance to reduce downtime is of great importance. CHPs larger
than 300 MW can have more than 100 pumps installed, varying in size and type, which speaks to the
complexity of the system [
]. A feedwater pump failing results in a large capacity reduction or full
shut-down depending on the system setup, which in turn presents an opportunity loss and safety
issue for asset operators. Avoiding this is desirable. Pumps are utilised in various ﬁelds and comes
in various forms, but tend to be critical components for operation. Examples are seen in seawage
industries and aviation [9,10].
1.2. The Field of Prognostics and Health Management
Continuous improvements in information and communication technology, ICT, have increased
data accessibility and in turn created a basis for Industry 4.0. This allows for increased automation
and data exchange, through increased deployment of Internet-of-Things, IoT, sensors [
]. The ﬁeld
dealing with these complex systems and beneﬁts from the development in technology is the study of
Prognostics and Health Management, PHM. PHM present tools to understand complex systems,
develop health indicators, and predict future complications. PHM has allowed for conducting
Condition-based Maintenance and Predictive Maintenance, PdM [1,2].
The prognostics within maintenance have been studied to a great degree within literature and
has been found to be gaining interest, as new papers are being released at a rapid pace. Figure 1
depicts the trend on three recognised databases considering PdM. This emphasizes the importance
of literature reviews. The focal point of literature reviews is to recap what has been investigated,
while also determining what challenges and unmet needs require examination.
Prior art conducting reviews within the ﬁeld of PHM can be found in Table 1. It was found that
literature reviews covered a great area of interests, such as algorithms, applications, decision-making,
implementation, etc. The thing they have in common is that they are scoped broadly to cover various
industries and ﬁelds. Thus, the purpose and novelty of this paper is the focus it provides on speciﬁc
topic. Carvalho et al. [
] presents a structured review of new literature presented in the years after
2014 and does so in a general manner, so the reader can get a general understanding of what moves
within the ﬁeld. Merkt et al. [
] studies various maintenance approaches and the different applications
that PdM might have. This is done by studying the full cycle of developing a successful model. Lee
et al. [
] outlines a framework for considering to what extent a company will be gaining value from
various PdM approaches. It more speciﬁcally dives into rotary components and their common failure
types, before presenting a set of case studies. This paper will provide an understanding of newly
published art within the PdM ﬁeld and then delve into PdM within pumping systems and CHPs. This
Sensors 2020,20, 2425 3 of 25
allows for a niche understanding of the differences between the general trends and the speciﬁc studied
ﬁeld. This approach can be applied to other ﬁelds or industries to evaluate gaps.
The published papers on predictive maintenance from 2000 to 2020 on MDPI, ScienceDirect
Table 1. An overview of a set of review papers in relation to the ﬁeld of PHM.
Reference Year Focus
 2019 Structured PdM literature review of 30 papers and their algorithms.
Applications for RF, ANN, SVM, k-means are presented.
 2019 A review of decision making algorithms and their applications.
 2019 Review and proposition of efﬁcient step-wise method for
companies to deploy models on historical data and digital twins.
 2019 Review of various maintenance strategies, their application
and use-cases with focus on the full cycle of creating PdM models.
 2018 An introduction to CDM with a strong overview over what
possibilities it provides.
A holistic overview of PHM. Present tools for companies
to consider to evaluate what degree of maintenance they need.
An overview of common fault types on rotary
equipment and a list of algorithms with their use-case.
This paper investigates the literature within the ﬁeld of PdM, creating a structured overview
of current applications. A deep dive into applications found on pumping systems will be provided,
before elaborating on how this is or could be applied to a CHP setting. The ﬁndings on pumping
systems can be extended to other ﬁelds. The paper will among other outline the current trends and
challenges. One of the focus points will be the scale-ability of the models as systems tend to rely on
more than one pumping system. The paper is structured as follows; Section 2introduces the concept
of PdM and puts it into context for the reader. Furthermore, shortly introduce how models can be
developed and what is required of the model owner. Following this, state-of-the-art applications will
Sensors 2020,20, 2425 4 of 25
be presented alongside the advantages and limitations of the algorithms considered. Section 3will
then delve into the applications found within pumping systems and CHPs, this is done by presenting
an initial overview of common faults within pumps. Section 4will be presenting the challenges that
PdM currently faces. Section 5will present the future trends found for PdM and what unmet needs
that require investigation. Section 6ﬁnalises the paper with a conclusion on the review.
2. Predictive Maintenance
2.1. Deﬁning Predictive Maintenance
To better grasp PdM, the authors have decided to initially deﬁne the concept. According to Oxford
Learner’s Dictionaries the adjective predictive is formally deﬁned as: “Connected with the ability to
show what will happen in the future” and maintenance is deﬁned as: “The act of keeping something
in good condition by checking or repairing it regularly” [
]. Hence, combining the two will allow
for the following deﬁnition of PdM; “the act and ability to keep something in a good condition by
understanding what happens in the future”.
The following deﬁnition of PdM is provided by Mobley [
]; “Predictive Maintenance is a
philosophy or attitude that, simply stated, uses the actual operating condition of plant equipment and
systems to optimize total plant operation. A comprehensive predictive maintenance management
program uses the most cost-effective tools (e.g., vibration monitoring, thermography, tribology) to
obtain the actual operating condition of critical plant systems and based on this actual data schedules
all maintenance activities on an as-needed basis.” Therefore, PdM strives to identify trends, anomalies,
degradation at an early stage, so that sufﬁcient counter measurements can be deployed. The term
PdM and CDM has found different use in various papers. Some deﬁne it to differ, while others use it
interchangeably. This paper will be using the later, as the ﬁeld of CDM has somewhat developed into a
PDM-CDM ﬁeld. The PDM-CDM considers a broader area of both fault detection, diagnostic, and
2.2. Methods of the Predictive Maintenance Field
Several methods of doing PdM exists, where each approach has a set of pros and cons. It is
commonplace to establish system models as they automate the prognostics and can continuously
monitor complex system effectively, furthermore, provide indicators of potential risks. Figure 2
presents an overview of common model types utilised. The physical model approach has in current
literature been referred to as a digital twin [
]. The physical laws and formulas governing the
system is utilised to establish a model that represents the machinery with great detail. The beneﬁt
of such models are the accuracy and knowledge of the system that it provides. White-box models
allow for the identiﬁcation of faults, as the full system is described. This type of modelling have
been found to have high accuracy. The weakness of digital twins is the time they consume before
being established [
]. Knowledge-based models enable the use of domain expert knowledge and
deploy it in models, this can be beneﬁcial as it can assist the prediction process [
models mainly apply machine learning, soft computing, and statistical theory to establish a model that
takes in historical operational or condition data. Though, in many cases not as accurate as physical
models, data-driven models can achieve high accuracy. Relative to the amount of time it takes to
establish a digital twin, a data-driven model can be developed fast. The weakness of data-driven
models is their need for a large amount of high-quality data [
]. If the data is faulty, it will be reﬂected
in the model and the accuracy will deteriorate. Pre-processing methods exist to prepare data and
up the quality. Physical and knowledge-based models are as an initial stance not considered to be
within the scope of the paper, but is still included to the extent that some papers propose a hybrid
approach. Hybrid models combine two approaches and hence utilises the strength of each method
to overcome the weaknesses of the other [
]. The focal point of this paper is within data-driven
models, but the authors saw the need to mention other applicable approaches and recommends the
Sensors 2020,20, 2425 5 of 25
interested reader to consider the following book by Lughofer and Sayed-Mouchaweh [
and Sayed-Mouchaweh [
] presents the concepts and applications of the various PdM approaches
and introduces a set of case studies.
Figure 2. A representation of model types that can be developed and utilised within the PdM ﬁeld.
2.3. Framework for Developing a Machine Learning Model
This section will give an overview of how literature tend to approach creating a data-driven model.
First, understanding what type of data you have at your disposition will allow for narrowing down
what type of model will be suitable and what information you can achieve. Figure 3gives an overview
of what type of approaches exists. Initially identify whether the data is labelled. If labelled, a broad set
of options are available within the realm of supervised methods. If unlabelled, the typical problem
type will be that of unsupervised, which allows for clustering. If a different type of information is
desired, then manual labelling is possible or it would require a change to the current management of
data, by creating a metric that identiﬁes conditional data of the monitored system.
An overview of various methods that can be chosen depending on what information is
desired and to what extent data is available.
It seems that a consensus on how a model should be build have somewhat been reached within
literature. Though, the order of the steps might be deﬁned or prioritised differently, the same
steps tend to be included. Figure 4depicts an overview of the common steps taken to develop
Sensors 2020,20, 2425 6 of 25
a machine learning model. Data Acquisition refers to the gathering of data. With the deployment
of communication technology, sensors have been a key source of data, sending and storing data
in databases. Data Preprocessing is the steps taken to prepare the data. A taxonomy, suggested by
], prepares six steps to consider and is displayed in Figure 5. The exact order depends on
the data being worked on. Choosing Model & Training considers the amount of available models and
the task of training it. In some literature, these are split into two steps, which the authors also consider
a possibility. The amount of models available has rapidly increased and gaining an overview can be a
challenge, thus ﬁnding an appropriate model requires testing. In the paper by
Lee et al. 
, a table over
a set of models and tools have been proposed alongside their corresponding strengths and weaknesses.
A paper by Fernandez-Delgado et al. [
] present an extensive study on various machine learning
techniques. Training can often be used as a basis for choosing a model, due to the training commonly
giving an indication of the model performance. Methods commonly deployed in literature to evaluate
training performance is the k-fold method and hold-out method. The k-fold method splits the data
such that a k
of the data is used to predict upon and the rest for training, this being iterated k times.
The hold-out method is often recommended for larger data sets. This method takes out a percentage
of the data to be predicted upon, where the rest is used for training. Model Evaluation considers the
performance of the model on an unseen test data set. This allows the developer to determine how
the model potentially will perform prior to being deployed. Analysing whether the model accuracy
found in training aligns with testing. Parameter Tuning allows for optimising the accuracy of a model.
Machine learning models operate with different parameters to allow for ﬂexibility and to increase
or reduce bias. The last two steps consider deploying and maintaining the model. Deploying the
system can relate to installation or integration of software, while it can also refer to soft values, such as
Change Management for the customer to adopt the new tool and use it for decision-making. Model
maintenance considers that systems often are dynamic, where a common assumption on a model is
that the system is static .
An overview of the process for developing a successful machine learning model, be it in a
PdM setting or another.
Sensors 2020,20, 2425 7 of 25
A suggested Taxonomy for Data Preprocessing. The exact order does not necessarily matter
and each step should be considered whether necessary for the application at hand.
Another common acknowledgment is that the individual case study requires a different approach.
Despite this, efforts have been invested in developing a structured approach to create and streamline
model construction. The authors recommend the interested reader to consider Lee et al. [
] for an
overview of what considerations should be done before determining whether predictive maintenance
is something that should be deployed, furthermore on how to approach it effectively. The above
discussed problem was popularised as the No Free Lunch Theorem by Wolpert and Macready [
optimisation. This refers to that no single algorithm is superior to the other, as they each serve a
purpose and is case dependent.
2.4. Overview of Methods
This section will be introducing relevant literature on successful applications of predictive
maintenance. The concept of each highlighted ML method will be introduced shortly before relevant
literature will be presented. Machine learning methods tend to be grouped by their “way of
functioning”, e.g., tree-based models, such as decision trees, or neural networks with their various
architectures. The authors have tried to take this grouping-by-similarity approach, but the authors are
aware that some might fall into two categories or not fall into any category. An overview of relevant
literature categorised can be seen in Table 2. The reader should be aware that papers can have applied
several types of methods for either testing or comparison, this allows for the paper to be mentioned
several times within the list. Listing the papers show a fair distribution of papers across all categories,
but a preference towards research within ANN & DL can be seen as it has been increasing in popularity.
Furthermore, it is found that a signiﬁcant amount of papers are presented in a experimental setting and
that approximately 30% of the studied literature utilised vibration data to determine the health status
of a system. Though, the majority of prior art is presented in a experimental setting, the applications
does span over a large area, which shows the ﬂexibility of the PdM ﬁeld.
2.4.1. Stochastic Algorithms
Stochastic data-driven models within prognostics are often considered within the Bayesian
category. Rather than giving a single estimated output on the current system health, it gives a
probability distribution of possible likely options [
]. In this way, the Bayesian method can present
the current state of the system, but can also evaluate future trends before a given threshold. Stochastic
algorithms are mainly used within degradation models and the most common Bayesian network
algorithms are Particle Filters, Kalman Filters, and hidden Markov models [29,30].
Sensors 2020,20, 2425 8 of 25
A list of studied papers categorised by the characteristics of the algorithm applied. The list is an effort on displaying what is moving within the ﬁeld, but does
not reﬂect all available algorithms.
Category Algorithm Area of Application Type of Machinery Data Reference
ANN & DL MLP Experimental Motor Vibration 
Aeropropulsion system & truck engine Nasa & Scania Trucks 
Battery Capacity 
Transportation Rail Oil-level, voltage, pressure, temperature 
ELM Wind Turbines Gear Vibration 
KELM Power Production Hydropower generator Vibration 
SW-ELM Experimental - PHM Challenge 2008 
CNN Experimental Aeropropulsion system & truck engine Nasa & Scania trucks 
Buildings Fans Electrical, mechanical, temporal values 
Milling Cutting machinery Force, feed rate, speed 
(LSTM)-RNN Experimental Aeropropulsion system & truck engine Nasa & Scania trucks 
Turbofan engine Nasa [40,41]
Motor bearings Vibration 
Battery Capacity 
TDNN Experimental PEMFC Voltage 
SFAM-ANN Rotational mechanical assets Bearings Vibration 
Stochastic DLM w. BN Transportation Aircraft aircondition Airplane Condition Monitoring System (temperature) 
Markov Experimental - - 
Transportation Rail - 
KF Experimental Battery - 
Voltage, current, capacity 
SUKF Experimental Bearings Vibration 
EKF Experimental Battery - 
PF Experimental Battery - 
Voltage, current, capacity 
DBF Stainless steel industry Hot rolling process (drums) Density, temperature, pressure, power, force 
Sensors 2020,20, 2425 9 of 25
Table 2. Cont.
Category Algorithm Area of Application Type of Machinery Data Reference
Statistical LR Experimental Industrial radial fans Vibration, rotational speed, temperature, pressure 
Milling Cutting machinery Vibration 
RFR Experimental Industrial radial fans Vibration, rotational speed, temperature, pressure 
Milling Cutting machinery Vibration 
SR Experimental Industrial radial fans Vibration, rotational speed, temperature, pressure 
ARMA Experimental PEMFC Voltage 
ARIMA Experimental Battery - 
RVM Experimental Battery - 
SVR Experimental Battery Capacity 
SVM Experimental Bearings Vibration 
LS-SVM Reﬁnery Destillation process - 
LS-SVM NAR Food industry Vertical form ﬁll and seal machine Vibration, thermal imaging 
Fault Detection & SVM Agro-industry Harvester Vibration 
Fault Classiﬁcation DT Transport Rail ERP system data, conditional data 
RF Transport Rail ERP system data, conditional data 
Milling Cutting machinery Speed, feed rate, depth 
RBF-SVM ion-Implanter tool Filament Current 
Sensors 2020,20, 2425 10 of 25
Degradation can be difﬁcult to determine, especially, when the progress status is hidden.
This means that a degradation can to some extent seem linear, but in reality contain non-linear
elements. Furthermore, data can be noisy to the extent that it greatly inﬂuences the predicted RUL and
hence reduces accuracy. Son et al. [
] delves into this issue. A method is proposed to utilise condition
monitored data to predict the RUL with an acceptable error. The RUL is estimated by deploying a
model based on a constrained KF model. To do so, a set of inequality restraints are set up to achieve
the desired accuracy. To verify the method, it is tested on a case study of an automotive lead-acid
battery. Qiu et al. [
] also studies a battery, but looks at both state of charge, state of health, and RUL.
To improve the accuracy of the prediction, BS-SRCKF is deployed for the state of charge, which in turn
is combined with MHKF and an EKF to do a joint estimate of the state of charge and state of health.
The state of health values are then utilised in a particle ﬁlter that relies on an improved cuckoo search
to determine the RUL.
In an alternative approach, Xue et al. [
] deploys an AUKF alongside genetic algorithm that
optimises the parameters for a SVR model. The AUKF serves the purposes of calculating the process
noise covariance and observation noise covariance, this is done continuously over the time series. The
method is veriﬁed on a battery data set provided by NASA.
In the study proposed by Ruiz-Sarmiento et al. [
], machinery degradation in the hot rolling
process is studied. This is done by applying a Bayesian Filter. This utilises expert knowledge along
side historical data to increase the accuracy of the predictions. The ﬁlter deployed is called a Discrete
Bayes Filter. The ﬁndings of the paper result that this hybrid model relying on expert knowledge and
machine learning theory increases the overall accuracy than they could individually.
2.4.2. Statistical Algorithms
Statistical data-driven models within prognostics are relatively simple trend extrapolation.
They rely on condition monitoring data to create a trend curve from a single dimension time series that
then reﬂects the degradation of health for the asset. Where stochastic algorithms give a probability
distribution as an output, the statistical algorithms give a single speciﬁed output. A common algorithm
is ARIMA, but it is also possible to apply regression techniques .
SVM sees its use in both classiﬁcation and regression problems. In the paper by
Yan et al. 
RUL of bearings is estimated based on a classiﬁcation. This is done by splitting the degradation into 5
categories. The method requires a dimensionless input. This is achieved by calculating the RMS for the
vibration data. To evaluate the model, data from a public data set is utilised. The data set considered
are from IMS and PRONOSTIA. The model performs equally to better than other models proposed to
In a paper by Susto et al. [
], SVM is utilised to classify whether the machinery is in need of
maintenance. The machinery in consideration is that of an ion-implanter tool. The base case is the
current PvM scheme. The study ﬁnds that the PdM approach can perform equally to better than
the presented PvM scheme. In an approach to detect faults early, LS-SVM Regression is applied to
a Vertical Form Fill and Seal. The paper by Langone et al. [
] initially applies an unsupervised
clustering algorithm, KSC, to initially identify anomalies within the data set, before applying the
LS-SVM regression for prognostics. The model successfully detects when the dirt is building up within
In the study by Li et al. [
], LS-SVM is applied in an alternative manner to a distillation process.
The data set is based on a moving-window and hence applies just-in-time maintenance. Here the work
focuses on creating an online soft sensor tool that allows for monitoring the distillation process. A
comparison between applying LS-SVM as a static and dynamic version is done for presenting the
results. The static LS-SVM arrives at a
of 0.995 with a RSME 1.7991, where the dynamic version,
with the largest moving window, presents a R2of 0.999 and a RMSE of 0.062.
An alternative approach proposed by Zhou et al. [
], considers algorithms across the categories
of statistical and ANN. The RUL prediction is conducted on a PEM fuel cell. The prediction is divided
Sensors 2020,20, 2425 11 of 25
into three phases. The ﬁrst phase deals with non-stationary trends by deploying PAM. The second
phase, the order of an ARMA model is estimated and deployed to ﬁlter the linear degradation elements
in the data. The ﬁnal phase utilised the leftover non-linear elements to train a TDNN. The paper ﬁnds
the model to predict RUL with high conﬁdence.
2.4.3. Artiﬁcial Neural Network and Deep Neural Network
The idea of ANN stems from the functionality of the human brain. The nervous system is the
driver behind everything that a human does and the nervous system consists of neurons. An ANN
replicates this by having a set of layers, where each layer consists of neurons. Each layer is connected
and can receive and transmit signals, corresponding to the synapse within the nervous system.
Each connector has a weight, the neurons have an activation function and a threshold value. These are
the parameters considered within ANN. As technology has developed and ANN has taken on
increasingly complex tasks, a new term has been given to more complex ANN, which is called
Deep Neural Network. The most distinct difference between the two terms is the complexity, referring
to the architecture of the networks and the number of layers that it may contain [22,63,64].
In a paper presented by Silva and Capretz [
], two fans are studied by applying CNN. CNN is
typically known for being applied to image recognition, but is in an alternative approach applied to
fault detection and predictive maintenance. The features used are divided into electrical, mechanical,
and temporal. Initially, features are one-dimensional, hence a data transformation is required. The
paper is inspired by Wang and Oates [
] and Chen et al. [
] to try GAF and Moving Average Mapping,
respectively, and converts the data into two-dimensions. The intent is to allow for better knowledge on
the relationship between features. CNN is applied and ﬁnds an accuracy of 95% and 98% for the two
fans. The proposed method reports that it outperforms traditional methods such as RF, SVM, and MLP.
In a study proposed by Markiewicz [
], predictive maintenance of induction motors is considered.
The paper touches upon the issues of data gathering and storage, as it is an energy intensive process.
Hence, Markiewicz [
] presents a solution where predictive maintenance is done locally on a set of
ultra-low power wireless sensors. This is possible with a reduced computational complexity from
applying a compressed RNN. For their cell, LSTM is applied as they argue that it does not suffer from
a vanishing or exploding gradient problem [
]. By applying a compressed RNN algorithm they found
an accuracy of approximately 92% and that this was comparable or at a better performance than other
algorithms e.g., KNN. This approach moreover beneﬁted from the lowered energy consumption.
Zhang et al. [
] investigates a degradation model on bearing performance by applying LSTM
RNN. The model builds on vibration data, but also introduces an alternative indicator, namely
“waveform entropy”. Entropy can be for a discrete instance of random vector X be deﬁned as follows:
is the entropy,
is the number of elements in X,
is a single element of X, and
the probability mass function. The paper sets out to develop a dimensionless indicator. This is due to
e.g., load and speed not only resembling the health of a system, but also reﬂects the current operating
condition. Inspired by Ali et al. , a waveform entropy can be calculated:
Here WFE is the waveform entropy,
is the waveform factor at time instance
length of the sliding window. This method beneﬁt from not requiring signal decomposition, such as
other time-frequent domain features. WFE is a local mean of logarithmic vibration energy. Hence,
a strong representation of the state of the vibration levels. The paper presents three steps for predictive
Sensors 2020,20, 2425 12 of 25
maintenance. Fault detection, fault classiﬁcation, and a degradation model. The WFE supports this
and gives great results.
Clustering represents a category of unsupervised algorithms, which objective is to ﬁnd clusters
within a data set. Algorithms typically rely on a centroid or hierarchical approach to determine the
clustering of data that reﬂects the shortest distance internally and the largest distance between clusters.
Examples of algorithms are k-means, Expectation Maximisation, Hierarchical Clustering.
As clustering is unsupervised, the main purpose of clustering algorithms is to ﬁnd golden nuggets
within the data. In a PdM setting, the main goal of clustering becomes to detect failure and anomalies
within data. In a paper by Uhlmann et al. [
], k-means is deployed in a supervised setting to see the
performance of detecting faults on a Selective Laser Melting machine tool.
Another approach is using clustering alongside other algorithms. This is done in the work by Cao
et al. [
]. Here fuzzy clustering is deployed to determine the severity of a failure, before deploying
semantic technologies to investigate the time of the failure. Similarly,
Daher et al. 
, applies fuzzy
C-means alongside ANFIS to detect the RUL of a distillation column.
2.5. Advantages and Limitations
In this section, some of the strengths and weaknesses of the investigated algorithms will be
elaborated upon. Considering that the various types differ greatly, each has its purpose and application.
Table 3lists some general advantages and limitations of the three categories. This is not a complete
list, as the individual architecture within ANN & DL and algorithm within regression present a set of
strengths and weaknesses. This is considered out of scope for this paper. For a look into benchmarking
of ML, the reader is recommended to visit Olson et al. [
Fernandez-Delgado et al. 
present a comparison of 179 machine learning algorithms on various data sets. Where Ahmed et
] looks at an empirical comparison of time series forecasting within the machine learning domain.
This paper will proceed to discuss some of the speciﬁc advantages and limitations found within the
literature of the various case studies.
An overview of some of the general advantages and limitations of the algorithms within the
given categories [
]. As individual algorithms present their own set of strengths and weaknesses
the reader is recommended to study the individual algorithm as well.
Category Advantage Limitation
ANN Great w. large dataset Black box model
Handles noisy data Requires a large dataset
Limited/no need for pre-processing Computational expensive
Deals w. nonlinear & complex tasks
Stochastic Can deal w. smaller datasets Require quality pre-processing
Result w. probability distribution Require accurate degradation modelling
Can operate alongside statistical approach Can struggle w. multidimensional data
Prior knowledge can be incrementally introduced
w. new data to achieve better performance
Can deal w. nonlinear tasks
Statistical Single precise RUL estimate Early prediction tend to be inaccurate
Various types of regression models Utilise a single dimension for prognostic
ANN solutions have been gaining speed as new architectures have been discovered [
ANN have enabled machine learning to complete more complex tasks that were not previously
possible or were difﬁcult by nature. The downside is that it is a black box, meaning low interpretability.
Sensors 2020,20, 2425 13 of 25
Sampaio et al. [
] found that the MLP architecture worked effectively on non-linear and complex
systems, furthermore it had great generalisation. On the downside, the convergence was found to
be somewhat slow and that the model had a tendency to overﬁt. Finally, when the proposed model
projected a RUL value near the end of the lifetime, it would slowly move further away.
Zhou et al. [
] claimed that the use of ANN provided great calculation speed. More speciﬁcally,
the KELM algorithm performed better than other algorithms in generalisation, due to it being able to
ﬁnd a least-square optimal solution. Another advantage being that it can achieve multiple outputs.
Though one weakness being that it is heavily dependent on the initial parameters, so if estimated
poorly it will be reﬂected in the model. Similarly, Javed et al. [
] ﬁnds that ELM has a fast iterative
tuning process for deciding on hidden layer parameters. This is due to the algorithm doing it in a
single-step and does not need human intervention. The downside is the random parameters could be
Markiewicz et al. [
] ﬁnds LSTM-RNN to beneﬁt from being relative stable, notably because it
does not suffer from the vanishing or exploding gradient problem [
Zhang et al. 
LSTM-RNN has good performance on sequential data due to the recurrent feedback. Finally, as argued
by Nguyen and Medjaher [
], LSTM-RNN beneﬁts from having a long-term memory meaning it can
keep important information for later application. This can be especially beneﬁcial for degradation
prognostics. Ruiz-Sarmiento et al. [
] decided on utilising a Bayesian approach by applying DAE, this
was due to the algorithm being robust against noisy and ﬂuctuating data. They further stated that the
reason for not utilising a KF was due to the linearity assumption within the underlying system of KF.
Several papers applied either SVM or SVR algorithms. Here, Li et al. [
] stated that LS-SVM had
a good performance on non-linear regression, especially in generalisation. Ruiz-Gonzalez [
a SVM classiﬁer due to the strong generalisation, furthermore, with a small data set the chances of
overﬁtting with SVM were relatively low, as well as having a good computation time. Langone et
] found that LS-SVM beneﬁtted from being able to achieve a global optimum due to least-square
presenting itself as convex. Zhou et al. [
] on the other hand states that SVR has a long computational
with larger data sets and constrained optimisation issues.
The paper has till now studied various state-of-the-art ML approaches on doing PdM and applying
algorithms. This general overview is needed when further delving into a speciﬁc topic to identify
challenges and unmet needs. The focus of this paper lies within pumping systems and with the
understanding of the current usage of predictive maintenance, the paper will move further towards
what has been accomplished within the ﬁeld of predictive maintenance on pumps. According to United
States International Trade Commission in 2005 [
], 80% of the pumps produced in the world are
centrifugal pumps and Karassik and McGuire  states that 90% of the pumps used in the chemical
industry are centrifugal pumps. Hence, this paper will primarily focus on applications on centrifugal
pumps. To ﬁnalise this section, understanding what has been conducted on PdM within a power plant
setting will also be elaborated upon. As power plants heavily rely on their pumps functioning. Notice
that power plants can potentially be using both centrifugal pumps and positive displacement pumps.
An article released by Collins and Davis on PowerEngineering [
], gives a great overview of things
to consider before investing in one type of pump and states that a power plant of 300 MW will on
average have approximately 100 pumps installed.
Sensors 2020,20, 2425 14 of 25
3.1. Common Faults in Pumping Systems
A centrifugal pump consists of several rotating and static components that has the task of
displacing liquids. Due to the interaction of moving parts and liquids interacting with a solid and static
surface, several errors can surface. A thorough investigation by Forbes [
] resulted in a conclusion
of dividing the main issues of a centrifugal pump into 13 types of faults. Each of the 13 types of
faults could then further be divided into more speciﬁc reasons for the fault happening. Table 4
presents an overview of detectable errors with their corresponding data tags that might be of interest.
To get a more detailed description of speciﬁc issues, the reader is recommended to visit Forbes [
The main components of a simple centrifugal pump will consist of an impeller, a suction nozzle, a
discharge nozzle, a shaft, a mechanical seal, and several bearings. A more detailed overview can
be seen in Figure 6. Due to the nature of the pump, the pump faults can be categorised into three
categories. Hydraulic, mechanical, and other failures. The hydraulic failures are caused by the liquid
ﬂowing within the piping. The state of the liquid can inﬂuence the performance of the pump, but also
damage parts of the pump. Mechanical failures refer to the interaction of moving and static parts, this
commonly causes wear or other types of fatigue. Finally, other type failures are a category for failures
that do not directly fall into the two other groups. The power consumption of a pump can often be an
indicator of whether the pump is performing desirably.
Table 4. Failure types by category for a centrifugal pump.
Name of Fault Category Relevant Data Tag Description
Cavitation Hydraulic Failures Vibration, pump efﬁciency,
noise, pressure, ﬂow rate
Formation of vapour bubbles that collapses and damages
the piping system. Can result in fatigue or erosion.
Pressure Pulsation Hydraulic Failures Vibration, pressure Can come from running frequencies of the pump,
be ressonance of the system, acoustic behaviour, etc.
Radial Thrust Hydraulic Failures Temperature Thrust directed towards the center of the
pump rotor. Typically occurs at low ﬂow rates.
Axial Thrust Hydraulic Failures Temperature Thrust imposed on the shaft in either an inboard or
outboard direction. Common consequence is fatigue failure.
Suction and Discharge
Recirculation Hydraulic Failures Pressure, noise
An unavoidable fault is the recirculation of some
water within the impeller, either in suction
or discharge of the impeller. Can often be determined by
observing pressure pulsations at inlet and outlet.
Bearing Failure Mechanical Failure Vibration, temperature,
(stress waves/shock pulses)
Can be due to various reasons, such as
contamination of bearing oil by water, other liquid,
high heat, introduction of solid particles.
Seal Failure Mechanical Failure Temperature
Opening of the lapped faces results in solids
entering, or solids sticks to the surface and
introduces severe wear on the hard face.
Lubrication Failure Mechanical Failure Temperature Excessive heat reduces the lubricating
ability and lifetime of the oil.
Excessive Vibrations Mechanical Failure Vibration
Stems from unbalanced moving parts, particles of
the liquid interacts with the pumping system. A large
source of errors should be evaluated to do identiﬁcation.
Consumption Other Failure Voltage, current,
A typical indication of the pumping system
having a failure somewhere in the system and
can have different sources for the error.
Blockage Other Failure Flow rate Clogging of piping system or the impeller
can result in the pump stopping to function.
Sensors 2020,20, 2425 15 of 25
A simple ﬁgure of a centrifugal pump. From left to right; water enters the impeller through
the suction nozzle, where kinetic energy will be applied to the liquid through turning of the shaft.
The chasing keeps the water within the system, while the mechanical seals make sure there is no
leakage. The bearing reduces the friction between the moving and stationary parts and is found in
several places within a pumping system. The water exits at the discharge nozzle.
3.2. Pumps and Thermal Power Plants
Pumps have been around for a long time and seen their application within various industries.
Displacing liquid has always been a need for humans to easily transport e.g., water and it is a key
part of the modern infrastructure. An example of a common pumping system is seen in CHPs.
District heating allows for increased efﬁciency of power plants by utilising excess heat and deliver
heat to costumers. When a critical pump breaks down, it inﬂuences the performance of the CHP and
hence maintenance becomes of importance to minimise failure and capacity reduction. Table 5lists
algorithms applied to pumps and CHPs found in the literature. It furthermore presents the objective
of applying the algorithm and what data was used. Literature on pumps are more commonplace
as they serve a general purpose, but it is in most literature limited to fault detection and diagnostic.
The literature presented on pumps is not limited to centrifugal pumps as they play a signiﬁcant role
in various systems. Utilising prior art will assist other industries relying on pumps to have an easier
time adopt similar systems. The literature found within the CHP domain was very limited, as most
of the power plant related material was referring to nuclear power plants, NPP [
]. The authors
recognise that the list is not complete and that the list would beneﬁt from a more exhaustive search,
but it is representative of current applications.
Sensors 2020,20, 2425 16 of 25
An overview of relevant literature for the domain of PdM within all types of pumps and CHP
units. It is not an exhaustive list, but it can give insight into where more efforts should be focused.
Application Algorithm Objective Data Type Reference
Pump GNB, SVM, RF, MLP, KNN Detect cavitation Vibration 
SVM Detect cavitation & blockage Vibration 
GMM Clustering Detect operation modes Vibration 
SOM NN Fault detection Vibration 
Polynomial Regression Fault detection & diagnostic Temperature, Pressure,
Mass Flow, Current 
PF Estimate RUL Vibration 
AO-PF Estimate RUL oil ﬂow 
LR, AHSM Estimate RUL Flow, vibration 
KF Estimate RUL Vibration 
LN, MLP, DAE Estimate RUL Flow, pressure, stress 
CHP LR, MLP, SVR, RF Performance estimation
Exhaust Vacuum, Full
EE, Autoencoders, IF Anomaly detection
Exhaust Vacuum, Full
FL, LR RUL estimate of turbine - 
The difﬁculty of creating models on pumps and CHPs stems from their ﬂexible nature. NPPs
will primarily be operating at high to full capacity at all times. If a NPP is not doing so, it becomes a
bad business case for the asset owner. The value of a CHP is, on the other hand, the ﬂexibility that it
offers. This is the same for pumps. Depending on the pump it can either freely regulate up and down,
or it can have set modes to operate in. Wang et al. [
] presents a model, which initially clusters the
operation modes of a centrifugal pump by applying GMM Clustering before estimating a RUL with a
PF. Such methods allow for easier fault detection and also eases the process of creating a RUL.
As CHPs are composed of various components, it becomes tricky to create just a single model for
the full CHP, but requires a model on component level. Hundi and Shahsavari [
] challenges this, by
creating a model which estimates the performance depending on various conditional and time-series
data, by applying various regression methods. This later allows for anomaly detection, by presenting a
The beneﬁt of CHPs being build of various components is that literature from other areas might
be applicable. e.g., bearings are commonplace within rotary machinery and state-of-the-art algorithms
have been applied to this area [
]. Another method of overcoming data reﬂecting a ﬂexible
operation is presented by Moleda et al. [
]. Moleda et al. [
] investigates a feedwater pump on a
CHP in Poland. An approach utilising a bag of regression models is suggested. This is done by creating
a regression model for each data tag and then calculate the RMSE between the estimated value of the
model and the actual value. Though it does not give a RUL estimate, this method allows for detecting
drift, anomaly, and outliers.
A trend found within the data sets utilised is that many rely on vibration
Vibrations are common within pumps, but also generally applied. This is due to vibrations giving a good
indication of the current state of health for the given system. Furthermore, given a little pre-processing
the degradation curve becomes fairly detectable. In a paper by Tse et al. [
], a slurry pump is studied.
Multiple vibration sensor data is extracted and fusioned. Following this a KF is applied to determine the
degradation. The proposed method is not reliant on run-to-failure data and can predict the RUL even
with one channel failing to provide data.
Sensors 2020,20, 2425 17 of 25
4. Challenges for Predictive Maintenance Applications
A set of challenges were identiﬁed, during the literature review. This section will outline some
general challenges that PdM faces currently and then some domain-speciﬁc challenges concerning
pumps and CHP units.
In Tabel 2a set of application areas was deﬁned. A large proportion was identiﬁed as being
]. Meaning that a component was investigated in an experimental
setting and hence not necessarily attached to a manufacturer or plant. This shows that even though
PdM has come a long way, there still exists a gap between experiments and industrial applications.
The ﬁeld of PdM would beneﬁt from applying algorithms to historic data from industrial processes [
This is reasoned that experimental data sets tend to be more ﬁne-tuned, whereas industrial historic
data tend to be noisy and have a larger amount of missing data. A sensor might not have been installed
for the desired data tag. As seen in Section 3.2, most pump models were developed on vibration data.
If vibration data is not available other methods have to be deployed. Furthermore, a component can
have multiple sensors as a source for a model, which requires the developer to align and pre-process
]. Furthermore, industrial data sets, typically, do not have a labelling metric, hence it
requires the developer to dive into expert knowledge or company ERP systems to identify faulty data,
which in turn can introduce bias. Here, some models present themselves more lucrative, as they do
not require a large amount of pre-processing or small amounts of domain knowledge.
Data-driven models tend to make certain assumptions on the system that it reﬂects. One common
assumption is that it is static and does not develop over time [
]. This will in many cases be
proven to be wrong with many systems being dynamic. Especially within the ﬁeld of PdM, as a
component is replaced with another, it will require a certain maintenance for each model developed.
Several strategies exist to develop models, but currently most of the models are developed ofﬂine,
before implementation and does not consider maintenance. Here, some methods have been deployed
to look at a certain frame or period, e.g., the last year, thus a continuously moving frame. With the
deployment of more IoT equipment, new options open up. Some studies within the ﬁeld of automatic
maintenance and continuous improvement have been completed .
Another challenge lies in the task of determining an accurate RUL. Depending on the data
available different models and methods exist to tackle the problem [
]. In many cases, there is not
necessarily run-to-failure data available, but rather a degradation path or rather depend on expert
knowledge. The issue then becomes to estimate the RUL from the available data. Some components
can have a steady linear degradation at the start of their lifetime, but then suddenly start dropping by
the end of it. This is common within e.g., batteries [
]. Hence, an issue of determining a hidden
In regards to the pump and CHP domain, a challenge was identiﬁed concerning the operation.
Most experimental data can be measured at a ﬁxed operation, but as CHPs and pumps have
variable operation [
], errors may be hidden within the mass of data. Here techniques need to
be extended to handle such issues. Another problem is the scale-ability and streamlining of models.
As industrial process might have several similar or identical components, it is important to investigate
the scale-ability of developed models for them to be used for plug-and-play. This should ease the
implementation of PdM into the industrial sector. This is challenged by the No free lunch theorem, as a
component at one plant might differ slightly from another and then, in turn, prove that another model
might be more suitable.
5. Future Trends
Following Figure 1, the trend within the ﬁeld of PdM seems obvious. An increasing interest
for research seems to take place. This follows the increase in deployment of ICT enabling data and
knowledge sharing. Some companies are already dealing with prior mentioned challenges, such as
streamlining and scale-ability of algorithms. e.g., Oracle presents the MSET2 tool that can be deployed
for process monitoring [
]. Another company stated that its monitoring tool could signiﬁcantly reduce
Sensors 2020,20, 2425 18 of 25
maintenance costs [
]. These services will be increasingly common. As tools for easily development
and implementation become more common, industrial applications will follow naturally, as it becomes
commonplace to utilise the state of the art of PdM to stay competitive.
Another trend that Industry 4.0 brings, is the possibility for automatic maintenance of models [
Models will be self-sustainable even with new components being installed, though it will be presenting
new issues to deal with as well. Following Moore’s Law, a doubling in computational power can be
expected to happen every second year [
]. This will continuously increase the possibilities to solve
increasingly complex tasks and open up for various new architectures and algorithms from the ML
This paper did not extend to physical and knowledge-based models, only to the point that it
was introduced through hybrid models. Though, it seems that increasing amounts of literature is
considering combining either data-driven models with physical or knowledge-based models, as this
can improve accuracy and trust in the developed model [
]. This can be reasoned that
data-driven to a certain degree lack interpretability, where e.g., knowledge-based models would allow
for domain experts to inﬂuence the model [
]. This might be ideal for the industry moving forward,
as it would present them with a larger toolkit to operate within.
This paper presented a literature review on current state-of-the-art applications within the ﬁeld of
predictive maintenance. The paper initially presented a thorough introduction to what PdM offers and
gave a framework on how to prepare and develop PdM models. Here literature was recommended for
further reading to deepen the understanding of how model construction could be done. Then literature
with applications in various areas were introduced, this allowed for the identiﬁcation of advantages
and limitations of certain algorithms. It further allowed for identifying certain trends within the
predictive maintenance ﬁeld. This was compared to the speciﬁc domain of pumps and combined heat
and power plants. Here certain gaps were identiﬁed, as limited material could be identiﬁed within the
combined heat and power area. A set of challenges for predictive maintenance were then presented
before outlining future trends.
For future work, the authors recommend that a more exhaustive literature review is conducted as
there are increasing amounts of literature being published at a rapid pace. General trends could be
identiﬁed from the studied literature, but the paper would beneﬁt from veriﬁcation of the exhaustive
search. This is especially related to domain-speciﬁc areas, as general literature does not seem to be
lacking. Furthermore, the literature review should be extended to include literature of other types of
models, such as physical and knowledge-based models.
Conceptualization, J.F.O. and H.R.S.; methodology, J.F.O.; Formal analysis, J.F.O.;
investigation, J.F.O.; data curation, J.F.O.; writing—original draft preparation, J.F.O.; writing, J.F.O.; review
and editing, H.R.S.; visualization, J.F.O.; supervision, H.R.S.; All authors have read and agreed to the published
version of the manuscript.
This paper was written in connection with a project on Predictive Maintenance within
thermal power plants. The authors would like to express their gratitude to Ørsted for the collaboration.
Conﬂicts of Interest: The authors declare no conﬂict of interest.
The following abbreviations are used in this manuscript:
ML Machine Learning
RUL Remaining Useful lifetime
ICT Information and Communication Technology
PHM Prognostic and Health Management
Sensors 2020,20, 2425 19 of 25
PdM Predictive Maintenance
PvM Preventive Maintenance
CDM Condition-based Maintenance
CHP Combined heat and power plant
NPP Nuclear Power Plant
ANN Artiﬁcial Neural Network
DL Deep Learning
MLP Multilayer Perceptron
ELM Extreme Learning Machine
KELM Kernel Extreme Learning Machine
SW-ELM Summation Wavelet Extreme Learning Machine
CNN Convolutional Neural Network
LSTM-RNN Long-Short Term Memory Recurrent Neural Network
TDNN Time Delay Neural Network
SFAM Simpliﬁed Fuzzy Adaptive Resonance Theory Map (Neural Network)
DLM Dynamic Linear Model
BN Bayesian Network
KF Kalman Filter
SUKF Switching Unscented Kalman Filter
EKF Extended Kalman Filter
PF Particle Filter
DBF Discrete Bayes Filter
LR Linear Regression
RFR Random Forest Regression
SR Symbolic Regression
ARMA Autoregressive Moving Average
ARIMA Autoregressive Integrated Moving Average
RVM Relevance Vector Machine
SVR Support-vector Regression
SVM Support-vector Machine
LS-SVM Least-squares Support-vector Machine
LS-SVM NAR Least-squares Support-vector Machine Nonlinear Autoregressive
DT Decision Tree
RF Random Forest
RBF-SVM Radial Basis Support-vector Machine
BS-SRCKF Backward Smoothing Square Root Cubature Kalman Filter
MHKF Multiscale Hybrid Kalman Filter
AUKF Adaptive Unscented Kalman Filter
RMS Root Mean Square
KSC Kernel Spectral Clustering
RSME Root Mean Square Error
R2Coefﬁcient of Determination
PEMFC Proton-Exchange Membrane Fuel Cell
GAF Grammian Angular Field
KNN k-Nearest Neighbour
ANFIS Adaptive Neuro-Fuzzy Inference System
GNB Gaussian Naive Bayes
GMM Clustering Gaussian Mixture Model
SOM NN Self Organizing Map Neural Network
Sensors 2020,20, 2425 20 of 25
AO-PF Adaptive-Order Particle Filter
AHSMM Adaptive Hidden Semi-Markov Model
LN Ladder Network
DAE Denoising Autoencoder
EE Elliptic Envelope
IF Isolation Forest
ERP System Enterprise Resource Planning System
MSET2 Multivariate State Estimation Technique
Uckun, S.; Goebel, K.; Lucas, P.J. Standardizing research methods for prognostics. In Proceedings of the
2008 International Conference on Prognostics and Health Management, Denver, CO, USA, 6–9 October 2008;
pp. 1–10. doi:10.1109/PHM.2008.4711437.
Lee, J.; Wu, F.; Zhao, W.; Ghaffari, M.; Liao, L.; Siegel, D. Prognostics and health management design for
rotary machinery systems—Reviews, methodology and applications. Mech. Syst. Signal Process.
42, 314–334. doi:10.1016/j.ymssp.2013.06.004.
Mobley, R.K. Role of Maintenance Organization. In An Introduction to Predictive Maintenance; Elsevier:
Amsterdam, The Netherlands, 2002; pp. 43–59. doi:10.1016/b978-075067531-4/50003-8.
Merkt, O. On the Use of Predictive Models for Improving the Quality of Industrial Maintenance: An
Analytical Literature Review of Maintenance Strategies. In Proceedings of the 2019 Federated Conference on
Computer Science and Information Systems (FedCSIS), Leipzig, Germany, 1–4 September 2019; pp. 693–704.
Phogat, S.; Gupta, A.K. Expected maintenance waste reduction beneﬁts after implementation of Just
in Time (JIT) philosophy in maintenance (a statistical analysis). J. Qual. Maint. Eng.
Lund, R.; Mathiesen, B.V. Large combined heat and power plants in sustainable energy systems. Appl. Energy
2015,142, 389–395. doi:10.1016/j.apenergy.2015.01.013.
Wang, J.; You, S.; Zong, Y.; Træholt, C.; Zhou, Y.; Mu, S. Optimal dispatch of combined heat and power plant
in integrated energy system: A state of the art review and case study of Copenhagen. Energy Procedia 2019,
158, 2794–2799. doi:10.1016/j.egypro.2019.02.040.
Collins, D.; Davis, J. Things Power Plant Engineers Need to Know About Pumps. Available online:
pumps/#gref (accessed on 23 April 2020).
Forbes, G. A review of major centrifugal pump failure modes with application to the water supply and
sewerage industries. In ICOMS Asset Management Conference Proceedings; Asset Management Council:
Oakleigh, Australia, 2011.
Li, T.; Wang, S.; Shi, J.; Ma, Z. An adaptive-order particle ﬁlter for remaining useful life prediction of aviation
piston pumps. Chin. J. Aeronaut. 2018,31, 941–948. doi:10.1016/j.cja.2017.09.002.
Isaksson, A.J.; Harjunkoski, I.; Sand, G. The impact of digitalization on the future of control and operations.
Comput. Chem. Eng. 2018,114, 122–129. doi:10.1016/j.compchemeng.2017.10.037.
Permin, E.; Lindner, F.; Kostyszyn, K.; Grunert, D.; Lossie, K.; Schmitt, R.; Plutz, M. Smart Devices in
Production System Maintenance. In Predictive Maintenance in Dynamic Systems: Advanced Methods, Decision
Support Tools and Real-World Applications; Lughofer, E., Sayed-Mouchaweh, M., Eds.; Springer International
Publishing: Cham, Switzerland, 2019; pp. 25–51. doi:10.1007/978-3-030-05645-2_2.
Carvalho, T.P.; Soares, F.A.; Vita, R.; da P. Francisco, R.; Basto, J.P.; Alcalá, S.G.S. A systematic literature
review of machine learning methods applied to predictive maintenance. Comput. Ind. Eng.
Bousdekis, A.; Lepenioti, K.; Apostolou, D.; Mentzas, G. Decision Making in Predictive Maintenance:
Literature Review and Research Agenda for Industry 4.0. IFAC-PapersOnLine
Sensors 2020,20, 2425 21 of 25
Werner, A.; Zimmermann, N.; Lentes, J. Approach for a Holistic Predictive Maintenance Strategy by
Incorporating a Digital Twin. Procedia Manuf. 2019,39, 1743–1751. doi:10.1016/j.promfg.2020.01.265.
Sakib, N.; Wuest, T. Challenges and Opportunities of Condition-based Predictive Maintenance: A Review.
Procedia CIRP 2018,78, 267–272. doi:10.1016/j.procir.2018.08.318.
Oxford Learner’s Dictionaries. Oxford Advanced Learner’s Dictionary. 2019.
(accessed on 23
Mobley, R.K. Impact of Maintenance. In An Introduction to Predictive Maintenance; Elsevier: Amsterdam, The
Netherlands, 2002; pp. 1–22. doi:10.1016/b978-075067531-4/50001-4.
Armendia, M.; Alzaga, A.; Peysson, F.; Euhus, D. Twin-Control Approach. In Twin-Control: A Digital Twin
Approach to Improve Machine Tools Lifecycle; Armendia, M., Ghassempouri, M., Ozturk, E., Peysson, F., Eds.;
Springer International Publishing: Cham, Switzerland, 2019; pp. 23–38. doi:10.1007/978-3-030-02203-7_2.
Schluse, M.; Priggemeyer, M.; Atorf, L.; Rossmann, J. Experimentable Digital Twins—Streamlining
Simulation-Based Systems Engineering for Industry 4.0. IEEE Trans. Ind. Inform.
Tinga, T.; Loendersloot, R. Physical Model-Based Prognostics and Health Monitoring to Enable Predictive
Maintenance. In Predictive Maintenance in Dynamic Systems: Advanced Methods, Decision Support Tools and
Real-World Applications; Lughofer, E., Sayed-Mouchaweh, M., Eds.; Springer International Publishing: Cham,
Switzerland, 2019; pp. 313–353. doi:10.1007/978-3-030-05645-2_11.
Bektas, O.; Marshall, J.; Jones, J.A. Comparison of Computational Prognostic Methods for Complex
Systems Under Dynamic Regimes: A Review of Perspectives. Arch. Comput. Methods Eng.
Liao, L.; Köttig, F. A hybrid framework combining data-driven and model-based methods for system
remaining useful life prediction. Appl. Soft Comput. 2016,44, 191–199. doi:10.1016/j.asoc.2016.03.013.
Lughofer, E.; Sayed-Mouchaweh, M. (Eds.) Predictive Maintenance in Dynamic Systems; Springer International
Publishing: Cham, Switzerland, 2019. doi:10.1007/978-3-030-05645-2.
Cernuda, C. On the Relevance of Preprocessing in Predictive Maintenance for Dynamic Systems. In Predictive
Maintenance in Dynamic Systems: Advanced Methods, Decision Support Tools and Real-World Applications;
Lughofer, E., Sayed-Mouchaweh, M., Eds.; Springer International Publishing: Cham, Switzerland, 2019; pp.
Fernandez-Delgado, M.; Cernadas, E.; Barro, S.; Amorim, D. Do we Need Hundreds of Classiﬁers to Solve
Real World Classiﬁcation Problems? J. Mach. Learn. Res. 2014,15, 3133–3181.
Mathew, J.; Luo, M.; Pang, C.K. Regression kernel for prognostics with support vector machines.
In Proceedings of the 2017 22nd IEEE International Conference on Emerging Technologies and Factory
Automation (ETFA), Limassol, Cyprus, 12–15 September 2017; pp. 1–5. doi:10.1109/ETFA.2017.8247740.
Wolpert, D.H.; Macready, W.G. No free lunch theorems for optimization. IEEE Trans. Evol. Comput.
1, 67–82. doi:10.1109/4235.585893.
Gerum, P.C.L.; Altay, A.; Baykal-Gürsoy, M. Data-driven predictive maintenance scheduling policies for
railways. Transp. Res. Part Emerg. Technol. 2019,107, 137–154. doi:10.1016/j.trc.2019.07.020.
Qiu, X.; Wu, W.; Wang, S. Remaining useful life prediction of lithium-ion battery based on improved cuckoo
search particle ﬁlter and a novel state of charge estimation method. J. Power Sources
Sampaio, G.S.; de Aguiar Vallim Filho, A.R.; da Silva, L.S.; da Silva, L.A. Prediction of Motor Failure Time
Using An Artiﬁcial Neural Network. Sensors 2019,19, 4342. doi:10.3390/s19194342.
Rengasamy, D.; Jafari, M.; Rothwell, B.; Chen, X.; Figueredo, G.P. Deep Learning with Dynamically
Weighted Loss Function for Sensor-Based Prognostics and Health Management. Sensors
Zhang, S.; Zhai, B.; Guo, X.; Wang, K.; Peng, N.; Zhang, X. Synchronous estimation of state of health
and remaining useful lifetime for lithium-ion battery using the incremental capacity and artiﬁcial neural
networks. J. Energy Storage 2019,26, 100951. doi:10.1016/j.est.2019.100951.
Sensors 2020,20, 2425 22 of 25
En, T.Y.; Ki, M.S.; Hui, N.T.; Jie, T.J.; Bin Mohamed Vusoff, M.A. Predictive Maintenance of a Train
System Using a Multilayer Perceptron Artiﬁcial Neural Network. In Proceedings of the 2018 International
Conference on Intelligent Rail Transportation (ICIRT), Singapore, 12–14 December 2018; pp. 1–5.
Pan, Y.; Hong, R.; Chen, J.; Singh, J.; Jia, X. Performance degradation assessment of a wind
turbine gearbox based on multi-sensor data fusion. Mech. Mach. Theory
Zhou, K.B.; Zhang, J.Y.; Shan, Y.; Ge, M.F.; Ge, Z.Y.; Cao, G.N. A Hybrid Multi-Objective Optimization Model
for Vibration Tendency Prediction of Hydropower Generators. Sensors
,19, 2055. doi:10.3390/s19092055.
Javed, K.; Gouriveau, R.; Zerhouni, N. A New Multivariate Approach for Prognostics Based on
Extreme Learning Machine and Fuzzy Clustering. IEEE Trans. Cybern.
Silva, W.; Capretz, M. Assets Predictive Maintenance Using Convolutional Neural Networks. In Proceedings
of the 2019 20th IEEE/ACIS International Conference on Software Engineering, Artiﬁcial Intelligence,
Networking and Parallel/Distributed Computing (SNPD), Toyama, Japan, 8–11 July 2019; pp. 59–66.
Gouarir, A.; Martínez-Arellano, G.; Terrazas, G.; Benardos, P.; Ratchev, S. In-process Tool Wear Prediction
System Based on Machine Learning Techniques and Force Analysis. Procedia CIRP
Nguyen, K.T.; Medjaher, K. A new dynamic predictive maintenance framework using deep learning for
failure prognostics. Reliab. Eng. Syst. Saf. 2019,188, 251–262. doi:10.1016/j.ress.2019.03.018.
Deng, Y.; Bucchianico, A.D.; Pechenizkiy, M. Controlling the accuracy and uncertainty trade-off in RUL
prediction with a surrogate Wiener propagation model. Reliab. Eng. Syst. Saf.
Markiewicz, M.; Wielgosz, M.; Boche´nski, M.; Tabaczy ´nski, W.; Konieczny, T.; Kowalczyk, L. Predictive
Maintenance of Induction Motors Using Ultra-Low Power Wireless Sensors and Compressed Recurrent
Neural Networks. IEEE Access 2019,7, 178891–178902. doi:10.1109/ACCESS.2019.2953019.
Yu, Y.; Hu, C.; Si, X.; Zheng, J.; Zhang, J. Averaged Bi-LSTM networks for RUL prognostics with non-life-cycle
labeled dataset. Neurocomputing 2020. doi:10.1016/j.neucom.2020.03.041.
Zhou, D.; Al-Durra, A.; Zhang, K.; Ravey, A.; Gao, F. Online remaining useful lifetime prediction of proton
exchange membrane fuel cells using a novel robust methodology. J. Power Sources
Ali, J.B.; Chebel-Morello, B.; Saidi, L.; Malinowski, S.; Fnaiech, F. Accurate bearing remaining useful life
prediction based on Weibull distribution and artiﬁcial neural network. Mech. Syst. Signal Process.
56–57, 150–172. doi:10.1016/j.ymssp.2014.10.014.
Sun, J.; Wang, F.; Ning, S. Aircraft air conditioning system health state estimation and prediction for
predictive maintenance. Chin. J. Aeronaut. 2019. doi:10.1016/j.cja.2019.03.039.
Curcurù, G.; Galante, G.; Lombardo, A. A predictive maintenance policy with imperfect monitoring. Reliab.
Eng. Syst. Saf. 2010,95, 989–997. doi:10.1016/j.ress.2010.04.010.
Son, J.; Zhou, S.; Sankavaram, C.; Du, X.; Zhang, Y. Remaining useful life prediction based on noisy
condition monitoring signals using constrained Kalman ﬁlter. Reliab. Eng. Syst. Saf.
Cui, L.; Wang, X.; Xu, Y.; Jiang, H.; Zhou, J. A novel Switching Unscented Kalman Filter
method for remaining useful life prediction of rolling bearing. Measurement
Saha, B.; Goebel, K.; Christophersen, J. Comparison of prognostic algorithms for estimating remaining useful
life of batteries. Trans. Inst. Meas. Control 2009,31. doi:10.1177/0142331208092030.
Ruiz-Sarmiento, J.R.; Monroy, J.; Moreno, F.A.; Galindo, C.; Bonelo, J.M.; Gonzalez-Jimenez, J. A predictive
model for the maintenance of industrial machinery in the context of industry 4.0. Eng. Appl. Artif. Intell.
2020,87, 103289. doi:10.1016/j.engappai.2019.103289.
Zenisek, J.; Holzinger, F.; Affenzeller, M. Machine learning based concept drift detection for predictive
maintenance. Comput. Ind. Eng. 2019,137, 106031. doi:10.1016/j.cie.2019.106031.
Sensors 2020,20, 2425 23 of 25
Traini, E.; Bruno, G.; D’Antonio, G.; Lombardi, F. Machine Learning Framework for Predictive Maintenance
in Milling. IFAC-PapersOnLine 2019,52, 177–182. doi:10.1016/j.ifacol.2019.11.172.
Du, J.; Zhang, W.; Zhang, C.; Zhou, X. Battery remaining useful life prediction under coupling stress based
on support vector regression. Energy Procedia 2018,152, 538–543. doi:10.1016/j.egypro.2018.09.207.
Yan, M.; Wang, X.; Wang, B.; Chang, M.; Muhammad, I. Bearing remaining useful life prediction
using support vector machine and hybrid degradation tracking model. ISA Trans.
Li, Q.; Xing, L.; Liu, W.; Ba, W. Adaptive Soft Sensor Based on a Moving Window Just-in-time Learning
LS-SVM for Distillation Processes. IFAC-PapersOnLine 2015,48, 51–56. doi:10.1016/j.ifacol.2015.12.099.
Langone, R.; Alzate, C.; Ketelaere, B.D.; Vlasselaer, J.; Meert, W.; Suykens, J.A. LS-SVM based spectral
clustering and regression for predicting maintenance of industrial machines. Eng. Appl. Artif. Intell.
37, 268–278. doi:10.1016/j.engappai.2014.09.008.
Ruiz-Gonzalez, R.; Gomez-Gil, J.; Gomez-Gil, F.; Martínez-Martínez, V. An SVM-Based Classiﬁer for
Estimating the State of Various Rotating Components in Agro-Industrial Machinery with a Vibration Signal
Acquired from a Single Point on the Machine Chassis. Sensors
,14, 20713–20735. doi:10.3390/s141120713.
Bukhsh, Z.A.; Saeed, A.; Stipanovic, I.; Doree, A.G. Predictive maintenance using tree-based classiﬁcation
techniques: A case of railway switches. Transp. Res. Part Emerg. Technol.
Ravikumar, S.; Ramachandran, K. Tool Wear Monitoring of Multipoint Cutting Tool using Sound
Signal Features Signals with Machine Learning Techniques. Mater. Today Proc.
Susto, G.A.; Schirru, A.; Pampuri, S.; Pagano, D.; McLoone, S.; Beghi, A. A predictive maintenance system for
integral type faults based on support vector machines: An application to ion implantation. In Proceedings
of the 2013 IEEE International Conference on Automation Science and Engineering (CASE), Madison, WI,
USA, 17–21 August 2013; pp. 195–200. doi:10.1109/CoASE.2013.6653952.
Xue, Z.; Zhang, Y.; Cheng, C.; Ma, G. Remaining useful life prediction of lithium-ion batteries with adaptive
unscented kalman ﬁlter and optimized support vector regression. Neurocomputing
63. Graves, A. Neural Networks; Springer; Berlin, Germany, 2012.
Bishop, C. Neural Networks For Pattern Recognition; Oxford University Press: Oxford, UK, 2005; Volume 227.
Wang, Z.; Oates, T. Imaging Time-Series to Improve Classiﬁcation and Imputation. arXiv
Chen, J.; Chen, W.; Huang, C.; Huang, S.; Chen, A. Financial Time-Series Data Analysis Using
Deep Convolutional Neural Networks. In Proceedings of the 2016 7th International Conference
on Cloud Computing and Big Data (CCBD), Macau, China, 16–18 November 2016; pp. 87–92.
Zhang, B.; Zhang, S.; Li, W. Bearing performance degradation assessment using long short-term memory
recurrent network. Comput. Ind. 2019,106, 14–29. doi:10.1016/j.compind.2018.12.016.
Uhlmann, E.; Pontes, R.P.; Geisert, C.; Hohwieler, E. Cluster identiﬁcation of sensor data for
predictive maintenance in a Selective Laser Melting machine tool. Procedia Manuf.
Cao, Q.; Samet, A.; Zanni-Merk, C.; de Bertrand de Beuvron, F.; Reich, C. An Ontology-based Approach for
Failure Classiﬁcation in Predictive Maintenance Using Fuzzy C-means and SWRL Rules. Procedia Comput.
Sci. 2019,159, 630–639. doi:10.1016/j.procs.2019.09.218.
Daher, A.; Hoblos, G.; Khalil, M.; Chetouani, Y. New prognosis approach for preventive and predictive
maintenance—Application to a distillation column. Chem. Eng. Res. Des.
Olson, R.; La Cava, W.; Orzechowski, P.; Urbanowicz, R.; Moore, J. PMLB: A Large Benchmark Suite for
Machine Learning Evaluation and Comparison. BioData Min. 2017,10. doi:10.1186/s13040-017-0154-4.
Ahmed, N.; Atiya, A.; Gayar, N.; El-Shishiny, H. An Empirical Comparison of Machine Learning Models for
Time Series Forecasting. Econom. Rev. 2010,29, 594–621. doi:10.1080/07474938.2010.481556.
Sensors 2020,20, 2425 24 of 25
Wang, X.; Qin, Y.; Wang, Y.; Xiang, S.; Chen, H. ReLTanh: An activation function with vanishing gradient
resistance for SAE-based DNNs and its application to rotating machinery fault diagnosis. Neurocomputing
2019,363, 88–98. doi:10.1016/j.neucom.2019.07.017.
Krishnan, R.; Jagannathan, S.; Samaranayake, V. Direct Error Driven Learning for Deep Neural Networks
with Applications to Bigdata. Procedia Comput. Sci. 2018,144, 89–95. doi:10.1016/j.procs.2018.10.508.
Webster, J.A.; McNay, D.A.; Lundy, D.; De Sapio, V.; De Sapio, V.; Fan, Q.; Fravel, D.; Houck, G.; Lenchitz, H.;
Mapes, C.; et al. Foundry Products: Competitive Conditions in the U.S. Market; United States International Trade
Commission: Washington, DC, USA, 2005.
Karassik, I.J.; McGuire, T. Centrifugal Pumps; Springer: New York, NY, USA, 1997.
Kim, G.; Kim, H.; Zio, E.; Heo, G. Application of particle ﬁltering for prognostics with measurement
uncertainty in nuclear power plants. Nucl. Eng. Technol. 2018,50, 1314–1323. doi:10.1016/j.net.2018.08.002.
Gohel, H.A.; Upadhyay, H.; Lagos, L.; Cooper, K.; Sanzetenea, A. Predictive Maintenance Architecture
Development for Nuclear Infrastructure using Machine Learning. Nucl. Eng. Technol.
Zhao, Y.; Tong, J.; Zhang, L.; Wu, G. Diagnosis of operational failures and on-demand failures in nuclear
power plants: An approach based on dynamic Bayesian networks. Ann. Nucl. Energy
Nguyen, H.P.; Liu, J.; Zio, E. A long-term prediction approach based on long short-term memory
neural networks with automatic parameter optimization by Tree-structured Parzen Estimator and
applied to time-series data of NPP steam generators. Appl. Soft Comput.
Hu, Q.; Ohata, E.F.; Silva, F.H.; Ramalho, G.L.; Han, T.; Filho, P.P.R. A new online approach for
classiﬁcation of pumps vibration patterns based on intelligent IoT system. Measurement
Panda, A.K.; Rapur, J.S.; Tiwari, R. Prediction of ﬂow blockages and impending cavitation in centrifugal
pumps using Support Vector Machine (SVM) algorithms based on vibration measurements. Measurement
2018,130, 44–56. doi:10.1016/j.measurement.2018.07.092.
Wang, J.; Zhang, L.; Zheng, Y.; Wang, K. Adaptive prognosis of centrifugal pump under variable operating
conditions. Mech. Syst. Signal Process. 2019,131, 576–591. doi:10.1016/j.ymssp.2019.06.008.
Tao, X.; Wang, Z.; Ma, J.; Fan, H. Study on fault detection using wavelet packet and SOM neural network. In
Proceedings of the IEEE 2012 Prognostics and System Health Management Conference (PHM-2012 Beijing),
Beijing, China, 23–25 May 2012; pp. 1–5. doi:10.1109/PHM.2012.6228817.
Moleda, M.; Momot, A.; Mrozek, D. Predictive Maintenance of Boiler Feed Water Pumps Using SCADA
Data. Sensors 2020,20, 571. doi:10.3390/s20020571.
Liu, Q.; Dong, M.; Lv, W.; Geng, X.; Li, Y. A novel method using adaptive hidden semi-Markov model
for multi-sensor monitoring equipment health prognosis. Mech. Syst. Signal Process.
Tse, Y.L.; Cholette, M.E.; Tse, P.W. A multi-sensor approach to remaining useful life estimation for a slurry
pump. Measurement 2019,139, 140–151. doi:10.1016/j.measurement.2019.02.079.
He, R.; Dai, Y.; Lu, J.; Mou, C. Developing ladder network for intelligent evaluation system: Case of
remaining useful life prediction for centrifugal pumps. Reliab. Eng. Syst. Saf.
Hundi, P.; Shahsavari, R. Comparative studies among machine learning models for performance
estimation and health monitoring of thermal power plants. Appl. Energy
Alamaniotis, M.; Grelle, A.; Tsoukalas, L.H. Regression to fuzziness method for estimation of
remaining useful life in power plant components. Mech. Syst. Signal Process.
Dong, S.; Luo, T. Bearing degradation process prediction based on the PCA and optimized LS-SVM model.
Measurement 2013,46, 3143–3152. doi:10.1016/j.measurement.2013.06.038.
Sensors 2020,20, 2425 25 of 25
Koutroulis, G.; Thalmann, S. Challenges from Data-Driven Predictive Maintenance in Brownﬁeld Industrial
Settings. In Business Information Systems Workshops; Abramowicz, W., Paschke, A., Eds.; Springer International
Publishing: Cham, Switzerland, 2019; pp. 461–467.
Wotawa, F. Reasoning from First Principles for Self-adaptive and Autonomous Systems. In Predictive
Maintenance in Dynamic Systems: Advanced Methods, Decision Support Tools and Real-World Applications;
Lughofer, E., Sayed-Mouchaweh, M., Eds.; Springer International Publishing: Cham, Switzerland, 2019; pp.
Gross, K.C.; Li, D. Machine Learning Innovation for High Accuracy Remaining Useful Life (RUL) Estimation
for Critical Assets in IoT Infrastructures. In Proceedings of the International Conference Internet Computing
and Internet of Things, Las Vegas, NV, USA, 30 July–2 August 2018; pp. 101–107.
Predictive maintenance tool monitors pump and motor with wireless access and protective alerts.
World Pumps 2019,2019, 11. doi:10.1016/s0262-1762(18)30411-5.
Kuniavsky, M. Introduction The middle of Moore’s law. In Smart Things: Ubiquitous Computing User Experience
Design; Elsevier: Amsterdam, The Netherlands, 2010; pp. 3–11. doi:10.1016/b978-0-12-374899-7.00001-1.
2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).