Challenges To IoT-enabled Predictive Maintenance

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2957029, IEEE Internet of
Things Journal
IEEE INTERNET OF THINGS JOURNAL, VOL. , NO. , MONTH YEAR 1
Challenges to IoT-enabled Predictive Maintenance

for Industry 4.0
Michele Compare, Piero Baraldi, and Enrico Zio, Senior Member, IEEE
Abstract—The Industry 4.0 paradigm is boosting the relevance prognostic tasks is often referred to as Prognostics and Health
of Predictive Maintenance (PdM) for manufacturing and pro- Management (PHM, [8], [9], [10], [11], [12]). The capability
duction industries. PdM strongly relies on Internet of Things of performing these tasks with sufficient accuracy provides
(IoT), which digitalizes the physical actions allowing human-to-
human, human-to-machine and machine-to-machine connections the opportunity of setting efficient, just-in-time and just-right
for intelligent perception. Several issues still need to be addressed maintenance strategies: in other words, providing the right part
for reaching the maturity stage for widespread application of to the right place at the right time. This opportunity is big,
PdM. To do this, IoT needs to be empowered with data science ca- because doing this would maximize production profits and
pabilities, to reach the ultimate objective of digitalization, which minimize all costs and losses, including asset ones ([13]).
is supporting decision making to optimally act on the physical
systems. In this paper, we present a comprehensive outlook of Boosted by the intuitive and appealing potential of PdM,
the current PdM issues, with the final aim of providing a deeper the industry is making significant investments for equipping
understanding of the limitations and strengths, challenges and itself with the elements necessary for deploying PdM. For
opportunities of this dynamic maintenance paradigm. This is example, the investments by the Italian industry in research &
done through extensive research and analysis of the scientific and development & innovation for Industry 4.0 increased by 15%
technical literature. On this basis, the work outlines some main
research issues to be addressed for the successful development in 2017, a significant part of which allocated to PdM [17],
and deployment of IoT-enabled PdM in industry. and similar investments are reported in other countries (e.g.,
[18]). This situation has sparked the birth of a large number of
Index Terms—Predictive Maintenance, IoT, Industry 4.0.
PdM specialized companies, commercial softwares, dedicated
journals and conferences, etc.
I. I NTRODUCTION The Internet of Things (IoT) is a main pillar of PdM ([6],
[7]), as it allows translating physical actions from machines
Industry 4.0, the fourth industrial revolution ([1], [2], [3]),
into digital signals used for PdM. Namely, IoT continuously
aims at creating smart factories, equipped with disruptive tech-
streams data from sensors such as temperature, vibration, etc.
nologies such as advanced robotics, 3-D printing, high com-
and from other sources, such as a machine Programmable
puting power and connectivity, etc., which are integrated with
Logic Controller (PLC), Manufacturing Execution system
analytical and cognitive technologies that enable machine-to-
(MES) terminals, Computerized Maintenance Management
machine (M2M) and machine-to-human (M2H) communica-
systems (CMMSs, [14], [15], [16]), or even an Enterprise
tion. The smart factory provides the opportunity of offering
Resource Planning (ERP) system. These pieces of information
new services and products to customers, with efficiency, stan-
provide the basis for setting PdM approaches.
dards of quality and reliability higher than before. These allow
Up to now, the focus of the effort made has been mainly on the
expanding the value chain by generating new business models
development of hardware (i.e., IoT, smart meters, etc. [6], [7],
that create value for customers and revenue for manufacturing
[19], [20]) and software (e.g., PHM tools, platforms for IoT
companies ([4], [5]).
interconnection and clouding, etc. [21], [22], [23]), for tracking
One of the opportunities (among others) most spoken of in
the health state of monitored components. On the other hand,
Industry 4.0 is Predictive Maintenance (PdM), which makes
the industrial-scale deployment of PdM involves many other
use of condition monitoring data to detect anomalies (i.e.,
aspects and impacts various sectors of the workplace involved
recognize deviations from normal operating conditions) in
in maintenance (i.e., workers can use smart systems, main-
production processes, manufacturing equipment and products,
tenance engineers can analyze big data for the maintenance
diagnose (i.e., characterize the occurring abnormal state) and
process), logistics (spare parts and warehouse management
prognose (i.e., predict the future evolution of the abnormal
can be driven by the PHM results), Occupational Health,
state up to failure). The set of detection, diagnostic and
Safety & Environment (OHSE, smart system information can
Michele Compare, Piero Baraldi and Enrico Zio are with the Department be used for updated monitoring of risks), design (the use of
of Energy, Politecnico di Milano, Italy. e-mail: michele.compare@polimi.it, smart components may lead to different reliability allocation
piero.baraldi@polimi.it, enrico.zio@polimi.it solutions), top management (new business opportunities can
Michele Compare and Enrico Zio are with Aramis s.r.l., Milano, Italy. e-
mail: michele.compare@aramis3d.com, enrico.zio@aramis3d.com arise in services), etc. [6]. To bridge the gap, IoT needs to be
Enrico Zio is an Eminent Scholar, Department of Nuclear Engineering, integrated with data science and modeling capabilities, to reach
College of Engineering, Kyung Hee University, Republic of Korea and with the ultimate objective of digitalization, which is supporting
MINES ParisTech, PSL Research University, CRC, Sophia Antipolis, France.
e-mail: enrico.zio@mines-paristech.fr decision making to optimally act on the physical systems.
Manuscript received TBD; revised TBD. In this paper, we present a comprehensive outlook of the
2327-4662 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/JIOT.2019.2957029, IEEE Internet of
Things Journal
Fig. 1. PdM development activities
current PdM issues, with the final aim of providing a deeper so. Rather, the opportunity for maintenance in Industry 4.0
understanding of the limitations and strengths, challenges and lies in the possibility of defining the optimal maintenance for
opportunities of this dynamic maintenance paradigm. This is every component, taking into account its specificity within the
done through extensive research and analysis of the scientific system, e.g., applicable safety and environmental legislation,
and technical literature. On this basis, the work outlines quality standard, importance for business, physical and func-
some main research issues to be addressed for the successful tional characteristics, etc.
development and deployment of IoT-enabled PdM in industry. Reliability Centered Maintenance (RCM, [27], [28], [29]) was
The remainder of the paper is presented according to the PdM proposed in the 1970’s (i.e., at the beginning of the third
development activities represented in Figure 1. The selection industrial revolution determined by automation) to address
of the components which would benefit most from PdM is a the issue of selecting the best maintenance strategy for every
fundamental issue to address for the effective application of component in a system. Nowadays, RCM is standardized
PdM. This is overviewed in Section II. Once the components for the different industrial sectors (e.g., [25], [30]) and is
eligible for PdM have been identified, the next step is to supported by the availability of advanced CMMS, with many
properly design the IoT infrastructure in support to PdM. The success cases reported (e.g., [27], [31], [32]).
issues related to this topic are presented in Section III. Section The main idea of RCM is to concentrate the maintenance ef-
IV focuses on the issues related to the development of the al- forts on the components of the asset most critical for safety and
gorithms and methods for PdM, which are generally addressed business, and apply to them the most effective maintenance
once the monitoring data from IoT are available. The last step approach, as resulting from the analysis of their reliability
concerns the exploitation of IoT-enabled monitoring, to really characteristics. To do this, RCM relies on a decision flowchart,
ensure that PdM brings an added value. The decision making whose first question is about the possibility of monitoring the
issues to achieve this objective are presented in Section V. condition of the component, i.e., a physical variable indicative
Finally, conclusions of the work are given in Section VI. of the component degradation state, and defining a threshold
value for it, at which to do maintenance on the component
II. W HICH COMPONENTS FOR PREDICTIVE MAINTENANCE to avoid its failure with major consequences ([25], [30]). In
Maintenance approaches are generally divided into two case of affirmative answer, CBM can be considered technically
main groups: corrective maintenance (CM) and preventive feasible; otherwise, the decision flowchart proceeds with other
maintenance. Under CM, the components are operated until questions about the reliability characteristics of the component
failure; then, repair or renovation actions are performed. to check the applicability of scheduled maintenance; if also
Preventive maintenance, instead, encompasses all actions per- this is not applicable, the component is inevitably run to failure
formed in an attempt to retain an item in specified conditions, and taken care of by corrective maintenance.
by providing systematic inspection, detection and prevention The rationale underlying RCM is applicable to Industry 4.0,
of incipient failures (e.g., [8], [24], [25]). Accordingly, preven- but with some major limitations:
tive approaches can be further divided into three sub-groups • The first RCM question on the possibility of condition
([8], [26]): Scheduled Maintenance (SM), if the actions are monitoring for the applicability of CBM can be mislead-
performed based on a pre-fixed basis, Condition-Based Main- ing in the practice of Industry 4.0: whilst it goes without
tenance (CBM), which uses condition monitoring to identify saying that CBM is doable in case of affirmative answer
problems at an early stage and perform maintenance when the to the question, a negative answer does not necessarily
degradation level reaches a threshold, and PdM, which can be imply that CBM must be abandoned. In fact, PHM
regarded as an advancement of CBM: the degradation of the approaches have been developed (e.g., Principal Compo-
component is predicted in the future and its Remaining Useful nent Analysis (PCA), Auto-Associative Kernel regression
Life (RUL) is estimated. (AAKR), Self Organizing Maps (SOM), etc. [33], [34],
A tempting misconception within the Industry 4.0 paradigm [35], [36], [37]) for detecting early failures in a com-
is that PdM is always the best maintenance policy. This is not ponent based on multiple signals not directly measuring
Things Journal
the condition of that component. Indeed, PHM methods the uncertainty in the estimations of the degradation state
of feature extraction and selection (e.g., wavelet tranform thanks to the measurement collected by sensors or even
[38], [39], [40]) can find combinations of features from inspections, i.e., VoI is used for selecting explorative and
the available signals that although not directly measuring inspection actions ([51], [52], [53]). Then, it is a relative
the component degradation state, can infer it, and CBM value, which does not allow for a fair comparison of PdM
can be developed on this basis. with the other approaches not based on sensor monitoring.
• PdM does not enter the decision flowchart for main- The second issue limiting the VoI approach application is
tenance selection. A positive answer to the first RCM that the algorithms adopted are very time consuming and
question on CBM does not necessarily mean that also applicable to scaled-down case studies, only, in which the
PdM is feasible, as the condition monitoring for CBM number of state-action pairs is not large.
may not provide the information needed for PdM (e.g., A model for evaluating the system-level value of PdM has
[41]). been proposed in [55], within a real options framework.
• Cost-effectiveness of the maintenance strategy must be PdM is seen as a tool for the Decision Maker (DM) to
considered, as CBM and PdM require investment costs invest options of performing maintenance actions in the
in software, instrumentation, knowledge, etc. which must future and a cost–benefit–risk model is developed. Some
be justified by the benefits they can yield. issues remain for its application to industrial practice,
From these considerations, it emerges that RCM needs to be including the need of estimating the difference in the costs
extended for its application to the Industry 4.0 context: clear of performing CM instead of a RUL-driven maintenance.
and solid ways are needed to guide the decision makers in the Moreover, although [55] considers time-dependent RUL
identification of those components for which PdM would be predictions, these are not linked to the performance of the
the right maintenance choice. predictive algorithms (e.g., accuracy, precision, etc.), and a
Brownian motion process is used to describe the evolution of
the economic indicators related to RUL predictions entering
A. Economics of PdM the options model. In addition, the model considers CM
Development of IoT-enabled PdM for Industry 4.0 makes as the only possible alternative to PdM, and not other
sense if it is proved to be more profitable than the other preventive maintenance approaches: however, there can be
maintenance approaches. Maintenance cost models must, cases in which the economic performances of CM and
then, be developed to evaluate the economic benefits of PdM. SM are superior to those of CBM and PdM ([8]). Refined
However, only few attempts have been made in this sense analytical methods are developed in [42], for the cost-benefit
([42]), in spite of relevance that this issue for the decision ı̀s analysis of canary-based PHM; in [56], [57], to maximize the
in investing in IoT for PdM. component resilience, which is defined as a combination of
A few works (e.g., [43], [44], [45], [46], [47], [48], [49]) reliability and restoration, the latter being a function of the
have attempted to evaluate the cost-benefit of PdM through PHM characteristics; in [58], where a life-cycle maintenance
the commonly used financial metrics such as Return on cost analysis framework is developed, which considers
Investment (RoI), Net Cash Flow, Cumulative Cash Flow, time-dependent false and missed alarms for fault diagnosis;
Payback, Net Present Value, and Internal Rate of Return. in [59] and [60], where time-variant metrics of the literature
These works, however, rely on simulation instead of ([61]) are linked to component reliability and availability,
developing general analytical approaches ([42]). respectively, to derive the economic performance of PHM
Another cost-benefit metric proposed is the Technical Value capabilities of different quality levels. These analytical
(TV, [50]), which accounts for the performance in detection, approaches, however, do not fully capture the dynamics of
diagnostics and prognostics of critical failure modes and the the CBM context, where a decision must be taken every time
costs associated with false alarms. However, TV contains the PHM algorithms are run.
cost terms that are difficult to estimate (e.g., the savings The enhancement of the economics models is a mandatory
realized by isolating a fault in advance) and it makes use of condition for the industry to unleash investments in IoT for
constant performance metrics, i.e., independent on time (e.g., PdM. The reviewed literature is schematized in Table II-A.
the probability of a failure mode). Finally, TV does not give
due account to erroneous detection, diagnosis and prognosis.
Partially Observable Markov Decision Processes (POMDP,
[51], [52], [53]), have been used to estimate the Value of B. PdM for production and product
Information (VoI, [54]) of data measured by sensors installed To answer the topical question ”is PdM convenient for this
on civil infrastructures, accounting for the uncertainty in equipment?”, we need to distinguish the case where PdM is
the condition monitoring. Roughly speaking, VoI is the considered for a product from that in which it is applied to
maximum cost a decision-maker is willing to pay for getting the equipment of a production process.
the information, which is worth acquiring only if its value In the former case, the economic justification can be relatively
is above its cost ([54]). Although this framework seems simple: several companies of different industrial sectors (i.e.,
very promising, there are two main issues preventing its manufacturing [62], aviation [13], [63], [64], mining [65],
application to Industry 4.0. Namely, the VoI definition relates energy [22], [66], etc.) look into PdM simply because it
to the expected savings that can be yielded by reducing gives commercial competitiveness. Furthermore, new sources
Things Journal
Fig. 2. scheme of a production process in the automotive industry
RoI TV Cost VoI Resilience

[43], [44] [45], [46], [47]
Time independent PHM metrics [50] [42] [51], [52], [53] [56], [57]
[48], [49]
Time dependent PHM metrics [58] [59] [60]
TABLE I
E CONOMICS OF P D M: SYNOPTIC
of income can generate thanks to new opportunities of added PdM. Then, this is doable only if we are able to both estimate
values in service, by taking over portions of the clients’ the other indirect costs of failure, such as costs for re-filling
business risks and other (financial) burdens: the possibility of the buffer, costs of warehouse, costs related to conservative
new business may be a sound justification per se for investing settings of scheduled maintenance intervals that result in over-
in PdM. maintenance expenses, etc., and prove that these costs are
In case of manufacturing processes, the value of PdM and, thus large enough to justify significant investments in PdM. This
of the IoT infrastucture, is more difficult to assess. To show emphasizes the need for sound cost models encoding PdM for
this, we consider the example of a manufacturing process in IoT investment justification.
the automotive industry but draw some general considerations.
Figure 2 shows the scheme of a manufacturing process made III. I OT INFRASTRUCTURE AND DATA MANAGEMENT FOR
of different steps, possibly spaced by buffers. Generally speak- PDM
ing, the more stringent the application of the just-in-time A major misbelief in Industry 4.0 concerns the assump-
paradigm, the smaller the buffers; the later the process step, tion that larger amounts of acquired data and, thus, more
the larger the value of the half-processed units and, thus, the widespread and performing IoT networks, always result in
smaller the buffers. From the PdM perspective, the earlier better performance of PdM. This is not so, as acquiring,
production phases (i.e., shell manufacturing through welding, storing, maintaining and analyzing data entail a cost that
milling, etc.) are the most promising ones, as these are per- increases with the amount of data. As pointed out in [11],
formed in the capital-intensive parts of the plant, with robots, the final objective of digitalization should be that of acquiring
transportation means, welding systems, etc. On the contrary, smart data, rather than big data. To show this, we briefly
in the latest step (i.e., assembly), where the production flow is report about an experience concerning the data acquisition
more time-sensitive, there are mainly screwdrivers, traveling from bearings installed in a manufacturing plant, to outline
cranes, etc., in relatively large redundancy and with the largest general considerations. In that plant, the raw bearing vibration
manning level. data are acquired at a frequency of 1.6 kHz. Due to data
In this scheme, the value of PdM heavily depends on the storage limitations, these raw data are not stored into the
buffers, whose level Bf to withstand a downtime of D hours servers. Rather, only two features are extracted from the raw
of the upstream production step can be estimated as: signal, i.e., Root Mean Square (RMS) acceleration and peak-
to-peak vibration, which are then averaged on a period of 0.5
1
Bf = ×D (1) seconds and recorded in the data storage system.
takT These two features have proved to be effective in identifying
where takT is the takt time in hours (i.e., the average time abrupt failures. Nonetheless, their informational content is not
between the start of production of one unit and the start of useful for developing a PdM approach. To wit, Figures 3b and
production of the next unit). To consider reasonable values, 3c refer to a bearing case study and show the available bearing
we can conservatively assume an extremely long downtime acceleration RMS and peak-to-peak values, respectively, over a
D = 10 h and takT = 6 min = 0.1 h; then, we get Bf = 100. time window of almost 130 samples (i.e., almost 65 seconds),
If Cp is the cost of the product at the end of the production whereas Figure 3a shows the raw signal data relevant to the
step, the mobilized capital reads M c = Cp ×Bf . For example, first part of almost 5 seconds of the same time window.
if Cp = 3000 e, then M c = 3000 000 e. Assuming a capital From their comparison, it is clear that the averaging leads to
cost of 10%/year, it turns out that 300 000 e per year is the cost hiding the information contained in the raw signal, as signals
that PdM has to payoff to avoid business interruptions. This averaged on relatively long time windows encode different
value is much smaller than any massive investment in IoT for working conditions with variable loads and speeds, which
Things Journal
good sense, nontheless a general framework for optimizing

the management of the IoT infrastructure supporting PdM is
needed. This is still lacking, to the authors best knowledge.
Then, we give insights to formalize the issue, while leaving
its solution as a urgent challenge for researchers.
Consider a simplified model of a data acquisition chain from
a piece of equipment monitored by S sensors, s = 1, ..., S,
which can be positioned in locations l ∈ L = {1, ..., L}
(Figure 4). We introduce matrix P, whose (s, l) entry is set
to 0 when sensor s is not positioned in location l, and to
(a) 1, otherwise. Each sensor acquires data at a bit rate fs · b
Gbit/h, where fs is the sampling frequency in h−1 and b the
bit resolution in Gbit. We consider vector f = [f1 , ..., fS ].
Sensor data are transmitted to a local computing unit at a
maximum rate R1 Gbit/h (second block in Figure 4). The local
computing unit extracts sets of features Φs = [Φ1s , ..., Φφs s ],
which are appended to vector Φ = [Φ1 , ..., ΦS ] (third block
in Figure 4). The features are extracted on a time window
of ∆t hours and every feature extraction requires time dtsj ,
s = 1, ..., S, j = 1, ..., φs . The extracted features are sent to a
server of memory capacity M Gbit, without data compression
(last two blocks in Figure 4). The transferring rate for this
(b) second transmission line is R2 Gbit/h.
Finally, the measurements are performed every τ > ∆t
hours, over a conservatively (i.e., longer) estimated component
lifetime of T hours.
Notice that for the sake of simplicity, the units of measurement
have been kept coherent and proper coefficients are required
to use the units normally adopted in practice. For example, the
sampling frequency is usually measured in KHz: then, we need
to multiply this value by 30 6000 000 to get the corresponding
frequency value in h−1 ; similarly, features are calculated on
time windows of a few seconds, whereby ∆t expressed in
seconds must be divided by 3600 to get the corresponding
(c) value in hours.
The optimization of the sensors allocation can be considered
Fig. 3. (a) Bearing acceleration raw data. (b) Bearing acceleration RMS. (c)
Bearing acceleration peak-to-peak. within the PDA framework ([70]), which seeks the optimal
portfolios of sensor allocation solution X = [P, f , Φ, τ ] such
that
the vibration signals are sensitive to. The analysis of these
signals for prediction can lead to misunderstanding in their
interpretation.
Given that RMS and peak-to-peak values cannot be used for
prediction, it is clear that storing them for long time windows
(the bearing life duration is 3-4 years in the application
considered) is not cost-effective, as we do not need to rely
on values relevant to old conditions to capture abrupt changes
([67], [68], [69]).
Based on these considerations, the final strategy proposed can
be to i) collect vibration RMS and peak-to-peak features;
ii) store them for relatively short time windows (i.e., a few
months); iii) complement this information with features useful
for prediction, extracted from the raw data (i.e., wavelet
transforms in the specific case), acquired every couple of
weeks in a tailored, baseline reference setting of working
conditions.
Although the proposed strategy has been, indeed, effective in
the application considered and seems to be of engineering
Things Journal
should be extended to include costs and benefits related to

overall system resilience [56], flexibility [74], risk, etc. and the
max V oI(X) (2)
X other decision variables of X. In fact, the acquisition frequency
L
X is also a fundamental parameter for the PHM algorithm to
P(s, l) ≤ 1 s ∈ {1, ..., S} (3) give information valuable for PdM: the frequency should be
l=1 large enough to catch the information relevant for PdM, as
XS emerged from the bearing application mentioned above; on
P(s, l) ≤ 1 l ∈ {1, ..., L} (4) the contrary, if it is too large, there can be an overload on
s=1 both the transmission link and the local computing unit, with
L
!
X consequent increments of costs cr and cc . In this respect, Eq.
fs ≤ P(s, l) · fmax s ∈ {1, ..., S} (5)
5 states that rate fs is lager than 0 only when sensor s is
l=1
L
installed, and it is always smaller than fmax , whereas Eq. 6
X indicates that features can be extracted from sensor s only if
φs ≤ P(s, l) s ∈ {1, ..., S} (6)
this has been installed. Moreover, Eq. 7 sets a constraint on
l=1
S the capacity of the data transmission link, which must be large
enough to allow continuous data transmission. The larger the
X
b· fs ≤ R1 (7)
s=1 value of R1 , the larger the cost: cr should be thought of as
S
XXφ s the result of an embedded optimization problem, which finds
max( dtsj , ∆tdeg ) ≤ ∆t (8) the technological solution that guarantees the fulfillment of
s=1 j=1 the constraint in Eq. 7 at the smallest cost. Yet, the larger
S the rate R1 , the larger the time required to process the data
1X
|Φs | · b ≤ R2 (9) by the computing unit that extracts the features, the larger
τ s=1 the computational costs: cc (∆t, Φ) is the minimum cost of
T X
S the technological solution that guarantees that features Φ are
· |Φs | · b ≤ M (10) calculated before a new dataset is acquired. Given Φ, the
τ s=1
smaller the value of ∆t, the larger the computational capability
S
X L
X required.
cs · P(s, l)+ The main requirement for the duration of data collections, ∆t,
s=1 l=1 is that it be large enough to capture all the characteristics of
+cr (R1 ) + cc (∆t, Φ) + cr (R2 ) + cm (M ) ≤ B (11) the component behavior (Eq. 8). We refer to this minimum
where cs is the cost of a sensor, cr , cc and cm are functions that duration as ∆tdeg , whose value depends on the specific
link the cost to the required transmission rate, computational application. For example, in the bearing case study, ∆tdeg
capability and memory capacity, respectively, whereas B is must be large enough that the vibrational signal encodes all
the available budget, which must not be exceeded (Eq. III). In the vibration conditions. If the acquisition time interval is too
words, the VoI function maps the variables in solution X onto short, then the collected signal is not able to represent the
the maximum investment in PdM that a DM is willing to pay bearing functional behavior, whereas if it is too large there can
(Eq. 2). be limitations to the storage of the collected data on the local
Eqs. 3 and 4 state that, respectively, every sensor can be computing unit. Therefore, also in this case, it is necessary to
installed in a single location, at most, and every location can identify a compromise between the richness of information and
accommodate a single sensor, at most. The position and the the storage burden. Moreover, as expressed in Eq. 8, the value
number of sensors are fundamental drivers for PdM to be of ∆t must be larger than the computational time required to
profitable: the larger the number of sensors, the larger the calculate the features.
information available and, thus, the chances of identifying Finally, the time interval τ between two successive data
and extracting information relevant for the development of collections is also an important decision variable, which should
effective PdM, also by exploiting correlations among the be tuned to the specific characteristics of the degradation
signals [71]. Obviously, the larger the number of sensors, the process to ensure that the maximum rate R2 is not exceeded
larger the investment costs. (Eq. 9). In the bearing application, the degradation process is
slow (T > 4 years), whereby a compromise solution must be
found between the granularity of the information over time,
the required data transmission rate and the storage capacity,
which must be large enough to store data over the whole life
of the component T (Eq. 10).
To make the final investment decision, the optimal solutions
Fig. 4. Data acquisition and storage chain corresponding to different budget levels are, then, found and
compared (e.g., [75], [76]).
With respect to the sensor positioning, the relationship with This optimization issue is further challenged by another trend
VoI has been investigated in [72], building on a relatively wide within the Industry 4.0: edge computing ([23], [77]). Namely,
literature (e.g., see [73] for an overview). The model in [72] data computation can be done at the “edge”, meaning that
Things Journal
pre-processing can be performed on the machines where data are collected in experiments and/or on-field, and can be
are gathered from to directly inform machine operators and exploited when the understanding of component operation
maintenance technicians. As data is beginning to approach the is not straightforward or when the component is so
zettabytes (i.e., 1021 bytes), edge computing can be exploited complex that developing an accurate physical model is
to reduce the overall burden on a computer network by prohibitively expensive ([89]). IoT-enabled PdM relies
properly distributing the processing effort to a network outer on this class of algorithms, for which a taxonomy is
nodes, which alleviates the core network traffic and improves proposed in Table II. Given the huge number of works in
application performances ([77]). the field, the reference list is certainly partial, although
it includes literature reviews on specific classes of algo-
A. Dependability of IoT rithms. Notice that in some cases the boundaries between
An additional relevant topic deserving further investigation the algorithms classes become weak.
is the interdependence of PdM and the dependability of IoT, • Model-Based methods, in which physical models of the
which is defined in [78] as ‘the ability to deliver services that component are used for the estimation of its healthy
can justifiably be trusted’. conditions and the prediction of its degradation. The
Intuitively, IoT communications for PdM have to satisfy strin- benefit of resorting to these models lies in that they
gent requirements in terms of timeliness and correctness, as the can be applied to components for which data from
information they exchange is critical for ensuring an effective abnormal operating conditions are lacking (e.g., safety-
and safe behavior of the monitored components. Hence, the critical systems, capital parts, equipment conservatively
communication network must be engineered to meet stringent maintained, etc.). In these cases, data-driven models can
delay deadlines, be robust to packet losses and, finally, be safe neither diagnose the anomalous behavior of the compo-
and resilient to damages [79]. nent nor predict its failure trajectory. On the contrary,
To this aim, different technologies are currently being devel- physics-based or physics-of-failure models (e.g., [6]) can
oped, especially within the 5G paradigm, which are critically be developed for simulating the degradation mechanisms
reviewed in [79], [80], [81], [82], [83], to cite a few. affecting the component (e.g., [64], [114], [115], [108])
In spite of these advancements, however, a fundamental re- and used for RUL prediction. In the Industry 4.0 era,
search work is still required to include in PdM modeling and these models are at the basis of the development of
analysis the trust to the IoT. In fact, in industrial practice the Digital Twins. However, the development of physics-
dependability of communication is often characterized through based models is not always practicable because it is very
parameters such as packet delivery ratio, outage probability, costly and, also, these models often do not fully take into
Signal to Interference and Noise Ratio (SINR), Bit Error Rate account the effects of the external conditions and rely on
(BER) [84]. Although these metrics are intuitively related to parameters that are difficult to estimate ([116]) .
the conventional understanding of dependable communication, Notice that these models are not relevant for IoT-enabled
nonetheless they are not sufficient to fully characterize the PdM, which rely on data provided by IoT.
capability of IoT for supporting PdM. This is due to the fact
that IoT are extremely complex and distributed Cyber-Physical B. Challenges
systems of systems (CPSoS), with a multitude of intercon-
nections, also with the human environment, under strict legal A virtuous loop of research and industry is sustaining this,
and regulatory constraints ([85], [86]). This means that to whereby research solutions continue to provide opportunities
fully capture the trustworthiness of IoT for PdM applications, of improvement to industry, while industry provides new chal-
we must integrate in PdM modeling several concepts such as lenges to research. Despite the availability of PHM algorithms,
cyber-security, reliability, resilience, etc. ([85]). To the authors’ the companies that want to benefit from Industry 4.0 still
best knowledge, this is an almost unexplored field ([88]), need to trade off the opportunities of PdM against the capital
especially for the 5G connectivity technology, which promises expenditures required to purchase the necessary instrumenta-
to be at once truly ubiquitous, reliable, scalable and cost- tion, software and specialized knowledge. This downside is
efficient. perceived large at the beginning of the development of PdM,
when real data of normal and abnormal equipment behaviors
IV. PHM A LGORITHMS FOR P D M are lacking or scarce, and in case of new systems, when there
is no experience on their operation. This situation can lead the
A. PHM algorithms taxonomy companies to distrust the investment in PdM solutions.
A wide range of methods have been developed for detection, For a systematic and rationale decision making on PdM
diagnostics and prognostics, as extensively discussed in the investment, the actual challenge is to embedd the cost mod-
literature ([9], [10], [89], [90], [91], [92], [93], [94]), and els presented in Sections II and V in adaptive and robust
with many successful applications reported (e.g., see [6] for frameworks for guiding PdM development: these should allow
an overview). updating and adjusting the PHM algorithms for PdM on the
In general, PHM methods can be divided into two main basis of the Knowledge, Information and Data (KID, [11]) that
classes, although hybrid methods exist too ([95]): incrementally become available as the development goes on
• Data-driven methods, which use monitored operational from the design to its operation, which tends to continuously
data related to the component health conditions. These evolve, due to deterioration of components and sensors, main-
Things Journal
TABLE II
PHM ALGORITHMS OUTLOOK
Algorithm Brief Description Pros Cons

Markov Models (MM) A Markov model is a sequence of • Appropriate when degra- • Can be computationally
including: infinitely many states representing the dation can be described expensive and can re-
• Hidden Markov Models component by discrete states; quire a large number of
(HMM); degradation from new to failed. • Simple in the analytical simulations;
• Semi-Markov Models Transition probabilities among the definition and easy to be • Require the definition
(SMM); states depend on the current state and understood even by non- of the degradation states
• Hidden SMM; not on the path followed to reach it. expert analysts; and the estimate of the
([96], [97], [98], [99]). Differently from MM, the SMM transi- • Can rely on a sound lit- transition state probabil-
tions depends also on the sojourn time erature. ities.
in the current state.
HMM assumes the degradation to be
not directly observable.
Artificial Neural Networks ANN consist of processing elements • Provide good functional • ANN require large
(ANN), including: called neurons, which interact with mappings between input amount of training
• Convolutional NN; each other through numerically and output data points in data that have to
• Extreme Learning Ma- weighted connections among the many practical PHM in- be representative of
chines (ELM); input, hidden and output layers. stances ([100]). true data range and
• Radial Basis Networks Training data are used to build a variability [95];
(RBN; regression model by adjusting the • Performance depends
• Recurrent Neural Net- connection weights between neurons also on the capability of
works (RNN); to reduce the errors between the the user to identify the
• Echo State Networks network and the target outputs [107]. optimum setting (i.e.,
(ESN); The trained ANN process new data number of neurons,
• Auto-Encoders; and give an estimate of the expected layers, activation
• Self Organizing Maps output [95]. functions, etc.);
(SOM); RNN and its advanced versions (ESN, • The operating and
• Long Short Term mem- LSTM, etc.) are ANN, whose neurons training processes are
ory (LSTM); contain feedback connections from “black boxes”, as the
([64], [100], [101], [102], [103] the hidden or output layers to the understanding of the
[104], [105], [106], [107], preceding layers. These connections built models, except
[108]) add to the ANN the ability of from qualitative, is hard
processing temporal dependencies to catch ([92]);
between the inputs and the outputs • ANN can have a slow
and, thus, dynamic information. convergence during the
Auto-Eorders are ANN used to training process [89].
learn efficient data codings in an
unsupervised manner.
Statistical techniques, Rely on both the Bayesian and fre- • Rigorous theoretical • Lots of data required for
including: quentist frameworks, thus giving a background; frequentist approaches;
• Principal Components probabilistic interpreation to the re- • Uncertainty on parame- • Bayesian approaches can
Analysis; sults. ters estimation. be computationally ex-
• Regression Models (Lin- pensive.
ear, Logistic, etc.);
• ...
([33], [34], [93], [109], [110]).
Instance-Based methods, Rely on stored data as training set; • Efficient with both small • Parameters tuning
including: when predicting a value of a new in- and large datasets; strongly affects the
• Fuzzy Similarity; stance, they compute its distances from • Can provide real-time performance;
• K-Nearest Neighbors; or similarities to the available training analysis and guarantee a • Heavy memory usage for
• Kernel Machines instances. good generalization per- storing all training in-
(e.g., Support Vector formance; stances;
Machines, Relevance • Can handle non-linear • Risk of overfitting.
Vector Machines, and complex system
Gaussian Fields, Auto modelling;
Associative Kernel • Models are built directly
Regression); from the training in-
stances themselves.
([33], [38], [111], [112],
[113]).
tenance activities, upgrading plan involving the use of new refer to predicting the labels of samples drawn from a target
components and system architectures, and the modifications domain (e.g., system of a fleet working in a new environment),
of the operational and environmental conditions. These mod- given labeled samples drawn from a source domain (e.g., data
ifications of the system behavior, which are typically referred from a system of the same fleet, with longer experience)
to as concept drifts or operation in an Evolving Environment and unlabeled samples drawn from the target domain itself
(EE, [69]), challenge the PHM algorithms development. (i.e., data from the new system). Algorithms for this domain
PHM in EE has been recently addressed by transfer learning adaptation are carefully revised in [117], where relavant chal-
[117] and incremental learning approaches [118]. The former lenges are also outlined. These mainly refer to computational
Things Journal
and analysis burdens, and to reduce the analyst knowledge be implemented in safety critical systems ([131], [132]).
about the specific application of interest that is often required To develop this modeling framework, on the one hand one can
to select an appropriate transformation among many possible build on the model in [59], in which the relationship between
alternatives. the PHM algorithms sustaining PdM and the probability of
Incremental learning approaches can be divided into passive failure has been formally developed. This allows defining the
and active approaches. The former adapt the empirical model values of the thresholds for a set of performance metrics that
every time new batches of data become available. This is time guarantee a desired level of safety, with adequate margins
consuming and not always doable, as it requires the availablity related to the uncertainties.
of labeled time series and empirical model retraining. On the On the other hand, PHM can be embedded in dynamic
contrary, active approaches allow adjusting the models only Probabilistic Risk Assessment (PRA) models (e.g., [133]),
when the occurrence of a concept drift is detected. They are to integrate the dynamic predictions, and their uncertainties,
typically classified into the following categories [119]: with the actions performed by operators and automatic control
• Sequential analysis-based approaches, which analyze the systems. A first attempt is proposed in [134], which, however,
newly acquired signals one by one, until the probability does not consider the dynamic character of predictions.
of observing the subsequence under a new distribution is The capability of modeling the impact of predictions on
significantly larger than that under the original distribu- safety also allows to balance reliability allocation schemes by
tion [120]. installation of PdM capabilities and by redundancy. This topic
• Data distribution-based drift detection approaches, which has been partially addressed in [56], [57], [135], but there is
consider distributions of raw data from two different time- still research work to do for investigating the impact of PdM
windows: a fixed window containing information of the on the optimal reliability allocation for safety.
past time series behavior and a sliding window containing Finally, the impacts of IoT on safety critical aplications is still
the most recent acquired data [121]. an unexplored research area, to the authors best knowledge.
• Learner output-based drift detection approaches, which
are based on the development of a learner (classifier) and B. Decision making for business
the tracking of its error rate fluctuations [122]. A significant part of the value of PdM comes from indirect
A drawback of the application of active approaches in PHM is consequences of the prediction capabilities. For example, the
that the activities of concept drift detection, data labeling, and benefit of PdM for wind farms may come not solely from the
empirical model updating are sequentially and independently obtained increase in availability, but also from the improve-
performed. This requires the use of different algorithms which ment in the logistics for maintenance operations enabled by
exploit the same information, contained in the time series data the knowledge of the component RULs (e.g., [136], [137]).
stream, for different purposes and at different times. In a manufacturing plant, economic benefit from PdM may
To conclude, the adaptivity characteristic of the methodologies come from the warehouse management, which can rely on
and algorithms give the possibility of tracking the development the RUL knowledge to set a just-in-time logistic support that
of the PHM system and the improvement of its performance. reduces the stored spares.
However, this requires computational and analysis burdens. In the car market, the business of PdM relates to the marketing
The challenging issue is on how to simplify and, thus, make opportunities of selling a car with this appealing technology,
faster and cheaper the development of PHM solutions. which provides the driver with the current health state of
the car and the remaining time up to failure (e.g., brake
V. D ECISION MAKING WITH P D M pads consumed). Cross-selling opportunities come from the
Once the PHM algorithms have been developed and their workshop services: the after-sales department can propose a
performance validated, the information about the RUL of the service which directly makes an appointment at the preferred
equipment is exploited for PdM under different perspectives. workshop, which is prepared for receiving the car, for a very
We consider the following three: safety, business, and Opera- fast intervention with discounted spares. This also enhances
tion and Maintenance (O&M). the customer loyalty. Additional benefits come from the con-
trol of the dealers’ operations, which gives the possibility of
A. Decision making for safety both improving the replenishment plans of the spare depots
serving the dealers, and allocating the after-sales budgets to
Intuitively, prediction capabilities can strongly impact on
the dealers based on their actual selling performances. Finally,
safety, as they allow monitoring the risk of failure of the
the prediction capabilities allow proposing a business model of
components giving, thus, the opportunity of preventing failures
selling the run kilometers instead of the car. A similar example
by PdM.
is that of the turbine engines for aircrafts [138], in which
Although many experiences are reported in the literature about
the manufacturer sells the fired hours instead of selling the
the possible applications of PdM to safety critical contexts
turbines, in a win-win setting.
(e.g., nuclear [127], [128], [129], aerosapce [130], [131]), a
structured modeling approach that quantifies the benefit of
PdM for safety is still lacking, as witnessed by the fact that C. Decision making for O&M
safety standards still consider that many enhancement steps To fully exploit the prediction capabilities, the PdM
are necessary to make the PdM technology mature enough to analytics must enter the asset-level management decision
Things Journal
making, which must consider the impact of maintenance [3] R. Drath and A. Horch, ”Industrie 4.0: Hit or Hype?, IEEE Industrial
on logistics, safety, costs, etc. An holistic management of Electronics Magazine, Vol. 8, no 2, pp. 56-58, 2014.
[4] Accessed on Nov. 2019,
the asset with respect to maintenance is also referred to as https://www.gtai.de/GTAI/Content/EN/Invest/ SharedDocs/Downloads
Prescriptive Maintenance [139]. This entails the integration /GTAI/Brochures/Industries/industrie4.0-smart-manufacturing-for-the-
of various asset management and maintenance systems to future-en.pdf
[5] R.Y. Zhong, X. Xun, E. Klotz and S.T. Newman, ”Intelligent Manufac-
prescribe optimal solutions to manage maintenance at the turing in the Context of Industry 4.0: A Review”, Engineering, Vol. 3,
asset level. For example, whilst predictive maintenance might pp. 616-630, 2017.
recommend that the bearings of the application mentioned [6] D. Kwon, M.R. Hodkiewicz, J. Fan, T. Shibutani and M.G. Pecht,
”IoT-Based Prognostics and systems Health Management for Industrial
in Section III be overhauled, a prescriptive view may also Applications”, IEEE Access, Vol. 4, pp. 3659-3670, 2016.
inform the DM that if the equipment loading conditions [7] ”Industrial Internet of Things: Unleashing the potential of
are slowed down, the time to failure can increase up to connected products and services”, World Economic Forum,
Cologny, Switzerland, Technical Report, 2015. [Online]. Available:
the already scheduled maintenance window, also verifying http://reports.weforum.org/industrial-internet-of-things/
whether the planned production requirements can still be met. [8] E. Zio and M. Compare, ”Evaluating maintenance policies by quantitative
To do this, prescriptive systems must be ‘cognitive’, which modeling and analysis”, Reliability Engineering & system Safety, Vol.
109, pp. 53-65, 2013.
means relying on advanced technology at the intersection of [9] M.M. Simões, J.M. Gomes and C.F. Yasin, ”A literature review of
big data, machine learning, and artificial intelligence analytics maintenance performance measurement: A conceptual framework and
([139]). directions for future research”, Journal of Quality and Maintenance
Engineering, Vol. 40, pp. 116–137, 2011.
Some developments are proposed in [140], [141], [142], [143], [10] H. Wang, ”A survey of maintenance policies of deteriorating systems”,
but on specific, scaled-down applications, not transferable to Europena Journal of Operational Research, Vol.139, no. 3, pp. 469–489,
industrial practice. This confirms the need for further research 2002.
[11] E. Zio, ”Some Challenges and Opportunities in Reliability Engineering”,
and development on this. IEEE Transactions on Reliability, Vol. 65, no.4, pp.1769-1782, 2016.
[12] M. Pecht, Prognostics and Health Management of Electronics, John
Wiley & Sons, Encyclopedia of Structural Health Monitoring, 2009.
[13] K. Pipe, ”Practical prognostics for condition based maintenance”, 2008
VI. C ONCLUSIONS International Conference on Prognostics and Health Management, 2008.
[14] A. Rastegari and M. Mobin, ”Maintenance decision making, supported
In the Industry 4.0 era, the available smart and connected de- by computerized maintenance management system”, Proceedings - An-
vices give companies the opportunity of redefining their busi- nual Reliability and Maintainability Symposium, art. no. 7448086, 2016.
nesses by rethinking nearly everything they do. In particular, [15] M. Wienker, K. Henderson and J. Volkerts,”The Computerized Main-
tenance Management system an Essential Tool for World Class Mainte-
the new technologies, including IoT, enable the development of nance”, Procedia Engineering, Vol. 138, pp. 413-420, 2016.
PdM, which is transforming the way of thinking maintenance: [16] N.A.J. Hastings and J.Y.T. Ang, ”Developments in computer-based
from cost to business opportunity. reliability analysis”, Journal of Quality in Maintenance Engineering, Vol.
1, no. 1, pp.69-78, 1995.
IoT-enabled PdM is attracting considerable investment from [17] Italian Ministry of Economy and Finance, Piano
industries and increasing attention by research, as witnessed, Nazionale Impresa 4.0: Risultati 2017-Linee Guida 2018,
for example, by the many initiatives and confrontation fo- http://www.sviluppoeconomico.gov.it/images/stories/documenti/impresa
2040 19 settembre 2017.pdf.
rums established by academy and industry for discussion and [18] https://ec.europa.eu/growth/tools-databases/dem/monitor/sites/default/files
experience-sharing of PdM solutions, in various industrial /DTM Industrie%204.0.pdf
fields. [19] P.C. Evans and M. Annunziata, ”Industrial Internet: Pushing the bound-
aries of minds and machines - Technical Report”. [Online]. Available:
The development for PdM has thus far mainly concerned www.ge.com/docs/chapters/Industrial Internet.pdf, 2012.
hardware and software for remote tracking the health state of [20] J. Siryani, B. Tanju and T.J. Eveleigh,”A Machine Learning Decision-
monitored equipment. Indeed, we have shown that the PdM Support system Improves the Internet of Things’ Smart Meter Opera-
tions”, IEEE Internet of Things Journal, Vol. 4, no. 4, pp. 1056-1066,
value chain is much longer, including many activities upstream 2017.
and downstream the collection of data and the execution of [21] F. Färber et al., ”The SAP HANA database.: An architecture overview”,
the maintenance labour. For the full develoment of PdM and IEEE Data Engineering Bulletin, Vol. 35, no. 1, pp. 28-33, 2012.
[22] L. Winig, ”GE’s big bet on data and analytics”, in Proceedings MIT
its deployment in practice, it is important to build integrated Sloan Management Review, Vol. 57, pp. 1-16, 2016.
cost-benefit models that include the impact of the PdM on the [23] W. Shi, J. Cao, Q. Zhang, Y. Li and L. Xu, ”Edge Computing: Vision and
entire asset management. Challenges”, IEEE Internet of Things Journal, Vol. 3, no. 5, pp. 637-646,
2016.
[24] United States - Department of Defense, MIL-STD 721C, Definitions of
ACKNOWLEDGMENT terms for Reliability and maintainability, 1981.
[25] International Electrotechnical Commission, IEC 60300–3–11, Depend-
The authors thank the anonymous reviewers for their helpful ability management – Part 3–11: Application guide – Reliability centered
comments, which strongly contributed to the improvement of maintenance, Geneva, 2009.
[26] A. Grall, L. Dieulle, C. Berenguer, and M. Roussignol, Continuous-
the work. time predictive-maintenance scheduling for a deteriorating system, IEEE
Transactions on Reliability, Vol. 51, no. 2, pp. 141-150, 2002.
R EFERENCES [27] F.S. Nowlan and H.F. Heap, ”Reliability Centered Maintenance”,
UNITED AIR LINES INC SAN FRANCISCO CA, Rep. ADA066579, 1978.
[1] M. Hermann, T. Pentek, and B. Otto, ”Design Principles for Industrie 4.0 [28] M. Rausand, Reliability Centered Maintenance, Reliability Engineering
Scenarios”, 49th Hawaii International Conference on system Sciences, and system Safety, Vol. 60, pp. 121-132, 1998.
pp. 3927-3937, 2016. [29] M. Rausand and J. Vatn, ”Reliability centered maintenance”, In: Com-
[2] H. Lasi, P. Fettke, H.-G. Kemper, T. Feld and M. Hoffmann, ”Industry plex systems Maintenance Handbook Part B. Springer, London, 2008.
4.0”, Business and Information systems Engineering, Vol. 6, no. 4, pp. [30] Society of Automotive Engineers , SAE JA1012, A Guide to the
239-242, 2014. Reliability-Centered Maintenance (RCM) Standard, 2002.
Things Journal
[31] J. Moubray, Reliability Centered Maintenance, 2nd Ed., Industrial Press, [54] H Raiffa and R. Schlaifer, Applied statistical decision theory, New York,
NY, 1997. NY, John Wiley&Sons, 1961.
[32] F. Backlund and P.A. Akersten, ”RCM introduction: process and re- [55] G. Haddad, P.A. Sandborn and M. Pecht, ”Using maintenance options
quirements management aspects”, Journal of Quality in Maintenance to maximize the benefits of prognostics for wind farms”, Wind Energy,
Engineering, Vol. 9, no. 3, pp. 250-264, 2003. Vol. 17, pp. 775–791, 2014
[33] P. Baraldi, F. Di Maio, D. Genini and E. Zio, ”Comparison of Data- [56] B.D. Youn, C. Hu, and P. Wang, ”Resilience-Driven system Design
Driven Reconstruction Methods for Fault Detection”, IEEE Transactions of Complex Engineered systems”, Journal of Mechanical Design, Vol.
on Reliability, Vol. 64, no. 3, art. no. 7115190, pp. 852-860, 2015. 133.10, art. no. 101011, 2011,
[34] W. Li, T. Shi, G. Liao and S. Yang, ”Feature extraction and classification [57] J.T. Yoon, B.D. Youn, M. Yoo and Y. Kim, ”A newly formulated
of gear faults using principal component analysis”, Journal of Quality in resilience measure that considers false alarms”, Reliability Engineering
Maintenance Engineering, Vol. 9, no. 2, pp.132-143, 2003. & system Safety, Vol. 167, pp. 417-427, 2017.
[35] J. Ma and J. Jiang, ”Applications of fault detection and diagnosis [58] J.T. Yoon, B.D. Youn, M. Yoo, Y. Kim and S. Kim, ”Life-cycle
methods in nuclear power plants: A review”, Progress in Nuclear Energy, maintenance cost analysis framework considering time-dependent false
Vol. 53, no. 3, pp. 255-266, 2011. and missed alarms for fault diagnosis”, Reliability Engineering and system
[36] J.W. Hines, D.J. Wrest and R.E. Uhrig, ”Plant wide sensor calibration Safety, Vol. 184, pp. 181-192, 2019.
monitoring”, Proceedings of the 1996 IEEE International Symposium on [59] M. Compare, L. Bellani and E. Zio, ”Reliability model of a component
Intelligent Control, pp. 378-383, 1996. equipped with PHM capabilities”, Reliability Engineering and system
[37] J.W. Hines and R. Seibert, Technical Review of On-Line Monitoring Safety, Vol. 168, pp. 4-11, 2017.
Techniques for Performance Assessment, Vol. 1, State-of-the-Art, 2006. [60] M. Compare, L. Bellani and E. Zio,, ”Availability model of a PHM-
[38] P. Baraldi, F. Cannarile, F. Di Maio and E. Zio, ”Hierarchical k- equipped component”, IEEE Transactions on Reliability, Vol. 66, no. 2,
nearest neighbours classification and binary differential evolution for fault pp. 487-501, 2017.
diagnostics of automotive bearings operating under variable conditions”, [61] A. Saxena, J. Celaya, B. Saha, S. Saha and K. Goebel, ”Evaluating
Engineering Applications of Artificial Intelligence, Vol. 56, pp. 1-13, algorithm performance metrics tailored for prognostics”, IEEE, Aerospace
2016. conference, pp. 1-13, 2009.
[39] T.C. Akinci, N. Ekren, S. Seker and S. Yildirim, ”Continuous wavelet [62] F. Cannarile, P. Baraldi, M. Compare, D. Borghi, L. Capelli, M.
transform for ferroresonance detection in power systems”, Electrical Cocconcelli, A. Lahrache and E. Zio, ”An unsupervised clustering method
Power and Energy systems, Vol. 44, pp. 403-409, 2013. for assessing the degradation state of cutting tools used in the packaging
[40] P. Jayaswal, S.N. Verma and A.K. Wadhwani, ”Application of ANN, industry”, The 2nd International Conference on Engineering Sciences and
Fuzzy Logic and Wavelet Transform in machine fault diagnosis using vi- Technologies, 2017.
bration signal analysis”, Journal of Quality in Maintenance Engineering, [63] A. Hess, G. Calvello and P. Frith, ”Challenges, Issues, and Lessons
Vol. 16, no. 2, pp.190-213, 2010. Learned Chasing the “Big P”: Real Predictive Prognostics Part 1”, IEEE
[41] P. Baraldi, G. Bonfanti and E. Zio, ”Differential evolution-based multi- Aerospace Conference, 2005.
objective optimization for the definition of a health indicator for fault [64] M. Rigamonti, P. Baraldi, E. Zio, K. Goebel and S. Poll, ”Ensemble
diagnostics and prognostics”, Mechanical systems and Signal Processing, of optimized echo state networks for remaining useful life prediction”,
Vol. 102, pp. 382-400, 2018. Neurocomputing, Vol. 281, pp. 121-138, 2018.
[42] W. Wang and M. Pecht, ”Economic Analysis of Canary-Based Prognos-
[65] How KOMTRAX Works, [online] Available:
tics and Health Management”, IEEE Transactions on Industrial Electron-
http://www.komatsu.com/CompanyInfo/profile/product supports/, 2015.
ics, Vol.58, no. 7, pp. 3077-3089, 2011.
[66] Remote Monitoring and Diagnostics, [online] Available:
[43] L. Goodman, S. Wood and A. Turner, ”Return-on-Investment (ROI)
https://powergen.gepower.com/content/dam/gepower-pgdp/global/en
for Electronic Prognostics in Mil/Aero systems”, IEEE, Autotestcon,
US/documents/service/gas%20turbine%20services/monitoring-
Orlando, FL., 2005.
diagnostics-fact-sheet.pdf.
[44] S. Wood and D. Goodman, ”Return-on-Investment (ROI) for Electronic
Prognostics in High Reliability Telecom Applications”, Annual Interna- [67] Y. Hu, P. Baraldi, F. Di Maio and E. Zio, ”A systematic Semi-Supervised
tional Telecommunications Energy Conference, 2006. Self-adaptable Fault Diagnostics approach in an evolving environment”,
[45] K. Feldman, T. Jazouli and P.A. Sandborn, ”A methodology for deter- Mechanical systems and Signal Processing, Vol. 88, pp. 413-427, 2018.
mining the return on investment associated with prognostics and health [68] A. Tsymbal, ”The problem of concept drift: definitions and related
management”, IEEE Transactions on Reliability, Vol. 58, no. 2, pp. 305- work”, Computer Science Department, Trinity College Dublin, 2004.
316, 2009. [69] G. Ditzler, M. Roveri, C. Alippi and R. Polikar, ”Learning in nonstation-
[46] M.-H. Chang, M. Pecht and W.K.C. Yung, ”Return on investment asso- ary environments: A survey”, IEEE Computational Intelligence Magazine,
ciated with PHM applied to an LED lighting system”, PHM 2013 - 2013 Vol. 10, no. 4, pp. 12-25, 2015.
IEEE International Conference on Prognostics and Health Management, [70] A. Salo, J. Keisler and A. Morton Eds., ”Portfolio Decision Analysis
Conference Proceedings, art. no. 6621434, 2013. – Improved Methods for Resource Allocation”, International Series in
[47] T Formica and M. Pecht, ”Return on investment analysis and simulation Operations Research & Management Science, Vol. 162, Springer-Verlag,
of a 9.12 kilowatt (kW) solar photovoltaic system”, Solar Energy, Vol. 2011.
144, pp. 629–634, 2017. [71] J. Reifman, ”Survey of artificial intelligence methods for detection and
[48] J. Banks and J. Merenich, ”Cost Benefit Analysis for Asset Health identification of component faults in nuclear power plants”, Nuclear
Management Technology”, 2007 Annual Reliability and Maintainability Technology, Vol. 119, pp. 76–97, 1997.
Symposium, Conference Proceedings, pp. 95-100, 2007. [72] C. Malings and M. Pozzi, ”Value-of-information in spatio-temporal
[49] J. Banks, K. Reichard, E. Crow and K. Nickell, ”How engineers can systems: Sensor placement and scheduling”, Reliability Engineering &
conduct cost-benefit analysis for PHM systems”, IEEE Aerospace and system Safety, Vol. 172, pp. 45-57, 2018.
Electronic systems Magazine, Vol. 24, no. 3, pp. 22-30, 2009. [73] C. Malings, Optimal Sensor Placement for Infrastructure
[50] J.E. Dzakowic and G.S. Valentine, ”Advanced Techniques For The system Monitoring using Probabilistic Graphical Models and
Verification And Validation Of Prognostics & Health Management Capa- Value of Information, PhD thesis at Carnegie Mellon University,
bilities”, Machinery Failure Prevention Technologies (MFPT 60), Virginia http://repository.cmu.edu/cgi/viewcontent.cgi?article=1908
Beach, VA, 2007. &context=dissertations.
[51] M. Pozzi, D. Zonta, W. Wang and G. Chen, ”A framework for evaluating [74] K. Dykes and R. de Neufville, ”Real Options for a Wind Farm in
the impact of structural health monitoring on bridge management”, Wapakoneta, Ohio: Incorporating Uncertainty into Economic Feasibility
Proc. of 5th International Conf. on Bridge Maintenance, Safety and Studies for Community Wind”, World Wind Energy Conference 2008,
Management (IABMAS2010), Philadelphia, July 11–15, 2010, pp. on CD. Kingston, Ontario June 24-26, 2008.
[52] M. Memarzadeh and M. Pozzi, ”Value of information in sequential deci- [75] J. Liesiö, P. Mild and A. Salo, ”Robust portfolio modeling with incom-
sion making: Component inspection, permanent monitoring and system- plete cost information and project interdependencies”, European Journal
level scheduling”, Reliability Engineering and system Safety, Vol. 154, of Operational Research, Vol. 190, pp. 679-695, 2008.
pp. 37–151, 2016. [76] A. Mancuso, M. Compare, A. Salo and E. Zio, ”Portfolio optimization
[53] M. Memarzadeh, M. Pozzi, and J.Z. Kolter, ”Optimal planning and of safety measures for reducing risks in nuclear systems”, Reliability
learning in uncertain environments for the management of wind farms”, Engineering and system Safety, Vol. 167, pp. 20-29, 2017.
ASCE Journal of Computing in Civil Engineering, Vol. 29, no. 5, art. no. [77] https://www2.deloitte.com/content/dam/Deloitte/us/Documents/process-
04014076, 2015. and-operations/us-cons-predictive-maintenance.pdf.
Things Journal
[78] A. Avizienis, J.-C. Laprie, B. Randell and C. Landwehr, ”Basic concepts [101] N. Gebraeel, M.A. Lawley, R. Liu and V. Parmeshwaran, ”Residual life
and taxonomy of dependable and secure computing”, IEEE Transactions predictions from vibration-based degradation signals: A neural network
on Dependable and Secure Computing, Vol. 1, no. 1, pp. 11-33, 2004. approach”, IEEE Transactions on Industrial Electronics, Vol. 51, no. 3,
[79] M.R. Palattella, M. Dohler, A. Grieco, G. Rizzo, J. Torsner, T. Engel and pp. 694-700, 2004.
L. Ladid, ”Internet of Things in the 5G Era: Enablers, Architecture, and [102] S.-J. Wu, N. Gebraeel, M.A. Lawley and Y. Yih, ”A neural network
Business Models”, IEEE Journal on Selected Areas in Communications, integrated decision support system for condition-based optimal predictive
Vol. 34, no. 3, pp. 510-527, 2016. maintenance policy”, IEEE Transactions on systems, Man, and Cybernet-
[80] Z. Ma, M. Xiao, Y. Xiao, Z. Pang, H.V. Poor, and B. Vucetic, ”High- ics Part A:systems and Humans, Vol. 37, no. 2, pp. 226-236, 2007.
reliability and Low-latency Wireless Communication for Internet of [103] H UANG , R., X I , L., L I , X., R ICHARD L IU , C., Q IU , H., L EE ,
Things: Challenges, Fundamentals and Enabling Technologies”, IEEE J. (2007), ”Residual life predictions for ball bearings based on self-
Internet of Things Journal, Early access, 2019. organizing map and back propagation neural network methods”, Mechan-
[81] G. Arfaoui, J.M.S. Vilchez and J. Wary, ”Security and Resilience ical systems and Signal Processing, Vol. 21(1), pp. 193-207.
in 5G: Current Challenges and Future Directions”, 2017 IEEE Trust- [104] R.M. Hasani, G. Wang and R. Grosu, An ”Automated Auto-encoder
com/BigDataSE/ICESS, pp. 1010-1015, 2017. Correlation-based Health-Monitoring and Prognostic Method for Machine
[82] M. Emara, M.C. Filippou and I. Karls, ”Availability and Reliability of Bearings”. arXiv preprint arXiv:1703.06272, 2017.
Wireless Links in 5G systems: A Space-Time Approach”, 2018 IEEE [105] S. Zheng, K. Ristovski, A. Faraha and C. Gupta, ”Long Short-
Globecom Workshops (GC Wkshps), pp. 1-6, 2018. Term Memory Network for Remaining Useful Life estimation”, 2017
[83] C.J.A. Tokognon, B. Gao, G.Y. Tian and Y. Yan, ”Structural Health IEEE International Conference on Prognostics and Health Management,
Monitoring Framework Based on Internet of Things: A Survey”, IEEE ICPHM 2017, pp. 88-95, art. no. 7998311, 2017.
Internet of Things Journal, Vol. 4, No. 3, 2017. [106] F.O. Heimes, ”Recurrent neural networks for remaining useful life
[84] H.V.K. Mendis and F.Y. Li, ”Achieving Ultra Reliable Communication estimation”, 2008 International Conference on Prognostics and Health
in 5G Networks: A Dependability Perspective Availability Analysis in the Management, PHM 2008, art. no. 4711422, 2008.
Space Domain”, IEEE Communications Letters, Vol. 21, pp. 2057-2060, [107] M.S. Kan, A.C.C. Tan and J. Mathew, “A review on prognostic tech-
2017. niques for non-stationary and non-linear rotating systems”, Mechanical
[85] F. Coallier, ”A system of systems engineering perspective on IoT systems and Signal Processing, Vol. 62, pp. 1-20, 2015.
trustworthiness”, 2018 13th Annual Conference on system of systems [108] M. Rigamonti, P. Baraldi, E. Zio, A. Alessi, D. Astigarraga and
Engineering (SoSE), pp. 89-91, 2018. A. Galarza, ”Identification of the degradation state for condition-based
[86] R. Rogers, E. Apeh and C.J. Richardson, ”Resilience of the Internet maintenance of insulated gate bipolar transistors: A self-organizing map
of Things (IoT) from an Information Assurance (IA) perspective”, 2016 approach”, Microelectronics Reliability, Vol. 60, pp. 48-61, 2016.
10th International Conference on Software, Knowledge, Information [109] Y.G. Li and P. Nilkitsaranant, ”Gas turbine performance prognostic for
Management Applications (SKIMA), pp. 110-115, 2016. condition-based maintenance”, Applied Energy, Vol. 86, pp. 2152-2161,
[87] J. Ali-Tolppa, S. Kocsis, B. Schultz, L. Bodrog and M. Kajo, ”Self- 2009.
Healing and Resilience in Future 5G Cognitive Autonomous Networks”, [110] M. Compare, P. Baraldi, I. Bani,E. Zio and D. Mc Donnell, ”Develop-
2018 ITU Kaleidoscope: Machine Learning for a 5G Future (ITU K), pp. ment of a Bayesian multi-state degradation model for up-to-date reliability
1-8, 2018. estimations of working industrial components”, Reliability Engineering
[88] K.A. Delic, ”On Resilience of IoT systems”, Ubiquity, an ACM publi- and system Safety, Vol 166, pp. 25-40, 2017.
cation. Ubiquity Symposium, The Internet of Things, 2016. [111] E. Zio and F. Di Maio, ”A data-driven fuzzy approach for predicting the
[89] A.K.S. Jardine, D. Lin and D. Banjevic, ”A review on machinery remaining useful life in dynamic failure scenarios of a nuclear system”,
diagnostics and prognostics implementing condition-based maintenance”, Reliability Engineering & system Safety, Vol. 95, no. 1, pp. 49-57, 2010.
Mechanical systems and Signal Processing, Vol. 20, pp. 1483-1510, 2006.
[112] F. Di Maio, K.L. Tsui and E. Zio, ”Combining relevance vector
[90] O.E. Dragomir, R. Gouriveau, F. Dragomir, E. Minca and N. Zerhouni,
machines and exponential regression for bearing residual life estimation”,
”Review of prognostic problem in condition-based maintenance”, IEEE
Mechanical systems and Signal Processing, Vol. 31, pp. 405-427, 2012.
Control systems Society. European Control Conference, ECC’09, Aug
[113] E. Zio and F. Di Maio, ”A fuzzy similarity-based method for failure
2009, Budapest, Hungary, pp.1585-1592, 2009.
detection and recovery time estimation”, International Journal of Per-
[91] E. Zio, ”Prognostics and Health Management of Industrial Equipment”,
formability Engineering, Vol. 6, no. 5, 407, 2010.
Diagnostics and Prognostics of Engineering systems: Methods and Tech-
niques, Seifedine Kadry (Ed.), pp. 333-356, 2013. [114] P. Baraldi, F. Cadini, F. Mangili and E. Zio, ”Model-based and data-
[92] J. Lee, F. Wu, W. Zhao, M. Ghaffari, L. Liao and D. Siegel, “Prognostics driven prognostics under different available information”, Probabilistic
and health management design for rotary machinery systems - Reviews, Engineering Mechanics, Vol. 32, pp. 66-79, 2013.
methodology and applications”, Mechanical systems and Signal Process- [115] M. Corbetta, C. Sbarufatti, M. Giglio, A. Saxena and K. Goebel, ”A
ing, Vol. 42, no. 1-2, pp. 314-334, 2014. Bayesian framework for fatigue life prediction of composite laminates
[93] X.-S. Si, W. Wang, C.-H. Hu and D.-H. Zhou, “Remaining useful life under co-existing matrix cracks and delamination”, Composite Structures,
estimation - A review on the statistical data driven approaches”, European Vol. 187, pp. 58-70, 2018.
Journal of Operational Research, Vol. 213, no. 1, pp. 1-14, 2011. [116] Y. Hu, P. Baraldi, F. Di Maio and E. Zio, ”A particle filtering and ker-
[94] W. Zhang, D. Yang and H. Wang, ”Data-Driven Methods for Predictive nel smoothing-based approach for new design component prognostics”,
Maintenance of Industrial Equipment: A Survey”, IEEE Systems Journal, Reliability Engineering and system Safety, Vol. 134, pp. 19-31, 2014.
Vol. 13, no. 3, pp. 2213-2227, 2019. [117] W.M. Kouw and M. Loog, ”A review of single-source unsupervised
[95] J.Z. Sikorska, M. Hodkiewicz and L. Ma, “Prognostic modelling options domain adaptation”, CoRR, Vol. abs/1901.05335, arXiv 1901.05335,
for remaining useful life estimation by industry”, Mechanical systems and 2019.
Signal Processing, Vol. 25, no. 5, pp. 1803-1836, 2011. [118] G. Ditzler and R. Polikar, “Incremental learning of concept drift from
[96] R. Moghaddass and M.J. Zuo, ”An integrated framework for online di- streaming imbalanced data,” IEEE Transactions on Knowledge and Data
agnostic and prognostic health monitoring using a multistate deterioration Engineering, Vol. 25, no. 10, pp. 2283 - 2301, 2013.
process”, Reliability Engineering and system Safety, Vol. 124, pp. 92-104, [119] J. Gama, I. ŽliobaitĖ, A. Bifet, M. Pechenizkiy and A. Bouchachia,
2014. “A survey on concept drift adaptation,” ACM Computing Surveys, Vol.
[97] L.R. Rabiner, ”A Tutorial on Hidden Markov Models and Selected 46(4), Article No. 442014, 2014.
Applications in Speech Recognition”, Proceedings of the IEEE, Vol. 77, [120] A. Wald, “Sequential tests of statistical hypotheses,” Annals of Math-
no. 2, pp. 257-286, 1989. ematical Statistics, Vol. 16(2), pp. 117-186, 1945.
[98] R. Dawid, D. McMillan and M. Revie, ”Review of Markov models for [121] J.P. Patist, “Optimal window change detection,” in Proceedings - IEEE
maintenance optimization in the context of offshore wind”, Proceedings International Conference on Data Mining, ICDM, 2007.
of the Annual Conference of the Prognostics and Health Management [122] J. Gama, P. Medas, G. Castillo and P. Rodrigues, “Learning with Drift
Society, PHM, pp. 269-279, 2015. Detection”, Advances in Artificial Intelligence – SBIA 2004, pp. 286–295,
[99] P. Baruah and R.B. Chinnam, ”HMMs for diagnostics and prognostics 2004.
in machining processes”, International Journal of Production Research, [123] F. Cannarile, M. Compare and E. Zio, ”A fault diagnostic tool based
Vol. 43, no. 6, pp. 1275-1293, 2005. on a first principle model simulator”, Lecture Notes in Computer Science,
[100] A. Malhi, R. Yan and R.X. Gao, ”Prognosis of defect propagation based 2017.
on recurrent neural networks”, IEEE Transactions on Instrumentation and [124] Z. Zeng, F. Di Maio, E. Zio and R. Kang, ”A hierarchical decision-
Measurement, Vol. 60, no. 3, art. no. 5710193, pp. 703-711, 2011. making framework for the assessment of the prediction capability of prog-
Things Journal
nostic methods”, Proceedings of the Institution of Mechanical Engineers, Michele Compare Michele Compare is the CEO
Part O: Journal of Risk and Reliability, Vol. 231, no. 1, pp. 36-52, 2017. of Aramis, Milano, Italy, and the principal inves-
[125] A. Nair, S. Boring and J. Coble, ”High accuracy signal validation tigator of the projects on reliability, availability
framework for sensor calibration assessment in NPPs”, Transactions of and maintenance analysis, optimization and decision
the American Nuclear Society, Vol. 115, pp. 348-351, 2016. making. He received the M.Sc. degree in mechanical
[126] P. Baraldi, G. Gola, E. Zio, D. Roverso and M. Hoffmann, ”A engineering cum laude from University of Naples
randomized model ensemble approach for reconstructing signals from Federico II, in 2003, the PhD in nuclear engineering
faulty sensors”, Expert systems with Applications, Vol. 38, no. 8, pp. cum laude from Politecnico di Milano, in 2011. He
9211-9224, 2011. has been research assistant at Politecnico di Milano
[127] P. Baraldi, F. Mangili and E. Zio, ”A prognostics approach to nuclear since 2011. He worked as RAMS engineer and risk
component degradation modeling based on Gaussian Process Regression”, manager, and is coauthor of more than 60 papers on
Progress in Nuclear Energy, Vol. 78, pp. 141-154, 2015. international journals and conferences.
[128] Z. Welz, J. Coble, B. Upadhyaya and J.W. Hines, ”Maintenance-based
prognostics of nuclear plant equipment for long-term operation”, Nuclear
Engineering and Technology, Vol. 49, no. 5, pp. 914-919, 2017.
[129] J. Coble, P. Ramuhalli, L. Bond, J.W. Hines and B. Upadhyaya, ”A re-
view of prognostics and health management applications in nuclear power
plants”, International Journal of Prognostics and Health Management,
Vol. 6 (SP3), pp. 1-22, 2015
[130] M. Daigle, I. Roychoudhury, L. Spirkovska, K. Goebel, S. Sankarara-
man, J. Ossenfort and C. Kulkarni, ”Real-time prediction of safety mar-
gins in the national airspace”, 17th AIAA Aviation Technology, Integration,
and Operations Conference, 2017. Piero Baraldi Piero Baraldi received the Ph.D.
[131] I. Roychoudhury, L. Spirkovska, M. Daigle, E. Balaban, S. Sankarara- degree in nuclear engineering from Politecnico di
man, C. Kulkarni, S. Poll and K. Goebel, ”Predicting real-time safety of Milano, Milan, Italy, in 2006. He is currently an
the national airspace system”, AIAA Infotech @ Aerospace Conference, Associate Professor of nuclear engineering with the
2016. Department of Energy, Politecnico di Milano. He
[132] J.B. Coble, P. Ramuhalli, L.J. Bond, J.W. Hines and B.R. Upadhyaya, was a Technical Committee Co-Chair of the 2014
”Prognostics and Health Management in Nuclear Power Plants: A Review European Safety and Reliability Conference and
of Technologies and Applications”, U.S. Department of Energy, PNNL- Technical Program Chair of the 2013 Prognostics
21515, Washington D.C. and system Health Management Conference. He
[133] T. Aldemir, ”A survey of dynamic methodologies for probabilistic is a co-author of two books and more than 140
safety assessment of nuclear power plants”, Annals of Nuclear Energy, papers on international journals and proceedings of
Vol. 52, pp. 113-124, 2013. international conferences
[134] H. Kim, S.-H. Lee, J.-S. Park, H. Kim, Y.-S. Chang and G. Heo ,
”Reliability data update using condition monitoring and prognostics in
probabilistic safety assessment”, Nuclear Engineering and Technology,
Vol. 47, no. 2, pp. 204-211, 2015.
[135] M. Compare, L. Bellani and E. Zio, ”Optimal allocation of Prognostics
and Health Management capabilities to improve the reliability of a power
transmission network”, Reliability Engineering and system Safety, Vol.
184, pp. 164-180, 2018.
[136] C. Gundegjerde, I.B. Halvorsen, E.E. Halvorsen-Weare, L.M. Hvattum
and L.M. Nonås, ”A stochastic fleet size and mix model for maintenance
operations at offshore wind farms”, Transportation Research Part C:
Emerging Technologies, Vol. 52, pp. 74-92, 2015. Enrico Zio Enrico Zio received the M.Sc. degree in
[137] C.A. Irawan, D. Ouelhadj, D. Jones, M. Stålhane and I.B. Sperstad, nuclear engineering from the Politecnico di Milano,
”Optimisation of maintenance routing and scheduling for offshore wind in 1991, the M.Sc. degree in mechanical engineering
farms, European Journal of Operational Research, Vol. 256, no. 1, pp. from UCLA, in 1995, the Ph.D. degree in nuclear
76-89, 2017. engineering from the Politecnico di Milano, in 1996,
[138] N. Waters, ”Engine Health Management, in Proceedings Ingenia, pp. and the Ph.D. degree in probabilistic risk assessment
37-42, 2009. from MIT, in 1998. He is currently a Full Professor
[139] A. Goyal, E. Aprilia, G. Janssen, Y. Kim, T. Kumar, R. Mueller, D. with the Centre for Research on Risk and Crises
Phan, A. Raman, J. Schuddebeurs, J. Xiong, and R. Zhang, ”Asset health (CRC), Ecole de Mines, ParisTech, PSL University,
management using predictive and prescriptive analytics for the electric France, a Full Professor and the President of the
power grid”, IBM Journal of Research and Development, Vol. 4, pp. 4- Alumni Association, Politecnico di Milano, Italy, an
14, 2016. Eminent Scholar with Kyung Hee University, South Korea, a Distinguished
[140] M. Compare, P. Marelli, P. Baraldi and E. Zio, ”A Markov decision Guest Professor with Tsinghua University, Beijing, China, an Adjunct Profes-
process framework for optimal operation of monitored multi-state sys- sor with the City University of Hong Kong, Beihang University, and Wuhan
tems”, Proceedings of the Institution of Mechanical Engineers, Part O: University, China, and the Co-Director of the Center for REliability and
Journal of Risk and Reliability, Vol. 232, no. 6, pp. 677-689, 2019. Safety of Critical Infrastructures (CRESCI) and the Sino-French Laboratory
[141] R. Rocchetta, L. Bellani, M. Compare, E. Zio and E. Patelli, ”A of Risk Science and Engineering (RISE), Beihang University, Beijing, China.
reinforcement learning framework for optimal operation and maintenance He has authored or coauthored seven books and more than 500 papers in
of power grids”, Applied Energy, pp. 291-301, 2019. international journals. His research interests include modeling of the failure-
[142] S. Barde, S. Yacout and S. Shin, ”Optimal preventive maintenance repair-maintenance behavior of components and complex systems for the
policy based on reinforcement learning of a fleet of military trucks”, analysis of their reliability, maintainability, prognostics, safety, vulnerability,
Journal of Intelligent Manufacturing, pp. 1-15, 2016. resilience, and security characteristics, and the development and use of Monte
[143] N. Aissani, B. Beldjilali and D. Trentesaux, ”Dynamic scheduling of Carlo simulation methods, artificial techniques, and optimization heuristics.
maintenance tasks in the petroleum industry: A reinforcement approach”, He is the Chairman and Co-Chairman of several international conferences, an
Engineering Applications of Artificial Intelligence, Vol. 22, no. 7, pp. Associate Editor of several international journals, and a Referee of more than
1089-1103, 2009. 20 journals.
[144] L. Bellani, M. Compare, P. Baraldi and E. Zio, ”Towards Developing
a Novel Framework for Practical PHM: a Sequential Decision Problem
solved by Reinforcement Learning and Artificial Neural Networks”,
accepted with minor comments on International Journal of Prognostics
and Health Management, 2019.

Challenges To IoT-enabled Predictive Maintenance

Uploaded by

Copyright:

Available Formats

Challenges To IoT-enabled Predictive Maintenance

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Challenges To IoT-enabled Predictive Maintenance

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Challenges to IoT-enabled Predictive Maintenance

Fig. 1. PdM development activities

Fig. 2. scheme of a production process in the automotive industry

RoI TV Cost VoI Resilience

good sense, nontheless a general framework for optimizing

should be extended to include costs and benefits related to

Algorithm Brief Description Pros Cons

You might also like