Journal of Intelligent Manufacturing (2023) 34:107–121

https://doi.org/10.1007/s10845-022-01960-x

From knowledge-based to big data analytic model: a novel IoT

and machine learning based decision support system for predictive
maintenance in Industry 4.0
Riccardo Rosati1 · Luca Romeo1,2 · Gianalberto Cecchini3 · Flavio Tonetto3 · Paolo Viti4 · Adriano Mancini1 ·
Emanuele Frontoni1

Received: 10 September 2021 / Accepted: 30 April 2022 / Published online: 24 May 2022
© The Author(s) 2022, corrected publication 2022

Abstract
The Internet of Things (IoT), Big Data and Machine Learning (ML) may represent the foundations for implementing the
concept of intelligent production, smart products, services, and predictive maintenance (PdM). The majority of the state-of-
the-art ML approaches for PdM use different condition monitoring data (e.g. vibrations, currents, temperature, etc.) and run
to failure data for predicting the Remaining Useful Lifetime of components. However, the annotation of the component wear
is not always easily identifiable, thus leading to the open issue of obtaining quality labeled data and interpreting it. This paper
aims to introduce and test a Decision Support System (DSS) for solving a PdM task by overcoming the above-mentioned
challenge while focusing on a real industrial use case, which includes advanced processing and measuring machines. In
particular, the proposed DSS is comprised of the following cornerstones: data collection, feature extraction, predictive model,
cloud storage, and data analysis. Differently from the related literature, our novel approach is based on a feature extraction
strategy and ML prediction model powered by specific topics collected on the lower and upper levels of the production system.
Compared with respect to other state-of-the-art ML models, the experimental results demonstrated how our approach is the
best trade-off between predictive performance (MAE: 0.089, MSE: 0.018, R 2 : 0.868), computation effort (average latency of
2.353 s for learning from 400 new samples), and interpretability for the prediction of processing quality. These peculiarities,
together with the integration of our ML approach into the proposed cloud-based architecture, allow the optimization of the
machining quality processes by directly supporting the maintainer/operator. These advantages may impact to the optimization
of maintenance schedules and to get real-time warnings about operational risks by enabling manufacturers to reduce service
costs by maximizing uptime and improving productivity.

Keywords IoT · Decision support system · Predictive maintenance · Machine learning · Random Forest

Introduction uitous scenario of Industry 4.0. These technologies lay

the foundation for implementing the concept of intelligent
The Internet of Things (IoT) and cyber-physical system production, smart products, and services. A realistic applica-
(CPS) represent two fundamental pillars within the ubiq- tion of these technologies is predictive maintenance (PdM),
which aims to address and diagnose in advance maintenance
B Riccardo Rosati issues in order to minimize downtime and costs associated
[email protected] with it. Until now, factory managers and machine opera-
1 Department of Information Engineering, Marche Polytechnic tors have carried out scheduled maintenance and regularly
University, Via Brecce Bianche 12, Ancona 60131, Italy repaired machine parts to avoid downtime. In addition to
2 Computational Statistics and Machine Learning, Istituto consuming unnecessary resources and causing productivity
Italiano di Tecnologia, Via Enrico Melen, 83, 16152 Genoa, losses, half of all preventive maintenance activities may be
Italy ineffective. Accordingly, the lack of maintenance will lead
3 Sinergia Consulenze Srl, Viale Goffredo Mameli, 44, 61121 to more emergency breakdowns and downtime in production
Pesaro, Italy lines which affect the production capability by increasing
4 Benelli Armi Spa, Via della Stazione, 50, 61029 Urbino, Italy operational cost, decreasing production rate, and reducing

123
108 Journal of Intelligent Manufacturing (2023) 34:107–121

the profit margin Burhanuddin et al. (2011). The downtime and represents a key issue for designing ML or Deep Learn-
in production lines has more negative effects than the cost of ing (DL) approaches for accurately solving a PdM task.
repairing the failures. For example, if a filling machine fails in
a production line, the end products will spill over Burhanud- Main contributions
din et al. (2011), thus leading to labour safety issues and
business losses. It is therefore not surprising that PdM has The main contributions of this work to the field of novel
quickly established itself as an industrial 4.0 use case and applications and manufacturing support in Industry 4.0 are
IoT-enabled PdM is attracting considerable investment and related to Big Data registration, processes and analyses. In
attention from industries and research Compare et al. (2019). particular, this paper aims to introduce a DSS for solving
In this scenario, maintenance decision support systems a PdM task by overcoming the above-mentioned challenges
(DSS) empowered by IoT, Big Data and Machine Learn- while focusing on a real industrial use case. The selected
ing (ML) assume a salient role to ensure maintainability PdM task originated from a specific company’s demand 1 and
and reliability of equipment in industries by transforming the experimental setup (see Fig. 1) reflects a real industrial
large datasets into knowledge and actionable intelligence use case which includes advanced processing and measuring
Ayvaz and Alpay (2021); Chen et al. (2020); McArthur et al. machines.
(2018); Schmitt et al. (2020). Implementing industrial DSS to The proposed DSS comprised of five cornerstones by
monitor the health of industrial processes, optimize mainte- combining fundamental aspects of IoT, Big Data and ML:
nance schedules and get real-time warnings about operational data collection, feature extraction, predictive model, cloud
risks enables manufacturers to reduce service costs, maxi- storage, and data analytics. Differently from the related lit-
mize uptime and improve productivity Schwendemann et al. erature, our novel approach is based on a feature extraction
(2021). strategy and ML prediction model powered by specific top-
However, the application of PdM in production envi- ics2 published in the MQTT broker and collected on the lower
ronments involves overcoming several unsolved challenges. and upper levels of the production system. In particular, lower
These challenges include the need to aggregate several het- level topics are used to extract salient KPI predictors that
erogeneous data gathered from sensors dealing with dynamic are fed into a ML model while the upper-level topics reflect
operating environments and the necessity of context-aware the ground-truth variables that are closely related to process-
information targeted to a specific industrial domain Compare ing quality and thus to productivity losses and maintenance
et al. (2019); Dalzochio et al. (2020). Therefore, one of the issues.
main challenges in this scenario is the difficulty to retrieve Comparing the proposed ML approach with respect to
labeled training data of failure (i.e. annotated failure-related other state-of-the-art ML methods, the experimental results
datasets). This kind of data is relevant for training supervised demonstrated how this strategy allows improving the pre-
ML models to accurately and promptly identify problems and dictive performance (in terms of MAE, MSE, and R 2 ) and at
alarms. the same time the interpretability of the PdM task. Moreover,
the lower computation effort allows to easily retrain the ML
model also in the cloud environment.
Challenges and limitation of state-of-the-art The overall IoT-based sensing strategies and cloud stor-
age architecture allow automatizing the annotation procedure
The basic idea behind PdM is to extrapolate/predict knowl- by collecting a huge annotated dataset for training the ML
edge to determine the Remaining Useful Lifetime (RUL) model. On the other hand, the cloud storage framework lays
of components. The majority of the state-of-the-art ML the foundation for integrating an incremental learning strat-
approaches for PdM (as reported in “Related work” section egy Rebuffi et al. (2017), where the ML model is continuously
use different condition monitoring data (e.g. vibrations, cur- updated with new data, and it provides the data analytics for
rents, temperature, etc.) and run to failure data for predicting promptly monitoring the maintenance procedures.
the RUL. However, the annotation of the component wear The paper is organised as follows. In “Related work” sec-
is not always easily identifiable and traced across different tion, we provide a review as well as the main differences with
production cycles and operating conditions. Thus, the open respect to state-of-the-art approaches in the field of PdM and
issues include the difficulty of obtaining quality labeled data
1
and interpreting it Vollert et al. (2021). A severe portion of Benelli Armi spa.
available data has no annotations, presents missing values, 2 Each client that publishes a message to the MQTT broker, includes
and is poorly structured. This fact leads to the high demand a topic into the message. The topic is the routing information for the
broker. Each client that wants to receive messages subscribes to a certain
to have available a huge amount of annotated failure-related
topic and the broker delivers all messages with the matching topic to the
datasets. The automatization, stability, and robustness of the client. Therefore the clients don’t have to know each other, they only
annotation procedure is a challenge that still remains open communicate over the topic.

123
Journal of Intelligent Manufacturing (2023) 34:107–121 109

from the knowledge-based approach, our strategy allows to

completely automatize the feature extraction and prediction
step for computing salient KPIs, which represent the input of
our ML predictive model. Despite meta-heuristic optimiza-
tion approaches (Abualigah, Yousri, et al., 2021; Abualigah,
Diabat, Mirjalili, et al., 2021 are being introduced for solv-
ing a large scale combinatorial optimization problem related
also to PdM tasks Shahbazi and Rahmati (2021); Shah et al.
(2021), the proposal of IoT sensing technologies together
with the designing of feature extraction stage and the deploy-
ing of ML model ensure to directly learn from data for solving
the PdM task. This peculiarity allows monitoring, annotating,
and consequently processing a large amount of heteroge-
neous data thus leading our approach relevant for both the
ML and Big Data analytics area.
In the Big Data scenario, another open challenge is the
difficulty to ensure accurate predictive performance and high
interpretability at the same time Dalzochio et al. (2020). For
example, the solution proposed in Hegedűs et al. (2018) pro-
vides a pre-processing step so that data are usable for PdM.
Further work was done by Strauß et al. (2018), proposing a
framework based on IoT and low-cost sensors that enables the
monitoring and data acquisition on a heavy lift electric mono-
rail system. On the contrary, our approach aims to acquire
directly (i.e. without an additional pre-processing) a labeled
and structured dataset that can be representative of the PdM
task.
Fig. 1 Experimental setup in the real industrial use case: advanced In the context of supervised learning, one important
processing (a) and measuring (b) machines limitation is the lack of data showing the annotation of
anomaly state behavior Adhikari et al. (2018); Dalzochio
et al. (2020); Xu et al. (2019). In Gatica et al. (2016), the
ML. We describe the PdM task and the proposed real indus- authors approached this problem by proposing a top-down
trial use case in “Real industrial use case PdM task”. The strategy consisting ofrst understanding machine operation
description and implementation of the proposed DSS based and then taking action to deal with the problem. On the
on ML strategy is reported in “Method” section. In “Exper- contrary, here we build a data-driven model that can learn
imental procedure” section the evaluation procedure with anomaly situations closely related to productivity losses. The
respect to the state-of-the-art ML models is reported. The model is able to generalize across different production cycles
experimental results of our solution are reported in “Exper- and operating conditions.
imental results” section. Finally, the conclusions and future Starting from the fact that manufacturing plants are
work are reported in “Experimental results” section. dynamic environments, both the lack Nuñez and Borsato
(2018); Selcuk (2017) and the excess Gatica et al. (2016);
Sarazin et al. (2019) of data heterogeneity may negatively
Related work affect the ML model. In this context, our approach tried to
reduce the high dimension of collected data, by introduc-
The proposed approach is related to the general field of PdM. ing a feature extraction stage based on hand-crafted features.
The state-of-the-art work in this field poses different chal- Besides the fact that this step enhances the interpretability
lenges related to the downtime and maintenance-related costs and explainability of the data, experimental results demon-
and different solutions for improving production efficiency. strated how it increase the generalization power of the ML
According to the literature review proposed in Dalzochio model.
et al. (2020), the state-of-the-art solutions to deal with Generally, state-of-the-art work test different ML models
this generic challenge can be divide into four groups: (i) to evaluate which one is more suitable for a given situation.
knowledge-based, (ii) Big Data analytics, (iii) Machine For example in Schmidt and Wang (2018), the authors pre-
Learning models and (iv) Ontology and reasoning. Different sented a classification model to predict failure using as input

123
110 Journal of Intelligent Manufacturing (2023) 34:107–121

vibration limit value and used the accuracy of the models ties. In the latter, the graphical model is part of a framework
as an evaluation metric. Differently in Xu et al. (2019) and designed to predict failures and to measure the impact of
Nuñez and Borsato (2018), the computation cost for training such a prediction on the quality of production planning pro-
the ML models is considered as an important aspect to take cesses and maintenance costs. Another class of ML models
into account. In contrast with respect to the above-mentioned includes the application of auto-regressive models (e.g. Auto-
literature, our employed ML model represents the best trade- regressive Moving Average) for predicting future behavior by
off between the accuracy prediction, computation effort and using historical data. In Adhikari et al. (2018), the authors
model interpretability for optimizing the machining quality apply the auto-regressive integrated moving average model
processes in Industry 4.0 scenario. in a predictive maintenance framework to predict the remain-
In Peres et al. (2018) and Li et al. (2017), the artifi- ing useful life of components.
cial neural network model was employed for providing the The main differences with the proposed approach lie in: (i)
fault prediction and identifying abnormal behaviors. Another the prediction of processing quality and anomaly situations
work considers the implementation of auto-associative neural that are quantitatively annotated in our training dataset by a
networks for finding irregularity in railways Liu et al. (2018). 3D coordinate measuring machine; (ii) the integration of the
The recent advances of technology and the huge amount of proposed ML approach in our IoT platform that enables the
data have laid the foundations to apply DL methodologies for collection of a huge amount of data and provide actionable
solving PdM task Jin et al. (2017); Khan and Yairi (2018). The decision recommendations for resolving productivity losses
application of these solutions includes the implementation of and maintenance issue; (iii) the combination of a feature
a Recurrent Neural Network (RNN) and Long-Short-Term extraction technique with a Random Forest (RF) regression
Memory (LSTM) Cachada et al. (2018); Rivas et al. (2019) model for ensuring the best trade-off between the accuracy
for predicting failure by modeling spatio-temporal relation- prediction, computation effort, and model interpretability.
ship across historical data. In the context of our proposed
approach, the sequential DL approaches (such as RNN and
LSTM) may represent an affordable predictive model by Real industrial use case PdM task
learning spatio-temporal features. However, the potential of
DL approaches may be limited by the interpretability of the The proposed PdM task originated from a specific company’s
model Lipton (2018), which does not always allow to provide demand: the prediction of processing quality and anomaly sit-
clues on how and why the algorithm achieved the selected uations during the machining of a tool. The input parameters
prediction. Taking into account this consideration, as we shall were represented by topic at level 0, i.e. processing param-
see in the experimental section, our ML-based approach per- eters (e.g. acceleration, speed, position) collected from two
forms favorably over other standard (Multi layer perceptron) different machine centers (see Fig. 1a). The condition mon-
and sequential DL approaches (i.e. LSTM). itoring data were represented by topic at level 2 acquired
From ML perspective, most related to our work are the by a robotic part loading system for coordinate measuring
papers of Calabrese et al. (2020), Adhikari et al. (2018), machine (see Fig. 1b). The overall flexible integrated man-
Schmidt and Wang (2018), Zhou et al. (2018), Carbery et al. ufacturing system includes two operator loading/unloading
(2018), Ansari et al. (2020), which proposed the application stations, two robot loading/unloading stations, one auto-
of ML and DL model for predicting RUL and associated cost mated vertical parts store and one parts washing unit.
using sensor and event log data.
Common ML approaches for solving PdM task Bilski Machining centers
(2014); Calabrese et al. (2020); Schmidt and Wang (2018);
Zhou et al. (2018) include the application of standard clas- The processing parameters were acquired by two different
sifiers such as k-Nearest Neighbor (K-NN), Decision Tree machining centers (mc) (i.e. MCM3 clock 5-axis machining
(DT), Naive Bayes (NB), Support Vector Machine (SVM), centers). On a 5-axis mc, the cutting tool moves across the
Random Forest (RF), Boosting and XGBoost. In particular, X, Y and Z linear axes and rotates on the A and B axes to
the RUL prediction task was solved in Calabrese et al. (2020) approach the workpiece from any direction. The mc can be
using log-based data and XGboost algorithm. An evolution of configured for multitasking operations, such as milling, turn-
ensemble learning (i.e. graph based ensemble learning) was ing, grinding, boring, etc. Moreover, all mc can be configured
presented in Zhou et al. (2018) for modeling the behaviour of with a single pallet, pallet exchanger, multi-pallet systems or
different subsystems using different base learners. Graphical integrated in a flexible manufacturing system (FMS). The
models based on Bayesian network and dynamics Bayesian level of automation can be changed or increased during the
network were also proposed in Carbery et al. (2018) and service life of the plant, providing considerable flexibility.
Ansari et al. (2020) for learning causal relationships among
features and across time in terms of conditional probabili- 3 https://www.mcmspa.it.

123
Journal of Intelligent Manufacturing (2023) 34:107–121 111

Coordinate measuring machine for quality control • Data analytics: the ML model outcomes are displayed
task in a GUI-based data analytic interface for supporting the
maintainer/operator.
The condition monitoring data were acquired by a robotic
part loading system for coordinate measuring machine Data collection
(CMM) (i.e. Hexagon4 Manufacturing Intelligence Robotic
CMM part loading). The CMM system automatically iden- The central sensing and communication point is the MQTT
tifies in real-time parts that are out of tolerance and triggers broker (Mosquito), which is in charge of dispatching all mes-
alarm situations. The CMM can be easily used by opera- sages between the senders and the rightful receivers. The
tors with minimal training as a cost-effective automated part implementation of MQTT protocol in JFMX was performed
loading system that increases the throughput of CMM and by using the Paho Java library. Each client that published a
maximizes operational capacity. message to the broker, includes a topic into the message, that
represents the routing information for the broker. Each client
that wants to receive messages subscribes to a certain topic
Method and the broker delivers all messages with the matching topic
to the client. Therefore the clients don’t have to know each
The proposed ML-based strategy is conceived to solve the other, they only communicate over the topic. This architec-
above-mentioned PdM task by allowing the continuous col- ture enables highly scalable solutions without dependencies
lection of an annotated dataset and the provision of a data between the data producers and the data consumers. The pay-
analytic interface for supporting the maintainer/operator. load of messages are just a sequence of bytes, up to 256Mb,
Figure 2 describes the overall architecture of the proposed with no requirements placed on their format, and with MQTT
DSS, which is comprised of five IoT and ML cornerstones: protocol usually adding a fixed two bytes header to most mes-
sages. Other clients can subscribe to these messages and get
• Data collection: the IoT sensing technology is based on updated by the broker when new messages arrive.
the Message Queuing Telemetry Transport (MQTT) bro- The central concept in MQTT to dispatch messages are
ker5 Andrew Banks Ed Briggs and Gupta (2019). The topics. A topic is a simple string that can have more hierarchy
central concept in MQTT dispatcher is topics that collect levels. For example, a topic for sending status data of mc2 of
processing parameters (e.g. acceleration, speed, posi- an FMS is the following: JFMX/L1/fms/UNIT/mc1/STATUS.
tion) of the machining centers at a lower level (level 0). On one hand, the client can subscribe to the exact topic or on
Accordingly, the status and conditioning data collected the other hand use a wildcard. The wildcard (+) allows arbi-
by the Hexagon machine represents the topic at a higher trary values for one hierarchy while the multilevel wildcard
level (level 2). All data are synchronized and are collected (#) allows to subscribe to more than one level (e.g. the entire
in a SQL database and then in Azure Blob cloud storage. subtree).
• Feature extraction: the Trapezoidal Numerical Integra- The MQTT topic organization allows to guarantee the
tion (TNI) is performed to compute a Key Performance Quality of Service (QoS). The MQTT protocol handles
Indicator (KPI) for each processing parameter during retransmission and ensures the delivery of the message,
each MCM production cycle. regardless how unreliable the underlying transport is. In addi-
• Predictive model: a RF regression model is applied in tion, the client is able to choose the QoS level depending on
order to estimate the status and conditioning data using its network reliability and application logic:
the collected processing parameters as predictors.
• Cloud architecture: the ML model is deployed as a • QoS 0—at most once it guarantees a best effort delivery.
docker container with an API endpoint for testing unseen A message won’t be acknowledged by the receiver or
acquired data. Azure ML service is used for providing a stored and redelivered by the sender. This is often called
cloud-based environment for deploying and updating ML “fire and forget” and provides the same guarantee as the
model. Azure Blob storage is adopted to store the ML underlying TCP protocol.
model weights and the ML model outcomes (predicted • QoS 1—at least once it is guaranteed that a message
error %). will be delivered at least once to the receiver. The sender
will store the message until it gets an acknowledgement in
4
form of a PUBACK command message from the receiver.
https://www.hexagonmi.com.
5
• QoS 2—exactly once it guarantees that each message is
MQTT is a lightweight telemetry protocol, coming from the world of
M2M and now widely applied in IoT. The central communication point
received only once by the counterpart. It is the safest and
is the MQTT broker, which is in charge of dispatching all messages also the slowest QoS level. The guarantee is provided by
between the senders and the rightful receivers. two flows there and back between sender and receiver.

123
112 Journal of Intelligent Manufacturing (2023) 34:107–121

Fig. 2 Flowchart of the proposed approach: data collection, feature extraction, prediction phase based on Random Forest predictive model, cloud
storage and data analytics

Table 1 Topics related to the working step analyzer application embed-

ded in the L0-Machine Agent
Field Type Description

ts Date Start date of the episode according to the

clock of the mc
l2mc String mc code
l2wa String workArea code
ordNo String Order number
ptType String Identifier of the part type
opNo String Operation number
dType String Identifier of the workingstep message type
(”wsEv”: in case of sensor coming from
mc, ”accTrace”: in case of sensor
coming from Accelerometer)
Slot Int Subsection as indicated by the part
program
pt String Identifier of the part machined
tl String Identifier of the tool formatted as
<tooltype>/<tool serial number>
Fig. 3 Flowchart of the data layer: jFMX MQTT Namespace Life Int Life of the tool at the beginning of the step
Sensor String Name of the sensor
Unit String Measure unit for the specifc sensor
Sampling Int Milliseconds sampling interval
dataObj Object Complex object containing two elements:
jFMX MQTT Namespace The MQTT topic namespace was n[int] segment of the acquisition, data:
Array[Double] data acquired
defined to manage interactions with the IoT application run-
mc String Machine name
ning on the jFMX gateway hierarchy. Figure 3 shows the
hierarchy of the jFMX MQTT Namespace. wa String workArea name
Based on this criterion, our IoT application running on an
IoT gateway may be viewed in terms of the resources it owns
and manages as well as the unsolicited events it reports:

123
Journal of Intelligent Manufacturing (2023) 34:107–121 113

• account_name: Identifies a group of devices and users. The topics related to the dataObj field represent the con-
It can be seen as partition of the MQTT topic namespace. sidered processing parameters (see Table 2) acquired by mc
For example, access control lists can be defined so that (i.e. computer numerical control [CNC] and accelerometer).
users are only given access to the child topics of a given The CNC and accelerometer data were acquired by a sam-
account_name. pling frequency of 24 and 100 Hz respectively.
• client_id: Identifies a single gateway device within an
account (typically the MAC address of a gateway’s pri- L2 CMM Topic The L2 CMM topics represents the condition
mary network interface). The client_id maps to the Client monitoring data that were acquired by a robotic part loading
Identifier (Client ID) as defined in the MQTT specifica- system for coordinate measuring machine (see Table 3). The
tions. condition monitoring data reflect the quality of processing
• app_id: Unique string identifier for application (e.g., in terms of the measured deviation (deviation) with respect
”L0” for mc topics, ”L2” for CMM topics). to the optimal condition. The optimal condition highlighted
• resource_id: Identifies a resource(s) that is owned and no deviation compared to the planning processing. The alarm
managed by a particular application. Management of situation is triggered once the measured deviation overcomes
resources (e.g., sensors, actuators, local files, or con- the admitted tolerance.
figuration options) includes listing them, reading the All the L2 CMM Topic and L0 Machine Center Topic
latest value, or updating them to a new value. A were synchronized by considering the physical tool (tl,
resource_id is a hierarchical topic, where, for example, identifier of the tool formatted as <tooltype>/<tool serial
”fms/mc1/spindle/temp” may identify a temperature sen- number), the part machined (pt) and the type of processing
sor and”fms/sh/Y/pos” a position sensor. (frindex). Although the system could consider all the type
of processing we took into account the drilling procedure
L0 Machine Center Topic The jFMX MQTT publisher is (i.e. FRINDEX = 10,20,30). This procedure has the intrinsic
executed by the L0-Gateway (Flight Recorder) and deliv- advantage of being standard, i.e. independent of the tl. The
ers message related to different application included inside synchronized L2 CMM Topic together with the L0 Machine
the L0-MachineAgent. All the topics published by the L0 Center Topic were saved in a SQL database.
MachineAgent is related to the processing parameters of mc
and have an app_id defined as: JFMX/L0/workAreaName/ Feature extraction
unitName where the workAreaName is the absolute unique
name of the workarea and the unitName is the name of the All the computed KPIs (speed, pow, pos, curr) represent the
unit inside the workarea. predictors of the ML model for each observation/physical
For our PdM task we refer to the topics related to the Work- tool (i.e. specific triplet tl, pt and frindex). The output of the
ing Step Analyzer application embedded in the L0-Machine ML model was represented by the percentage measurement
Agent related to the two different mc (see Table 1) error (error %).

Table 2 Topics related to the

Sensor (CNC) Unit Description
dataObj field
SP_SPEED rpm Spindle rotation speed
SP_POW W Spindle power consumption
X_AXIS_CURR A x-axis current consumption
Y_AXIS_CURR A y-axis current consumption
Z_AXIS_CURR A z-axis current consumption
B_AXIS_CURR A B-axis current consumption
A_AXIS_CURR A A-axis current consumption
X_AXIS_POS μm X-axis position
Y_AXIS_POS μm Y-axis position
Z_AXIS_POS μm Z-axis position
B_AXIS_POS μm B-axis position
A_AXIS_POS μm A-axis position
FRINDEX None Type of processing
Sensor (accelerometer) Unit Description
velMOD_RMS mm/s 3 directions vibratory speed module

123
114 Journal of Intelligent Manufacturing (2023) 34:107–121

Table 3 L2 CMM topics

Field Type Description

ts Date Start date of the episode according to the clock of the mc

pt String Identifier of the part machined
tl String Identifier of the tool formatted as <tooltype>/<tool serial number>
FRINDEX Int type of processing
Measured Double Measured value
Deviation Double Measured deviation
Tolerance Double Admitted tolerance

Notation We let a candidate univariate time series of a as follows:

specific sensor S collected from a CNC sensor as X =
{x 0 , x 1 , . . . , x T } where T denotes the number of observa- deviation
error% = × 100 (2)
tions. Notice how this time series is relative to the signal of tolerance
a specific sensor and relative to a specific triplet comprised
In particular, an error % greater than 100 reflects an out of
of physical tool (tl), the part machined (pt) and the type of
tolerance machining while an error % lower than 100 cor-
processing (frindex). We denote the err or % as a direct quan-
respond to a machining that does not exceed the tolerance
titative measure about the machine quality and the deviation
limits.
and tolerance the measured deviation and tolerance reported
The final dataset consist of 438 observations/physical
in Table 3.
tools collected by two different mc from the 1st October
Our feature extraction strategy is based on a geomet-
2019 to the 31st May 2020.
ric area analysis (GAA) and trapezoidal area estimation
(TAE) procedure that is widely used for solving novelty and
anomaly detection task Bulirsch and Stoer (2002); Moustafa Predictive model
et al. (2019). The relative KPI is computed by temporally
normalizing the TAE as follows: For solving the regression task we have taken into account
predictive performance, interpretability, and predictive accu-
racy. These factors represent also the three fundamental
 T requirements defined by the company for solving the PdM
1
K PI = xt dt task. For this reason we selected the RF model for solving the
T 1
T  t
regression task. RF represents a variant of bagging proposed
1  by Breiman (2001) and consists of an ensemble of regression
= xt dt
T
t=1 t−1 trees (RTs) (i.e., n ◦ of RT) generated by independent identi-
cally distributed random vectors. RF is modeled by sampling
11 
T
≈ (t − (t − 1))[xt − xt−1 ] (1) from the observations, from the features (i.e., n ◦ of features
2T to be selected) and by changing two tree-parameters (i.e.,
t=1
max n ◦ of splits and max n ◦ of size) Kuncheva (2004). The
idea behind this sampling is to maximize the diversity among
All the computed KPI for each specific sensor S is depicted trees, by sampling from the features set and from the data set
in Table 4. For the position and current a global KPI was as well. In particular, the random feature selection is carried
extracted by computing the euclidean distance of X,Y,Z axis. out at each node of the tree, by choosing the best feature to
For each observation, the error % was computed by con- split within a subset of the original feature set. In a standard
sidering the measured deviation and the associated tolerance classification problem, the idea is to built an ensemble of DT
that aims to split the data into subsets that contain instances
with similar values (homogenous). Since we aim to solve a
Table 4 Extracted KPIs for regression problem, the best splitting features for each node
Sensor (CNC) KPI
specific sensor. KPIs are related was computed according to the sum of squared error.
to rotation speed (speed), power SP_SPEED speed Although RF allows learning a non-linear decision bound-
consumption (pow), position ary, the RF originated as ensemble tree based model ensures
SP_POW pow
(pos) and current consumption
(curr) X,Y,Z_AXIS_POS pos an intuitive notion of interpretability: it allows providing a
X,Y,Z_AXIS_CURR curr direct interpretation of the most discriminative KPIs. How-
ever, the degree of interpretability depends on the model size

123
Journal of Intelligent Manufacturing (2023) 34:107–121 115

(i.e., number of weak learners/regression tree and depth of the for instance, a drift situation was detected. All the prediction
tree) Molnar et al. (2019). Hence, the interpretability of RF results together with the model weights (i.e. decision rule of
was encouraged by constraining the number of weak learners the ensemble trees) were stored in the Azure blob storage.
and the depth of the tree in the validation set. This lead also
to control the computation effort for the training phase. Data analytics
The importance of a specific KPI in the RF model to
identify the percentage measurement error (error %) was A GUI interface was created to display the predicted error %
measured according to a permutation of out-of-bag feature over different tl, pt, frindex. In particular, the GUI was final-
observation Boulesteix et al. (2012). A KPI is considered rel- ized to provide a timely indication to the machine operator
evant to identify the error %, if permuting its values should when the error % exceeds a certain tolerance threshold that
affect the model error. On the other hand, if a KPI was not may be different for each tl, pt, frindex.
relevant, then permuting its values should not affect sig- Additionally the Azure Application Insights instance was
nificantly the model error. The permutation importance of enabled for providing a high level overview of the deployed
each feature is computed as: 1 − err or (after permuting the API in terms of featuring failed requests, response time, num-
feature values). Compared to the standard impurity-based ber of requests and availability. The log analytic feature of
importance the permutation approach is unbiased towards the Application insights allows to view and inspect the logs
high cardinality features and measure directly the ability of provided from our containerized model in terms of stdout
feature to be useful to make prediction Janitza et al. (2018). and stderr.

Cloud architecture
Experimental procedure
A container logic was adopted for packaging the ML applica-
tion and all its dependencies, so the application runs reliably
Experimental comparisons
from one computing environment to another. A docker image
is essentially a snapshot of a container. Microsoft Azure
We decided to compare our RF based PdM approach with
IoT portal was adopted for providing a cloud-based environ-
respect to other state-of-the-art ML approaches employed
ment based on virtualized containers. This environment can
for solving PdM task. In particular we have considered the
ensure hardware and software isolation, flexibility, and inter-
following models:
dependencies between the IoT devices and data collection,
features extraction, and prediction phases. These properties
are suitable for our industrial use case since the proposed DSS • Linear Regression with ridge penalty (LR ridge) Susto
is currently designed to work with four operating machines et al. (2012);
and it provides the capability to be scaled up to collect a huge • Linear Regression with elastic net penalty (LR elastic)
amount of data from different interconnected machines. This Susto et al. (2012);
advantage also lies the foundations to continuously update • Regression Tree Romeo et al. (2020); Zhou et al. (2018);
the model, once a new machine is connected to the system. • XGBoost Calabrese et al. (2020);
Our architecture is depicted in Fig. 2 Cloud Architecture. • SVM with Gaussian Kernel Bilski (2014)
We used a Python ML library for training and testing our • Multi layer perceptron (MLP) Jin et al. (2017); Khan and
feature extraction stage and RF model with respect to other Yairi (2018)
state-of-the-art ML approaches. Afterward the feature extrac- • Long Short Term Memory (LSTM) Cachada et al. (2018);
tion procedure and the containerized RF model was pushed Rivas et al. (2019)
to Azure Container Registry. During this step, we included
the azureml-monitoring and azureml-defaults for enabling Experimental design
respectively the data collection feature and the deployment
to Kubernetes. Consequently, the ML model was deployed to A 10-fold Cross Validation (10-CV) procedure was per-
Azure Kubernetes Service (AKS). For our purpose we config- formed in order to evaluate the performance of the RF model.
ured the AKS with 3 agent nodes of type Standard_D3_v2 (4 The hyperparameters were optimized by implementing a grid
vCores), thus leading to a total of 12 vCores. Additionally, in search in a nested 5-CV. Hence, each split of the outer CV
our AKS configuration, we explicitly enable data collection loop was trained with the optimal hyperparameters (in terms
(input data and predictions outcome). The L2 CMM Topic of mean squared error) tuned in the inner CV loop. Despite
together with the L0 Machine Center Topic were exported this model checking procedure is expensive in terms of com-
from SQL database to azure blob storage, by allowing a con- putation effort it allows to obtain an unbiased and robust
tinuous testing and update/retraining of the ML model once, performance evaluation Cawley and Talbot (2010).

123
116 Journal of Intelligent Manufacturing (2023) 34:107–121

Table 5 shows the different hyperparameters for the pro- • Mean Absolute Error (MAE): it measures the absolute
posed ML models and all competitors’ ML approaches, as difference between the predicted and the ground truth
well as the grid-search set. error %;
For the LR ridge the λ penalty controls the 2-norm regular- • Mean Squared Error (MSE): it measures the squared dif-
ization. For the LR Elastic α = λ1 + λ2 and l1_ratio= λ1λ+λ1
2
ference between the predicted and the ground truth error
where λ1 and λ2 control separately the 1-norm and 2-norm %;
regularizations. • R 2 score (coefficient of determination): it is a proportion
between the variability of the data and the correctness of
the model used. It varies in range: [−∞; 1] Di Bucchi-
anico (2008).

Evaluation metrics The statistical significance of the R 2 score was evaluated

at the 5% significance level with respect to the zero value. The
The following metrics are considered to evaluate the predic- R 2 score distribution over each CV fold was found to follow
tive performance of the regression task: a normality distribution according to the Anderson-Darling

Table 5 Predictive performance

Model Hyp Range
and computation effort of ML
approaches RF n ◦ of regression trees {5, 10, 15, 20, 25}
n ◦ of features to select {1, 2, 4}
LR ridge λ {10−4 , 10−3 , 10−2 , 10−1 }
LR elastic α {10−4 , 10−3 , 10−2 , 10−1 }
l1_ratio {10−4 , 10−3 , 10−2 , 0.1, 0.2, 0.3, 0.4, 0.5}
RT Max depth {5, 10, 15, 20, 25}
Min n ◦ of leaf size {5, 10, 20, 50, 100}
XGBoost Learning rate {0.001, 0.01, 0.10}
Max n ◦ of estimators {5, 10, 15, 20, 25}
Max depth {5, 10, 15, 20, 25, 50, 75}
n ◦ of features to select {1, 2, 4}
SVM Gaussian Box constraint {1, 5, 10, 50, 100, 500, 103 , 5 × 103 , 5 × 104 , 104 }
Kernel Scale {10−2 , 0.1, 1, 10, 102 , 103 , 104 }
MLP Learning rate {10−5 , 10−4 , 10−3 , 10−2 }
n ◦ of hidden layers {1, 2, 4}
n ◦ of units {4, 8, 16, 32}
LSTM Learning rate {10−5 , 10−4 , 10−3 , 10−2 }
n ◦ of hidden layers {1, 2, 4}
n◦ of units {4, 8, 16}

Table 6 Predictive performance

Model MAE MSE R2 Training + validation (s) Testing (s)
and computation effort of ML
approaches RF 0.089 0.018 0.868∗ 2.353 (0.328) < 10− 3 (0)
LR ridge 0.166 0.054 0.591∗ 0.521 (0.007) < 10− 3 (0)
LR elastic 0.166 0.054 0.591∗ 0.531 (0.058) < 10− 3 (0)
RT 0.100 0.023 0.835∗ 0.310 (0.022) < 10− 3 (0)
XGBoost 0.088 0.017 0.877∗ 80.740 (10.057) < 10− 3 (0)
SVM Gaussian 0.151 0.047 0.648∗ 1.223 (0.101) < 10− 3 (0)
MLP 0.161 0.051 0.618∗ 39.454 (0.651) 0.062 (0.002)
LSTM 0.156 0.044 0.667∗ 89.549 (2.618) 0.570 (0.023)
The best results for each metric are indicated in bold
∗ Whether the R2 distribution over the 10-fold is significantly higher than 0

123
Journal of Intelligent Manufacturing (2023) 34:107–121 117

test (A = 0.436, p = 0.246). Hence, we used the parametric

paired t-test (α = 0.05) to compare the performance of the
proposed approach with respect to state-of-the-art work.

Experimental results

In this section we show the experimental results for the the

proposed DSS specifically tailored for solving PdM task. The
prediction of the processing quality represents the main task
we aim to solve. All the results related to the predictive per-
formance and computation effort of the proposed approach
with respect to the state-of-the-art approaches are depicted
in “Predictive performance” and “Computation effort” sec- Fig. 4 Predictive performance of RF for the estimation of the error %
over a subset of observations (tl, pt, frindex): red line ground truth, blue
tions, while the results related to the model interpretability,
line prediction (Color figure online)
i.e. the most relevant KPIs, are shown in “Interpretability”
section. More details on the implemented GUI for the pro-
posed DSS are reported in “Data analytics: GUI interfaces”
section.

Predictive performance

The predictive performance of the RF model is shown in

Table 6. It can be noted that the best prediction results were
obtained for our RF and XGboost model (R 2 score 0.868 and
0.877 respectively), while the LR model achieved the lowest
predictive performance (R 2 score 0.591). Accordingly the
RF and XGboost models show similar and competitive per-
formance in terms of MAE (0.089 and 0.088 respectively)
Fig. 5 Feature/permutation importance of RF model(Color figure
and MSE (0.018 and 0.017 respectively). R 2 score distribu- online)
tion of RF is significantly higher ( p < 0.05) than ML based
regression model (i.e. LR ridge, LR elastic net, RT, SVM
Gaussian) and DL based regression model MLP and LSTM. of RAM and NVIDIA GeForce GTX 970. Although the pre-
In particular, the performance of sequential DL approaches dictive performance of RF is similar to XGboost, the training
might be limited by the low presence of a huge amount of and validation of RF model are significantly ( p < 0.05)
annotated sequential data in this PdM scenario, in order to faster than XGboost with a gain of 34x. This peculiarity
learn spatio-temporal dependencies. ensures the possibility to retrain the ML model in the cloud
Figure 4 shows the comparison between the predicted with an average latency of 2.353 s for learning from around
error % from RF and its real values obtained from the L2 400 new samples. At the same time the RF prediction latency
CMM Topic. We focused on a subset of 42 testing samples may be neglected < 10−3 s and the RF model can give a
(one fold of CV-10 procedure). timely and consistent prediction.

Computation effort Interpretability

Taking into account the high performance achieved by the The interpretability of the proposed RF model was measured
proposed approach, we decided to test the computation effort according to the feature/permutation importance (Figure 5).
with respect to other state-of-the-art ML approaches. Table 6 The speed KPI achieved, on average, the highest permutation
compares the time effort (test and training+validation stage) importance score, thus highlighting the most discriminative
of the proposed RF algorithm with respect to other state-of- power of this KPI with respect to the other features. The fea-
the-art algorithms. All the experimental comparisons were ture importance together with the predicted error % values
performed on Intel Core i7-4790 CPU 3.60 GHz with 16 GB are the salient ML outcomes of the proposed DSS for sup-

123
118 Journal of Intelligent Manufacturing (2023) 34:107–121

Fig. 6 Examples of GUI

interface of the proposed DSS
for four different observations
(tl, pt, frindex): the predicted
error % is represented by blue
line and reported below the
gauge chart, the black line
represents the admissible
tolerance threshold that can
change across different triplets.
Notice how an error % greater (a) (b)
than the admissible threshold
(red bar) represents a significant
machining error (i.e. alarm
event), while an error % that
falls within the yellow bar
correspond to a potential risk
situation (i.e. machining error
below but close to tolerance
limits). The green bar reflect
how the machining is properly
executed within the tolerance
limit (Color figure online) (c) (d)

porting the maintainer/operator during the machining quality ing the operators to (i) predict alarm situation (i.e. significant
task. For instance, taking into account the predicted error %, machining error) and (ii) interpret and localize the source of
the operator may exploit preventive action in order to avoid the error by focusing on the average and temporal value of
future errors during the machine processing. At the same the most discriminative KPI predictors (see Fig. 5).
time, the localization of the most discriminative KPI may
address the human operator to detect the source of the error,
while optimizing the overall equipment effectiveness, pro- Conclusions
ductivity, and quality of production.
We have shown that our DSS approach is effective in solv-
Data analytics: GUI interfaces ing the machining quality and PdM task in a real industrial
use case. The main pillars of the proposed DSS consist of
Figure 6 shows an examples of GUI interfaces of the pro- data collection, feature extraction, predictive model, cloud
posed DSS for four different observations (tl, pt, frindex). storage, and data analytics. The strict collaboration between
We represent the predicted error % and the tolerance limits, the company and University allowed to design, validate and
which can be different for each observation. In particular, test the proposed approach in a real industrial PdM use
a predicted error % greater than the admissible threshold case task, by considering the dataset described in “Fea-
(red bar) represents a significant machining error (i.e. alarm ture extraction” section. Taking into account the achieved
event), while an error % that falls within the yellow bar corre- experimental results, we demonstrated the effectiveness of
spond to a potential risk situation (i.e. machining error below our theoretical frameworks into a real industrial environ-
but close to tolerance limits). The green bar reflect how the ment. In fact, our DSS approach was demonstrated to be
machining is properly executed within the tolerance limit. the best trade-off between predictive performance, computa-
Taking into account the high predictive results and the tion effort, and interpretability. These peculiarities together
interpretability of the proposed approach, our GUI interface with the integration of our ML approach into the proposed
is not limited to show only the predicted error % (i.e. Fig. 6). cloud-based architecture would allow the optimization of
In fact, we go further by supporting the operator by showing the machining quality processes by directly supporting the
the average value of the KPI predictors (see Fig. 7) for a spe- maintainer/operator. These advantages may impact to the
cific tl across different pt and frindex. Additionally, the trend optimization of maintenance schedules and to get real-time
of KPI predictors together with the life parameter (i.e. life of warnings about operational risks by enabling manufacturers
the tool at the beginning of the step) is displayed for a spe- to reduce service costs by maximizing uptime and improving
cific tl and frindex across pt and across time in a separated productivity.
dashboard, as shown in Fig. 8. These GUI interfaces may A current limitation is represented by the employed ML
empower the overall machining quality process by support- model, which is re-trained every time from scratch when a

123
Journal of Intelligent Manufacturing (2023) 34:107–121 119

Fig. 7 Example of the proposed GUI interface displaying the average value of the KPI predictors for a specific tl across different pt and frindex

Fig. 8 Example of the proposed GUI interface displaying the trend of KPI predictors (blue line) together with the life parameter (red line) (i.e. life
of the tool at the beginning of the step) for a specific tl and frindex across pt (top graph) and across time (bottom graph)

certain amount of new data is stored. As future work, we Abualigah, Diabat, Sumari, et al., 2021) could be imple-
aim to integrate a fully-automated incremental learning pro- mented to help in selecting the optimal hyperparameters for
cedure by updating continuously the model parameter Chai ML model to improve the stability and the testing predictive
and Zhao (2020). The new data continuously acquired and performance Shah et al. (2021); Zhang et al. (2020). For the
stored over time in the proposed cloud framework can be full applicability of PdM in all the company’s tasks, another
used to refine ML model and improve its predictive per- future work direction could be addressed to build integrated
formance according to an incremental learning procedure. cost-benefit models that include the impact and the benefit of
As a further limitation, the proposed model only works for our approach on the entire asset management of the company
certain tools installed on the processing machines. How- Compare et al. (2019).
ever, as future work, for the proposed ML model we aim
to improve the generalization performance of the proposed Acknowledgements The authors would like to thank Benelli Armi Spa
approach across different tools and types of processing. In and Sinergia Consulenze Srl for their contribution.
addition, meta-heuristic algorithms (Abualigah et al., 2022;

123
120 Journal of Intelligent Manufacturing (2023) 34:107–121

Funding Open access funding provided by Università Politecnica delle Boulesteix, A.-L., Janitza, S., Kruppa, J., & König, I. R. (2012).
Marche within the CRUI-CARE Agreement. This work was supported Overview of random forest methodology and practical guidance
within the research agreement between Università Politecnica delle with emphasis on computational biology and bioinformatics. Data
Marche and Benelli Armi Spa for the ”4USER Project” (User and Mining and Knowledge Discovery, 2(6), 493–507.
Product Development: from Virtual Experience to Model Regenera- Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.
tion) funded on the POR MARCHE FESR 2014-2020-ASSE 1-OS Bulirsch, R. & Stoer, J. (2002). Introduction to numerical analysis (Vol.
1-ACTION 1.1-INT. 1.1.1. Promotion of industrial research and exper- 3). Springer.
imental development in the areas of smart specialisation -LINEA 2 Burhanuddin, M. A. , Halawani, S. M. & Ahmad, A. (2011). An efficient
-Bando 2019, approved with DDPF 293 of 22/11/2019. failure-based maintenance decision support system for small and
medium industries. In Efficient decision support systems: practice
and challenges from current to future 195. InTechOpen.
Declarations
Cachada, A. , Barbosa, J. , Leitño, P. , Gcraldcs, C. A. , Deusdado, L.
, Costa, J., Teixeira, C., Teixeira, J., Moreira, A. H. J., Miguel, P.,
Conflict of interest The authors declare that they do no have any conflict & Romero, L. (2018). Maintenance 4.0: Intelligent and predictive
of interest. maintenance system architecture. In 2018 IEEE 23rd international
conference on emerging technologies and factory automation
Open Access This article is licensed under a Creative Commons (ETFA) (Vol. 1, pp. 139–146).
Attribution 4.0 International License, which permits use, sharing, adap- Calabrese, M., Cimmino, M., Fiume, F., Manfrin, M., Romeo, L., Cec-
tation, distribution and reproduction in any medium or format, as cacci, S., & Kapetis, D. (2020). Sophia: An event-based IoT and
long as you give appropriate credit to the original author(s) and the machine learning architecture for predictive maintenance in indus-
source, provide a link to the Creative Commons licence, and indi- try 4.0. Information, 11(4), 202.
cate if changes were made. The images or other third party material Carbery, C. M., Woods, R. & Marshall, A. H. (2018). A Bayesian net-
in this article are included in the article’s Creative Commons licence, work based learning system for modelling faults in large-scale
unless indicated otherwise in a credit line to the material. If material manufacturing. In 2018 IEEE international conference on indus-
is not included in the article’s Creative Commons licence and your trial technology (ICIT) (pp. 1357–1362).
intended use is not permitted by statutory regulation or exceeds the Cawley, G. C., & Talbot, N. L. (2010). On over-fitting in model selection
permitted use, you will need to obtain permission directly from the copy- and subsequent selection bias in performance evaluation. Journal
right holder. To view a copy of this licence, visit http://creativecomm of Machine Learning Research, 11(7), 2079–2107.
ons.org/licenses/by/4.0/. Chai, Z., & Zhao, C. (2020). Multiclass oblique random forests with
dual-incremental learning capacity. IEEE Transactions on Neural
Networks and Learning Systems, 31(12), 5192–5203.
Chen, C., Liu, Y., Wang, S., Sun, X., Di Cairano-Gilfedder, C., Titmus,
References S., & Syntetos, A. A. (2020). Predictive maintenance using cox
proportional hazard deep learning. Advanced Engineering Infor-
Abualigah, L., Abd Elaziz, M., Sumari, P., Geem, Z. W., & Gandomi, matics, 44, 101054.
A. H. (2022). Reptile search algorithm (RSA): A nature-inspired Compare, M., Baraldi, P., & Zio, E. (2019). Challenges to IoT-enabled
meta-heuristic optimizer. Expert Systems with Applications, 191, predictive maintenance for industry 4.0. IEEE Internet of Things
116158. Journal, 7(5), 4585–4597.
Abualigah, L., Diabat, A., Mirjalili, S., Abd Elaziz, M., & Gandomi, Dalzochio, J., Kunst, R., Pignaton, E., Binotto, A., Sanyal, S., Favilla,
A. H. (2021). The arithmetic optimization algorithm. Computer J., & Barbosa, J. (2020). Machine learning and reasoning for pre-
Methods in Applied Mechanics and Engineering, 376, 113609. dictive maintenance in industry 4.0: Current status and challenges.
Abualigah, L., Diabat, A., Sumari, P., & Gandomi, A. H. (2021). Appli- Computers in Industry, 123, 103298.
cations, deployments, and integration of internet of drones (IOD): Di Bucchianico, A. (2008). Coefficient of determination (R 2 ). In Ency-
a review. IEEE Sensors Journal, 21(22), 25532–25546. clopedia of statistics in quality and reliability 1. Wiley.
Abualigah, L., Yousri, D., Abd Elaziz, M., Ewees, A. A., Al-qaness, Gatica, C. P. , Koester, M. , Gaukstern, T. , Berlin, E. & Meyer, M.
M. A., & Gandomi, A. H. (2021). Aquila optimizer: A novel (2016). An industrial analytics approach to predictive maintenance
meta-heuristic optimization algorithm. Computers & Industrial for machinery applications. In 2016 IEEE 21st international con-
Engineering, 157, 107250. ference on emerging technologies and factory automation (ETFA)
Adhikari, P. , Rao, H. G. & Buderath, M. (2018). Machine learning (pp. 1–4).
based data driven diagnostics & prognostics framework for aircraft Hegedűs, C. , Ciancarini, P. , Frankó, A. , Kancilija, A. , Moldován,
predictive maintenance. In 10th international symposium on NDT I., Papa, G., Poklukar, S., Riccardi, M., Sillitti, A., & Varga,
in aerospace. P. (2018). Proactive maintenance of railway switches. In 2018
Andrew Banks Ed Briggs, K. B., & Gupta, R. (2019). Mqtt version 5.0 5th international conference on control, decision and information
oasis standard. Retrieved March 7, 2019, from https://docs.oasis- technologies (CoDIT) (pp. 725–730).
open.org/mqtt/mqtt/v5.0/os/mqtt-v5.0-os.html Janitza, S., Celik, E., & Boulesteix, A.-L. (2018). A computationally fast
Ansari, F. , Glawar, R. & Sihn, W. (2020). Prescriptive maintenance variable importance test for random forests for high-dimensional
of CPPS by integrating multimodal data with dynamic Bayesian data. Advances in Data Analysis and Classification, 12(4), 885–
networks. In Machine learning for cyber physical systems (pp. 915.
1–8). Springer. Jin, L. , Wang, F. & Yang, Q. (2017). Performance analysis and opti-
Ayvaz, S., & Alpay, K. (2021). Predictive maintenance system for mization of permanent magnet synchronous motor based on deep
production lines in manufacturing: A machine learning approach learning. In 20th international conference on electrical machines
using IoT data in real-time. Expert Systems with Applications, 173, and systems (ICEMS) (pp. 1–5).
114598. Khan, S., & Yairi, T. (2018). A review on the application of deep learning
Bilski, P. (2014). Application of support vector machines to the induc- in system health management. Mechanical Systems and Signal
tion motor parameters identification. Measurement, 51, 377–386. Processing, 107, 241–265.

123
Journal of Intelligent Manufacturing (2023) 34:107–121 121

Kuncheva, L. I. (2004). Combining pattern classifiers: Methods and Schmitt, J., Bönig, J., Borggräfe, T., Beitinger, G., & Deuse, J. (2020).
algorithms. Wiley. Predictive model-based quality inspection using machine learning
Li, Z., Wang, Y., & Wang, K.-S. (2017). Intelligent predictive main- and edge cloud computing. Advanced Engineering Informatics,
tenance for fault diagnosis and prognosis in machine centers: 45, 101101.
Industry 4.0 scenario. Advances in Manufacturing, 5(4), 377–387. Schwendemann, S., Amjad, Z., & Sikora, A. (2021). A survey of
Lipton, Z. C. (2018). The mythos of model interpretability. Queue, machine-learning techniques for condition monitoring and pre-
16(3), 31–57. dictive maintenance of bearings in grinding machines. Computers
Liu, Z. , Jin, C. , Jin, W. , Lee, J. , Zhang, Z. , Peng, C. & Xu, G. (2018). in Industry, 125, 103380.
Industrial AI enabled prognostics for high-speed railway systems. Selcuk, S. (2017). Predictive maintenance, its implementation and latest
In 2018 IEEE international conference on prognostics and health trends. Proceedings of the Institution of Mechanical Engineers,
management (ICPHM) (pp. 1–8). Part B: Journal of Engineering Manufacture, 231(9), 1670–1679.
McArthur, J., Shahbazi, N., Fok, R., Raghubar, C., Bortoluzzi, B., & Shah, P. , Sekhar, R. , Kulkarni, A. J. & Siarry, P. 2021. Metaheuristic
An, A. (2018). Machine learning and BIM visualization for mainte- algorithms in industry 4.0. CRC Press.
nance issue classification and enhanced data collection. Advanced Shahbazi, B., & Rahmati, S. H. A. (2021). Developing a flexible man-
Engineering Informatics, 38, 101–112. ufacturing control system considering mixed uncertain predictive
Molnar, C. , Casalicchio, G. & Bischl, B. (2019). Quantifying inter- maintenance model: A simulation-based optimization approach.
pretability of arbitrary machine learning models through functional Operations Research Forum, 2, 1–43.
decomposition. Ulmer Informatik-Berichte 41. Preprint. Strauß, P. , Schmitz, M. , Wöstmann, R. & Deuse, J. (2018). Enabling
Moustafa, N., Slay, J., & Creech, G. (2019). Novel geometric area of predictive maintenance in the brownfield through low-cost sen-
analysis technique for anomaly detection using trapezoidal area sors, an IIOT-architecture and machine learning. In 2018 IEEE
estimation on large-scale networks. IEEE Transactions on Big international conference on big data (Big Data) (pp. 1474–1483).
Data, 5(4), 481–494. Susto, G. A. , Schirru, A. , Pampuri, S. & Beghi, A. (2012). A pre-
Nuñez, D. L., & Borsato, M. (2018). Ontoprog: An ontology- dictive maintenance system based on regularization methods for
based model for implementing prognostics health management ion-implantation. In 2012 semi advanced semiconductor manufac-
in mechanical machines. Advanced Engineering Informatics, 38, turing conference (pp. 175–180).
746–759. Vollert, S. , Atzmueller, M. & Theissler, A. (2021). Interpretable
Peres, R. S., Rocha, A. D., Leitao, P., & Barata, J. (2018). Idarts-towards machine learning: A brief survey from the predictive maintenance
intelligent data analysis and real-time supervision for industry 4.0. perspective. In 2021 26th IEEE international conference on emerg-
Computers in Industry, 101, 138–146. ing technologies and factory automation (ETFA) (pp. 01–08).
Rebuffi, S.-A. , Kolesnikov, A. , Sperl, G. & Lampert, C. H. (2017). Xu, Y., Sun, Y., Liu, X., & Zheng, Y. (2019). A digital-twin-assisted fault
ICARL: Incremental classifier and representation learning. Pro- diagnosis using deep transfer learning. IEEE Access, 7, 19990–
ceedings of the IEEE conference on computer vision and pattern 19999.
recognition (pp. 2001–2010). Zhang, P., Wu, H.-N., Chen, R.-P., & Chan, T. H. (2020). Hybrid meta-
Rivas, A. , Fraile, J. M. , Chamoso, P. , González-Briones, A. , Sittón, heuristic and machine learning algorithms for tunneling-induced
I. & Corchado, J. M. (2019). A predictive maintenance model settlement prediction: A comparative study. Tunnelling and Under-
using recurrent neural networks. In International workshop on soft ground Space Technology, 99, 103383.
computing models in industrial and environmental applications Zhou, C. & Tham, C.-K. (2018). Graphel: A graph-based ensemble
(pp. 261–270). learning method for distributed diagnostics and prognostics in
Romeo, L., Loncarski, J., Paolanti, M., Bocchini, G., Mancini, A., & the industrial internet of things. In 2018 IEEE 24th international
Frontoni, E. (2020). Machine learning-based design support sys- conference on parallel and distributed systems (ICPADS) (pp. 903–
tem for the prediction of heterogeneous machine parameters in 909).
industry 4.0. Expert Systems with Applications, 140, 112869.
Sarazin, A. , Truptil, S. , Montarnal, A. & Lamothe, J. (2019). Toward
information system architecture to support predictive mainte- Publisher’s Note Springer Nature remains neutral with regard to juris-
nance approach. In Enterprise interoperability viii (pp. 297–306). dictional claims in published maps and institutional affiliations.
Springer.
Schmidt, B., & Wang, L. (2018). Predictive maintenance of machine
tool linear axes: A case from manufacturing industry. Procedia
Manufacturing, 17, 118–125.

123
Journal of Intelligent Manufacturing is a copyright of Springer, 2023. All Rights Reserved.