The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

https://doi.org/10.1007/s00170-023-12158-7

ORIGINAL ARTICLE

Operations‑aware novelty detection framework for CNC machine

tools: proposal and application
Marco Macchi1 · Laura Cattaneo2 · Adalberto Polenghi1 · Danilo Bruno3 · Sofia Mosci3 · Fabio Ferriani4 ·
Roberto Caneva4

Received: 8 March 2023 / Accepted: 12 August 2023 / Published online: 1 September 2023
© The Author(s) 2023

Abstract
Digitisation offers manufacturing companies new opportunities to improve their operations and competitiveness in the market
by unleashing potentialities related to real-time monitoring and control of operating machines. Through condition-based
and predictive maintenance, the knowledge about the health state and probability of failure of the machines is improved for
better decision-making. Amongst them, CNC machine tools do represent a complex case from a maintenance viewpoint as
their operations are ever-changing and their high reliability brings to a lack, or limited set, of run-to-failure data. To address
the problem, the research work proposes an operations-aware novelty detection framework for CNC machine tools based on
already-in-place controllers. The framework is based on statistical modelling of the behaviour of the machine tools, namely
through gradient boosting regression and Gaussian mixture models, to identify the health state considering varying operations
through time. The proposed solution is verified on sixteen multi-axis CNC machine tools in a large manufacturing company.
The results show that the proposed solution can effectively support maintenance decisions by informing on the health states
while discerning between varying operations and abnormal/faulty states of interest. This solution represents a brick in a
cloud-edge-based industrial information system stack that can be further developed for shop floor-integrated decision-making.

Keywords Novelty detection · Condition-based maintenance · Prognostics and health management · Machine tool · CNC ·
Manufacturing

1 Introduction gather thanks to digitisation [1, 2]. Especially, on the shop

floor, new ways for production and maintenance manage-
The digital transformation is bringing new opportunities to ment are offered by advanced sensing solutions empowered
improve the business and operations of industrial companies. by data analytics to extract useful information [3]. This, in
New business models, better engagement of suppliers and turn, becomes essential to improve decision-making pro-
customers along the supply chain as well as enhanced opera- cesses that are no more only reactive, but rather proactive so
tions management are some of the benefits companies could as to anticipate future possible scenarios and promptly adjust

* Marco Macchi Roberto Caneva

[email protected] [email protected]
* Adalberto Polenghi 1
Department of Management, Economics and Industrial
[email protected]
Engineering, Politecnico di Milano, via Lambruschini 4/b,
Laura Cattaneo 20156 Milan, Italy
[email protected] 2
School of Industrial Engineering, Università Carlo Cattaneo
Danilo Bruno – LIUC, C.so Matteotti 22, 21053 Castellanza, Italy
[email protected] 3
CAMELOT Biomedical Systems Srl, Via al Ponte Reale
Sofia Mosci 2/20, 16124 Genova, Italy
[email protected] 4
Ansaldo Energia Spa, Via Nicola Lorenzi 8, 16152 Genova,
Fabio Ferriani Italy
[email protected]

13
Vol.:(0123456789)
4492 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

the current situation, if necessary [4]. Specifically, main- tools perform are time-varying in the light of the flexibility
tenance is perceived as one of the most impacted areas by of the operations required to make multiple product types;
digitisation, given the shifts from corrective and time-based (iii) CNC machine tools are highly reliable, which limits the
maintenance to condition-based and predictive maintenance availability of RTF data; and (iv) an integrated condition
[5]. These offer the possibility to have insights into the cur- monitoring system fit for the purpose of maintenance deci-
rent working state of an operating machine, its health state sion-making is not originally in place in extant machines.
and the remaining useful life (RUL) [6]. This information These characteristics lead to a challenging PHM application,
finally leads to improve maintenance performances in terms to be developed considering the following requirements:
of reduced downtimes, improved availability, increased PHM solutions should take into account the interactions
product quality as well as production performances, adapt- between subsystems and components of the machine tools
ing to contingent situations [7, 8]. The body of knowledge in order to improve the maintenance decision-making [17];
that collects all the steps (referred also as “levels”) necessary PHM solutions should be able to adapt to ever-changing
to build a system able to bring these benefits is often referred working conditions dictated by new part programmes intro-
to as Prognostics and Health Management (PHM) [9]. As it duced to satisfy continuously evolving production require-
is well-established, PHM can build on standards, such as the ments [20]; and PHM solutions should be capable to cope
ISO 13374-1 [10], to determine a methodological approach with any limitations concerning RTF and condition monitor-
for the PHM program development. ing data for maintenance decisions [14].
The cornerstones of PHM reside in suitable sensing tech- Overall, the reduction of options in terms of AI and sta-
nologies and data analytics, both fit for purpose. Sensors tistical algorithms is a consequence, and this research work
and transducers must be installed and configured to collect proposes to employ supervised learning models within the
condition monitoring data from the machine components, novelty detection framework, leaving aside diagnostic and
which is key to improving knowledge about failure modes prognostic analyses.
the machine may experience. Furthermore, work orders
and manually inputted data are essential to combine condi-
tion monitoring data from the field with events recorded by 1.1 Research scope and objective
operators. Once the data are available, proper multi-nature
algorithms should be put in place to transform raw data into The research scope this work addresses with peculiar con-
useful information for decision-making. Algorithms may be sideration is the so-called state detection (SD) level of the
either physics-based, artificial intelligence (AI) or statistical PHM process, where machine states need to be identified.
[11]; recently, also hybrid models, mainly based on physics- In this scope, the assumption is to rely only upon the data
informed AI algorithms, are gaining momentum [12]. fed by the operations run by the machine tool and made
Notwithstanding its potentialities, PHM suffers from available by means of controllers (such as CNC or PLC) or
some issues. Firstly, PHM program development should sensors and transducers already-in-place for the machining
exploit run-to-failure (RTF) data: these mainly refer to event process, without any need of integration with a condition
data that allow labelling condition monitoring data with monitoring system fit for maintenance tasks such as diagnos-
respect to the real states of the operating machine, finally tics or prognostics. Therefore, at the SD level, the research
distinguishing faulty states (complete or partial faults) from objective this work pursues is the development of a novelty
healthy state; nevertheless, this potentiality is limited when detection (ND) framework for CNC machine tools enabling
RTF data are not, or scarcely, available [13, 14]. Besides, three major tasks: to identify when a new operation is car-
even condition monitoring data may be not available if a ried out featuring different working conditions, to recognise
proper monitoring system fit for the purpose of mainte- if a failure is going to occur by means of the detection of
nance decision-making is not put in place, or a monitoring an anomaly (alias partial fault), and, eventually, to classify
campaign has not yet been established [15]. A further issue the health state of the machine, either being in healthy or
relates to the varying working conditions that may result unhealthy state.
from the intrinsic flexibility of the operations made by the In detail, the developed solution should consider the capa-
operating machine, which could undermine the reliability of bility to recognise ever-changing operations and to guaran-
the information provided by PHM tasks [16]. tee ND adequate for the maintenance decision-maker. The
All these issues are experienced in the case of extant CNC industrial viability is demonstrated through the application
machine tools. Indeed, CNC machine tools represent a com- to a fleet of sixteen multi-axis CNC machine tools of a large
plex case where to apply PHM programmes, for different manufacturing company in Italy. Indeed, the challenge faced
reasons [17–19]: (i) CNC machine tools are an ensemble in this research work is to improve the capabilities of old,
of electro-mechanical and hydraulic systems that interact less sensitised CNC machine tools so as to embrace the
with each other; (ii) the part programmes CNC machine Machine tool 4.0 paradigm [21].

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4493

The remainder of the paper is so organised: section 2 on the current state information; prediction may be trig-
revises current PHM approaches applied to CNC machine gered as soon as a novelty comes across so as to identify
tools, giving a particular emphasis to ND; section 3 depicts future trends [11, 29].
the proposed solution, section 4 introduces the case com-
pany and section 5 describes the application to the case; Overall, ND represents a critical point in the develop-
section 6 then provides discussion and lessons learnt from ment of effective PHM programmes since diagnostics and
the industrial experience; finally, section 7 draws some con- prognostics analyses are enabled by the identification of the
clusions and envisions future works. actual states in which the equipment/asset or related subsys-
tem/component is.

2 PHM program and novelty detection 2.2 Novelty detection applied to CNC machine tools
for CNC machine tools
A pool of scientific articles published in the last few years
The advent of PHM has been promoted by the availability is hereinafter reported in order to shed light on the research
of sensors and increased computational capability, ena- trend on ND applied to CNC machine tools; the pool excludes
bling a shift from “fail-and-fix” to “predict-and-prevent” those works focused on diagnosis and prognosis only.
[22], whose development is strongly based on the ISO Recently, in Schlagenhauf and Burghardt [30], an expert
13374-1 [9]. This norm structures the process into six lev- system is realised with the aim of detecting defects on a ball
els: data acquisition, data manipulation, state detection, screw drive through convolutional neural network (CNN) on
health assessment, prognostic assessment and advisory images. Also, the authors realised a prediction model able
generation. to forecast the area of the defect so as to anticipate future
The applicability of the PHM process is independent of evolution. Also in Denkena et al. [31], ball screw anoma-
the specific asset or component under analysis, showing lies are identified through a semi-supervised ND algorithm
validity throughout different indenture levels [11, 23–25], based on an ensemble approach, built on local outlier fac-
including the equipment/asset itself, like a CNC machine tor, k-nearest neighbours, angle-based outlier detection
tool, a centrifugal pump, or an industrial fan, and the sub- (LOF-kNN-ABOD). These are selected since no previous
systems/components, such as bearings, drive belts, or rotors. assumption about data distribution is needed. Experiments
Moreover, the reviews propose specific insights on each level are carried out to understand the variation of performance
of the ISO 13374-1, sometimes showing different but com- of the ensemble approach with a varying set of features from
patible aggregation of the levels. the measured variables, and the results show that frequency
domain-related features outperform time domain ones for
2.1 Role of novelty detection in the PHM program ball screws condition monitoring.
Another relevant component for CNC machine tools, and
ND can be defined as the “task of recognising that test data in general for any machine including rotating parts, is the
differ in some respect from the data that are available during bearing. In the research by Wu et al. [32], bearing faults
training” [26]. With respect to the PHM process, as defined are analysed. To this end, CNN is used to detect and clas-
by ISO 13374-1, this task applies to the SD level. The states sify bearings faults together with variational mode decom-
in which the machines could be are identified and then, dur- position (VMD) and autocorrelation function (ACF) as
ing machine operations, if new data are significantly differ- they allow for the reduction of the number of layers in the
ent from previously identified patterns, a “novelty” is alerted CNN. In Goyal et al. [33], kNN and WkNN (weighted kNN)
so that maintenance technicians and managers could take are used to detect and diagnose bearings faults using non-
proper actions. contact laser-based vibration sensors instead of accelerom-
The importance of the SD level (and ND) in PHM pro- eters. The work supports that vibration remains one of the
gram development is recognised for diagnostic as well as main diagnostic signals for CNC machine tools even though
prognostic purposes: improvements are foreseen both on the side of algorithms
and on the physical sensors.
• ND is relevant for diagnostics since the identified new pat- Regarding the hydraulic subsystem, Bernini et al. [34]
tern of data, i.e. a novelty, enables the diagnosis of what set up an experimental campaign based on a digital twin
has happened [27, 28]. Also, in case of missing failure that reproduces the behaviour of the targeted CNC machine
data, ND and abnormal state labelling almost coincide as tool subsystem to generate healthy and faulty data. A pool
anything new is a potentially damaging state [14]. of several algorithms is used to detect, diagnose and pre-
• ND is necessary also for prognostics since the variables’ dict hydraulic subsystem faults, like CNN, random forest
prediction and related RUL estimation, highly depend classifier (RCF), quadratic discriminant analysis (QDA)

13
4494 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

and AdaBoost classifier. (ADB). The application shows that without the need to perform any simulation. Nonetheless,
there is no unique answer for ND, but some algorithms per- differences between machines like different operations to
form better than others for different components: QDA for a perform and different signal properties prevent success at
pump, and CNN for sensors and valve faults. the current state-of-the-art. Indeed, the review by Baur et al.
Instead of relying on AI for ND, Du et al. [35] adopt [42] confirms the above-mentioned trends; they made clear
proportional hazards modelling to model failure events and that at the moment the implementation of PHM in all or
condition monitoring data, namely vibration, for electro- some of its steps is subordinated to the knowledge of faults,
spindle. Proportional hazards modelling allows extracting the machine experiences over its history. This somehow
health indicators that are controlled in real-time thanks to prevents the full exploitation of such methods in industrial
statistical process control (SPC) to trigger maintenance environments if no prior knowledge of faults and related
actions in case of anomalies. measurements is available.
As one of the most critical components, the spindle has Table 1 summarises the literature findings, synthesising
received a lot of attention in CNC machine tools as it guar- some relevant insights, grouped by relevant meta-variables,
antees product quality [36]. In this regard, many attempts namely the indenture level targeted by the ND solutions,
were made to improve monitoring and controlling of the the models adopted in such solutions, and the application
spindle for chattering, tool wear prediction and thermal error contexts where the solution is tested.
compensation [37]. Vibration monitoring is widely adopted Overall, several are the research streams and approaches
both in continuous and discrete checking modes, to later for ND, and PHM in general, regarding CNC machine tools.
compare with predefined thresholds defined as a function Firstly, given the complexity of such machines, specific sub-
of operations requirements or manufacturers’ manual [38]. systems, such as hydraulic or electrical subsystems, and com-
Recent reviews show current advancements and future out- ponents, such as bearings and ball screws, are usually analysed
looks for spindle-related applications. For example, a com- one at a time. There are efforts in relating failures of different
prehensive review of AI and ML has been developed by subsystems/components, but this research stream is at its early
Soori et al. [39], while the review by Kuntoğlu et al. [40] stage [43]. Secondly, feature extraction and selection are vital
investigated which signals are currently used for advanced to guarantee the downstream algorithms so that ND performs
applications, especially for spindle monitoring. well. Hence, signals, especially vibrations, are transformed in
Overall, the picture resulting from the selected scien- time, frequency and time-frequency domains, and features are
tific literature shows that CNC machine tools are widely selected based on the current fault to be detected [44]. Lastly,
addressed in all their units and components, ranging from AI algorithms are currently playing the lion’s share for ND of
ball screw and bearing diagnostics to the identification of machine tools. Specifically, CNN is often engaged due to its
anomalies for the hydraulic system and machining unit, ability to approximate even complex phenomena without the
focusing on the spindle. Nonetheless, most works rely on need to know the underlying failure mechanism [45]. Ensem-
laboratory experiments so as to have the entire set of faults ble and hybrid approaches are also used in order to combine
available, resulting in many applications being supervised the advantages of multiple algorithms and outperform the
in nature. In case these data are available, the proposed single, standalone application [17].
works enable deep investigation on the health state of the
CNC machine tools, with information related to possible
causes (diagnostics). On the other hand, in case of miss- 2.3 Concluding remarks and gaps
ing a priori knowledge, there are still gaps to overcome,
with some tentative to transfer knowledge from one CNC Despite the promising trends and approaches to ND evident
machine tool to another, as in the work by Denkena et al. in literature findings, there are some gaps that are needed to
[41], which is focused on how to transfer ND thresholds be filled in. Amongst the gaps, the following appear the most
specific for a certain CNC machine tool to similar machines relevant to remark:

Table 1  Summary of selected literature findings

Meta-variable Finding

Indenture level Developed ND solutions work at subsystem or component as indenture level; also, the influence of faults between multiple
subsystems/components is not well studied
Adopted model Current ND solutions embrace both statistical and AI algorithms; however, as the phenomenon’s complexity increases,
most authors look for AI as a way to approximate the failure mechanism
Application context Usually, ND solutions are developed and tested in controlled environment, as laboratories, or are developed based on
numerically generated dataset via multi-physics simulation

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4495

1. AI algorithms, such as CNN, require a huge amount of aspects rather than model performance, like reliability
data to be properly trained. However, most of the time, and usability of the provided information to the mainte-
RTF data are not available in the industry [13] and not nance decision-makers. Therefore, stemming from the
affordable to be replicated in practice or generated by gaps identified in the scientific literature and accord-
simulations [34]. Then, opportunities offered by AI ing to the authors’ experience in industrial contexts, the
algorithms appear at least challenging in their full reali- goal of this research work is to realise a framework with
sation when data are scarcely available. associated models, organised in an integrated, indus-
2. There are few works that consider the wide set of avail- trially viable solution for ND of CNC machine tools.
able data from CNC machine tools. Condition moni- This research considers CNC and PLC data also to tune
toring data such as vibrations and temperatures are the the PHM with the working conditions of the machines,
most used, but CNC and PLC could provide valuable besides supporting the ND task aimed at fault detec-
information regarding the part program and tool num- tion as required by maintenance decision-making. It is
ber that allow for improving algorithm performance [46, worth recalling that diagnostic and prognostic analyses
47]. Thus, there is a lack of emphasis in the extant scien- are considered out of scope in this research work and the
tific literature on how to better tune PHM thanks to the proposed operations-aware ND framework, whereas fault
knowledge of the working conditions of the machines. detection is addressed with the main purpose to sup-
3. The industrial viability of proposed solutions needs to be port information for the decision-maker that an anomaly
better investigated since, at present, datasets are generated arose and a functional failure is then incipient.
numerically through multi-physics simulation or from To tackle every aspect of the proposed solution, the
laboratory tests. Few are applications considering CNC remainder of the section is organised as follows: subsec-
machine tools working in a real industrial environment. tion 3.1 describes the proposed framework, the ground
4. Most of the work concentrates on diagnosis, usually theory it refers to, the overall structure and the peculiari-
referred to as fault detection and diagnosis (FDD) ties that make it novel with respect to the extant state of art
rather than prognosis. Indeed, given labelled or unla- solutions; subsections 3.2 and 3.3 deal with the SD level
belled data, many studies focus on the application of of PHM, clearing out the novel part of it; subsection 3.4
supervised or unsupervised approaches to identify which presents the health assessment and related health indicator
failure modes are evolving or which failure has been definition; and finally, subsection 3.5 depicts the working
experienced. Prognostic assessment may be empowered flowchart of the framework that enables it to be operative
in future developments. in an industrial environment.

In this research work, the focus is on the first three gaps,

while the last is out of scope, but worth for the future scien- 3.1 Overview of the framework
tific effort. Amongst the focused gaps, it is worth remark-
ing that the industrial viability of the method is a key fea- The ground theory of the proposed framework is provided
ture addressed by this work. Firstly, it relates to the need by the ISO 13374-1 [10], which allows structuring the
to demonstrate implementation of PHM in CNC machine solution around an agreed-upon flow to realise the PHM
tools; secondly, it implies the adoption of a real setting in program. The hypotheses to which the framework and the
the actual evolution of the production requirements, enabling related algorithms are constrained are listed hereinafter as
an industrial testbed inclusive of the variations of the actual long as the effects on the modelling decisions.
working conditions of the machines, which is a key item
under concern in the present work. • RTF data are assumed not to be available (or to be
scarcely available). It prevents embracing pure super-
vised learning built upon failure labels, as failures are
3 Proposal of the operations‑aware novelty not (or, at least, are poorly) recorded. Then, specifi-
detection framework for CNC machine cally for ND, this learning is sometimes referred to as
tools “semi-supervised” [26] as, on the one hand, it leverages
upon a kind of unsupervised modelling as there are no
The evidence from the literature shows several pre-established labels, but, on the other hand, the nor-
approaches to ND for CNC machine tools, to further mality is assumed to represent the healthy state, which
enable diagnostic and prognostic analyses. However, enables a labelling process, therefore a kind of super-
what emerged is that there are few examples that rely vised modelling. Within the ND framework, specific
on CNC and PLC data only and that show the indus- models—gradient boosting regression and Gaussian
trial viability of the proposed solutions, including other mixture models—are then implemented.

13
4496 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

• No condition monitoring data are adopted at the SD The framework is presented in Fig. 1, where the ISO
level, as it is assumed a limited machine retrofitting, 13374-1 is taken as a reference and extended in the level
without the need to integrate a condition monitor- named “state detection”. Health assessment is considered
ing system fit for maintenance purposes; hence, “tra- only as a matter of labelling the health state of the CNC
ditional” techniques for CNC machine tools, such as machine tool, without any effort towards diagnosis; also,
vibration analysis, are not applied. prognostic assessment is skipped as it is not within the scope
• Multiple machining operations are performed by the of this work.
CNC machine tool and the entire set of part programmes The first two levels entail the collection of data from sen-
is supposed not to be entirely known a priori as prod- sors, transducers and manual inputs, and the pre-processing
uct requirements may change; therefore, the developed and processing of data to guarantee high-quality data flow-
models must consider the variability of part programmes ing towards the state detection data-driven models. Then, the
to be realised so as to provide reliable information to SD level is implemented, which is focused on the developed
maintenance decision-makers. ND models of whatever nature to determine how many and
• CNC and PLC data are available so that operations, which states the machine may experience. Within this level,
determining the working conditions, may be recognised. the newly added step is designated “operations consistency
Particularly, CNC and PLC data available are those char- checking” (righthand grey-coloured box in Fig. 1); this is
acterising the part programmes and operations, such as combined with the development of “fault detection” (lefthand
active program, program status, actual spindle speed, grey-coloured box in Fig. 1). Both are data-driven models, fed
nominal spindle speed and actual current absorbed by by only CNC and PLC data, namely part programmes, opera-
the spindle. The models developed should integrate such tions ID, tool number and other program-related variables
CNC and PLC data to achieve relevant information for from the former, and linear speed, rotational speed, current
the maintenance decision-maker. drive load and other process-related variables from the latter.
The “operations consistency checking” is in charge of
It is worth remarking that ND is hereinafter focused on assessing if the results derived by running the data-driven
the spindle and axes as critical equipment of the machine model could be used or not to judge maintenance-related
tool. Nevertheless, the framework, in its inherent logic, decisions. In particular, only when the operations are compa-
could be applied to other critical subsystems/components rable (i.e. training data are comparable with testing data), the
that are monitored and for which an ND model exists or fault detection and health assessment may take place to then
could be developed, to finally come up with the definition support maintenance decisions. Otherwise, if the operations
of the health states of the CNC machine tool. are not comparable, the results are not directly adopted for a

Fig. 1  Proposed novelty detection framework for CNC machine tools

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4497

maintenance reason; however, they are used to inform about 2. Machining operations could be compared by analysing
a newly performed operation with respect to the history the distance between the distributions of the representa-
of the machine under analysis. In this way, failure-related tive variables. This enables to determine the statistical
information can be trusted by the decision-maker as they do comparability.
really represent an incipient functional failure and not pos- 3. Machining operations that are statistically compara-
sible changes due to production requirements, e.g. in terms ble determine the same working condition, which is
of material or processing parameters. This is assessed sta- assumed to lead to the same effects on the CNC machine
tistically as explained in subsection 3.2. Indeed, the “opera- tools and their degradation, and so health state.
tions consistency checking” step is not in charge of assessing
model accuracy and precision as this should be done during Therefore, as anticipated in Fig. 1, the operations con-
training and testing of the “fault detection” model itself. The sistency checking corresponds to proving the similarity
novelty this framework claims does not refer to the single between operations by comparing the distributions of the
data-driven statistical model, but relies on the combination representative variables, enabling robust information about
of those models for improved fault detection based on opera- health state, and incipient functional failure, delivered to the
tions consistency checking, as better described hereinafter. decision-maker.
The idea is detailed in Fig. 2, which also introduces
an index, between 0 and 1, that brings information about
3.2 State detection level with a focus on operations the operations consistency checking. This index is called
consistency checking “Affinity Index” and refers to the measurement of similar-
ity between varying operations through time. The higher
The SD level represents the cornerstone of every well-con- the Affinity Index, the higher the similarity between two
ceived PHM program. The recognition of the state is sub- operations, and vice versa. For the sake of simplicity, only
ordinated to the identification of the machining operation two Gaussian distributions are drawn in Fig. 2, whereas the
the CNC machine tool is undergoing [48]. Nevertheless, consistency check in its full deployment adopts a mixture of
given the specific goal of FDD, most studies from the lit- Gaussian distributions.
erature concentrate on the development of well-performing For a practical deployment, the solution relies on the
algorithms for specific failure modes/causes, whereas few definition of a threshold, designated as THaffinity, in order to
applications consider more than one machining operation at discriminate between comparable and not comparable opera-
a time. Besides, the current state-of-the-art shows that most tions; the reader should refer to subsection 5.3.1 for informa-
of the applications target ND, or equivalently fault detection tion on how to set THaffinity. Being based on statistical analy-
in FDD, of CNC machine tools in a controlled environment sis, the consistency check then verifies if similar machining
or leveraging upon simulated datasets from multi-physics operations are performed by the analysed machine over dif-
models. Therefore, to deal with the variations from a real ferent monitoring periods, then bringing reliable and usable
industrial context, the current proposal aims at accompany- information for maintenance purposes.
ing the application of the data-driven fault detection model
with automatic checking of machining operations consist- 3.3 State detection level with a focus on fault
ency; this aims to inform the decision-maker if the provided detection
information is usable or not, for maintenance purposes.
Given the nature of the CNC machine tools, i.e. character- The development of the fault detection model follows the
ised by the ability to perform multiple operations, a relevant approach by Pimentel et al. [26], which is inspired by sta-
assumption must be stated: machining operations whose rep- tistical process control: if the difference between the pre-
resentative variables values are closer to each other have the dicted value by the ND model defined for fault detection,
same effect on the health state of the machine tools. The impli- and the current measured value monitored from the running
cations of this general assumption are hereinafter detailed: machine, exceeds a predefined threshold, then an anomaly
has arisen. This is meant as a partial fault, activating a con-
1. Machining operations should be described with the dis- dition-based maintenance intervention.
tributions of the variables and machining parameters To determine the predicted value, alias the response, a
(such as rotational speed of the spindle, feed rate, and gradient boosting regression (GBR) is selected as the ND
axis movement). These distributions can be known after model. The selection of GBR comes from a comparison with
one or more monitoring periods. They also do take into other statistical models, namely, linear regression and sup-
account exogenous variables not directly collected, such port vector regression. GBR is indeed recognised as one of
as material, which influences the operations, and so the the most popular regression algorithm for tabular datasets.
variables, measured at runtime. Its capability lies in its ability to effectively capture complex

13
4498 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

Training data Real time data

HIGH AFFINITY INDEX:
The undergoing operation is comparable
to the ones used for training.
Results of the model could be used for
health assessment.

Training data Real time data

LOW AFFINITY INDEX:
The undergoing operation is not
comparable to the ones used for training.
Results of the model could NOT be used
for health assessment.

Training data distribution on Current value of the data

which the models were trained collected from field

TBN: the example above is based on two gaussian distributions for which statistical testing may be performed (e.g. paired t-
test). They are just used to explain the concept of Affinity Index, since each operation is characterised by multiple distributions
and the comparison is based on mixture of gaussian distributions.

Fig. 2  Evaluation of Affinity Index by means of Gaussian mixture models

and nonlinear relationships between the target variable and [26]. Grubbs identified that the mean plus three times the
the input features. Moreover, it offers remarkable usability standard deviation represents a well-established definition
by handling missing values, outliers, so mitigating the need for the novelty thresholds [51], as confirmed by industrial
for any special treatment [49, 50]. experiences [52, 14]. It is anyhow important to underline
The goodness of the GBR model is validated via that the novelty threshold definition is an open challenge and
the coefficient of determination R 2, which is generally depends on available a priori knowledge about probabilistic
expressed as in Eq. 1. distributions of the signals [26].
� � Here in the present work, ND also comes the contribution
(current value − response)2 of the Affinity Index: the ND score has validity only when-
∑
2
R = 1− ∑
(current value − mean(current value))2 ever data comes from comparable machining operations, as
(1) summarised in subsections 3.4 and 3.5.
The model is trained and then tested over historical data
3.4 Health assessment level with health indicator
gathered during past operations of the machines in healthy
definition
state. Then, the model is used at runtime to evaluate the ND
score, looking for anomalies, i.e. partial faults, by comput-
Once the SD level has been formalized, it is possible to
ing the difference between the predicted value (so called
assess the health state of the CNC machine tool, or, more
response) of the target variable and the measured actual
precisely, of the monitored subsystems and for which ND
value (alias the current value as it is measured), as in Eq. 2.
models are available. Since no RTF data are available,
score = current value − response (2) a ND learning approach is implemented [26]: only the
healthy state is characterized and, as soon as a novelty
In this way, the score acts as the ND metric later used to (anomaly) arose, then a new degraded state (abnormal
assess possible anomalies. Indeed, once the score has been state) is experienced by the machine.
evaluated on the test data, its distribution is computed, such In practice, it is also worth observing that true-neg-
as its standard deviation σS. These values will be used in the ative or false-positive may happen due to uncontrolla-
ND algorithm to define when an anomaly arose, i.e. when ble variations, as the operations on the machine are not
the actual value is different from the response more than 3 deterministic and the production operators may change
times the σS (which means the score, measured during the parameters due to contingent situations. As such, mainte-
machine monitoring, is higher than 3 times the σS). This nance decision-makers may be mistakenly alerted along
decision is commonly applied for novelty detection purposes the day even for several reasons, not necessarily requiring

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4499

a maintenance intervention. Therefore, the modelling deci- ing hypotheses, no RTF data are available, hence, it is
sion is to smooth this behaviour by setting a proper fre- only possible to set these thresholds according to the risk
quency of alerts as a relevant feature for decision-making, the company would take. Therefore, the decision is to
in the case under study daily frequency is decided. The model these thresholds as follows, considering healthy,
daily frequency of alerts falert has been modelled as fol- abnormal, or faulty states for discretizing the behaviour
lows in Eq. 3: of the CNC machine tool:
Nalerti a. Healthy state, when HI < μ + 3σ, where σ is the
falerti = (3) standard deviation of HI and μ is the mean value.
Nscorei
b. Abnormal state, when HI ≥ μ + 3σ and HI < μ + 4σ.
where: c. Faulty state, elsewhere.

• Nalerti is the daily number of alerts during the day i; This builds on similar considerations expressed for the
• Nscorei is the daily number of evaluated scores during the day i. novelty threshold definition (subsection 3.3). Besides, it
should be noted that the number of health states is case-
Building on the score definition (Eq. 2), the alert is dependent; the authors assume a three-state model so to
released every time the score exceeds the 99.7% confi- include a “warning” period described by the abnormal state,
dence interval of the response of the ND model for fault before the failure (faulty state).
detection, which corresponds to the anomaly occurrence. The information on the health state is then provided to the
It means that the alert is released whenever the difference maintenance decision-maker in order to judge the decision
between the measured and predicted values is higher than to make. Indeed, the information is not provided directly,
3 times the standard deviation of the error evaluated on but is firstly filtered by joining together the Affinity Index
the test dataset. as expressed in Table 2.
The health state indicator (HI) is based on the daily
alert frequency and is modelled as its moving average 3.5 Working flowchart of the proposed framework
(MA(k)) as in Eq. 4:
∑k The working flowchart, characterized by the logical
f sequence of data management and computational steps, and
HI = MA(k) =
i=0 alerti (4)
k decision points, completes the specification for the PHM
program application according to the proposed approach in
where k is the order of the moving average MA(k). the case of CNC machine tools. Therefore, Fig. 3 reports the
To properly define the health indicator so as to be useful for overall flowchart, also drawing, on its left-hand side, the ISO
decision-making, two additional open points need to be solved: 13374-1 used as reference structure to build the framework.
Starting from the set of CNC and PLC data, the first step
1. The assessment of the order k of the moving average. The is the data pre-processing, integration, and homogenisation.
selection depends on the required promptness of change This step is fundamental so that the data later analysed by
of the indicator while avoiding useless spikes, thus, it the ND models are of high quality. Especially, data fusion is
should result after analysing real data available from the central as all data samples must be related to the very same
shop floor to properly run the health assessment. timestamp. Later, the data are used to evaluate the score,
2. The definition of the thresholds of the health indica- for fault detection, and the Affinity Index, for operations
tors so as to label the values of HI in a discretized way, consistency checking. Both results are used in the first two
treatable for decision-making. According to the work- decisional points:

Table 2  Information provided to the maintenance decision-makers according to Affinity Index

Health indicator (HI)
HI < μ + 3σ HI ≥ μ + 3σ and HI < μ + 4σ elsewhere

Affinity Index Not comparable operations (< THaffinity) The operation/s performed by the CNC machine tool is/are not comparable to those
known a priori. No information on the health indicator is available.
Comparable operations (≥ THaffinity) The operation/s performed by the CNC machine tool is/comparable to those known a
priori. Information on the health indicator is available.
Healthy Abnormal Faulty

13
4500 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

Fig. 3  Working flowchart of the Data

ISO 13374-1 levels

CNC/PLC data gathering
proposed framework acquisition
Data Data preprocessing
manipulation and homogeneization
Operations
Fault Evaluation of Affinity
Evaluation of Score consistency
detection Index
checking

State NO Affinity Index NO

Score trend?
detection ≥ THaffinity
YES
Release & storage YES
of the alert

Evaluation of Health
Indicator

Heath
assessment Health Indicator above NO
thresholds?

YES
Advisory
No alert release Unhealthy state Healthy state Different operation
generation

• The score for fault detection is analysed so to identify 4 Company overview and machines
possible anomalies. If the score is within the pre-defined under study
thresholds (three times the standard deviation, building
on Eq. 2), then there is no release of the alert and the The proposed framework for operations-aware ND of CNC
monitoring of the machine still goes on; otherwise, if machine tools, is tested in the real case of a large manufac-
the score exceeds thresholds (more than three times the turing company. This section 4 presents the company and
standard deviation, building on Eq. 2), an alert is released the machines under study, while section 5 describes how the
and stored to be later used for the feature evaluation; the framework is applied in this context.
feature is then evaluated (both alert daily frequency, see In fact, as industrial deployment, the solution has been
Eq. 3, and moving average HI, see Eq. 4), this happens if developed during a three-year collaborative project based
and only if also the Affinity Index is verified. on action research where multiple parties are involved, such
• The Affinity Index is compared with a threshold THaffinity (in as an Italian manufacturing company, researchers, and small
this case, set to 75%, that is the probability that the gathered and medium-sized enterprises (SMEs) each with specific
data are similar to those used to train the model is higher than objectives. The Italian manufacturing company produces
75%, please see subsection 5.3.1 for further information). If critical components for gas and steam turbines, later used in
the Affinity Index is higher, then the score could be used to the power generation sector to generate electric energy for
evaluate the alert daily frequency and the health indicator for several applications. The company has a long-lasting story
the decision-maker; otherwise, the information is not thrown in this field, with more than 100 years of experience in the
away, but it is used to inform the decision-maker (mainte- design, production, and maintenance of turbines. The prod-
nance technician or manager), that the current operation is ucts are sold all over the world, making the company one of
too different from those within the training dataset of the the global leaders in this sector.
model, so nothing could be said on the health state. The Italian plant where the project was performed
includes the machine tools under study, i.e. those in charge
Every time both conditions are satisfied, an alert is released of producing the blades for the gas/steam turbines. The qual-
and stored to evaluate the alert daily frequency. Then, the health ity related to this kind of product must be at the highest
indicator is computed and, if the feature goes over the average possible standards; therefore, the health state of the machine
value plus 3 times the standard deviation (HI ≥ μ + 3σ), the tools producing the blades must be kept under control and
unhealthy state is identified, and an alarm is released to trigger any possible anomaly be noticed to avoid waterfall effects
a call out to the maintenance decision-maker. Otherwise, the on the quality of the final assembly, i.e. the turbine. Specifi-
healthy state is set, and no alarm is released. cally, the CNC machine tools under study are the following:

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4501

• 4-axis CNC machines tools and 5-axis CNC machine 5 Application of the operations‑aware
tools (5 and 1 machines, respectively, for a total of 6 novelty detection framework for CNC
machines) in charge of realising the blade roots; machine tools
• 5-axis CNC machine tools (total of 8 machines) in charge
of realising the blades; and This section illustrates the development of the solution,
• Newly bought 6-axis CNC machine tools (total of 2 following ISO 13374-1 and applied to the company case.
machines) in charge of realising the blades. Some insights on the data acquisition are provided in sub-
section 5.1; the data manipulation is described in subsec-
The implementation of the proposed solution for tion 5.2; the state detection is presented in subsection 5.3;
operations-aware ND required the definition of a robust and finally the health assessment in subsection 5.4.
methodology, grounded on the ISO 13374-1 and PHM
body of knowledge. Indeed, it was within the goal of the 5.1 Data acquisition
3-year project to realise a PHM solution almost from
greenfield, i.e. machines were not connected, and no As one of the main assumptions of this research work, the
platform was available for data storage and analysis. only available data from the CNC machine tools are those
Therefore, the solution described in section 5 required coming from the CNC and related PLC controlling various
some propaedeutic activities. subsystems. Hence, for streamlined data acquisition, those
Amongst them, it is worth remarking on the need for data must be automatically read and made available for fur-
criticality analyses, in this case, carried out by a failure ther use, i.e. data storage and analysis. Therefore, the CNC
mode effects and criticality analysis (FMECA), aimed to and the PLC are connected via a gateway that is in charge of
understand the machines and their critical items; moreo- collecting the data and sending them to the cloud after some
ver, the development of an ICT architecture was required computations on edge, as represented in Fig. 4.
in order to collect, elaborate and store data. In particular, Given the constraint on the data effectively stored in the
it is worth pointing out that the results of the FMECA cloud, higher sampling frequencies required an aggregation
show that the spindle and the axes are those with the computed on edge.
highest criticality in terms of risk priority number (RPN) Specifically, the setup of a correct sampling frequency
considering occurrence, severity and detectability. Other for each variable is of paramount importance to guarantee
details of the propaedeutic activities will not be treated the correct functioning of the solution. This is combined
as they are out of the scope of the insights provided in with the identification of meaningful features from the meas-
this research work. ured variables, as obtained in the data manipulation phase.

Fig. 4  Data acquisition system

13
4502 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

Hence, an iterative approach was established between the Overall, the final value of the working state percentage
data acquisition and the data manipulation, to identify the aims at guaranteeing high-quality data for the actual machin-
best trade-off between the number of variables/features and ing operation, assuming that the higher the working state
their sampling frequency. percentage, the more representative the description of the
As a result of the iterations, the sampling frequencies machining operation and its impact on the health state.
were fixed in the range from every second to every 2 min Building on high-quality data, the identification of mean-
depending on the variable, while the aggregation was ingful variables/features requires typical methods such as
required at 2 min due to the constraint in the cloud storage. dimensionality reduction looking for constant-value vari-
ables and highly correlated variables. It is worth mention-
5.2 Data manipulation ing that, for the correlation analysis, a combined physics/
data-driven analysis was performed in this case, adopting:
In PHM programmes, data manipulation typically entails (i) a physical analysis based on empirical rules that relate
data preparation, data cleaning and feature extraction and manufacturing variables, like, e.g. the feed rate and the spin-
selection. dle speed; and (ii) a data-driven analysis based on the linear
To enable high-quality data, an important data cleaning correlation coefficient. Overall, this enables the selection of
was made. Hereinafter, it is worth discussing one specific relevant variables, for each CNC machine tool: the current
objective: the isolation of only active operation (i.e. metal of spindles and axes, the drive load of spindle and axis, the
cutting) by the CNC machine tool. In fact, during the execu- rotary speed of the spindle and the absolute velocity of the
tion of a part program, the machine tool undergoes both axes, together with the operation identifier.
active and non-active operations from a product perspec-
tive: an active operation implies contact between the tool- 5.3 State detection
ing head and the workpiece, while a non-active operation
refers to correct positioning or tool changing. Given that State detection runs the ND data-driven models for oper-
non-active operation may impact the model performance and ations checking and fault detection. These entail the
the constraint of 2-min aggregation time for cloud storage, it following:
was decided to use an ad-hoc variable named “working state
percentage” to estimate how much time, in a 2-min interval, 1. Definition of an index to measure the affinity/similarity
the CNC machine tool spent in active operation or not. The between the operations; and
working state percentage ranges from 0 to 1, where 1 means 2. Definition of a score to measure the existence of an
that all gathered data points within the 2 min represent active anomaly, i.e. a functional failure.
operation. The evaluation of the working state percentage
works as follows: The following subsections 5.3.1 and 5.3.2 provide
insights on the application, completing what was presented
1. For each data point, the “working state” is evaluated, in subsections 3.2 and 3.3.
which is based on the analysis of the CNC program
string and its specific codes: if the CNC machine is 5.3.1 Definition of the Affinity Index
in active operation, working state is equal to 1, 0 else-
where. Knowing that the CNC machine tool is in active operation—
2. The ratio between the number of data points with a as obtained from the working state percentage—does not
working state equal to 1 over total data points provides prevent obtaining a result that is not coherent if the new
the working state percentage. operation is significatively different from those in the train-
ing dataset. As reported in Fig. 5, the probability density
The working state percentage threshold, above which the functions of three relevant variables collected by the CNC
interval was considered good for the prediction of the health machines, namely, the spindle drive load, the spindle actual
state, was 0.7, i.e. most of the time the CNC machine tool speed value and the current on one of the axes, depend on
was in active operation. Furthermore, two additional con- the specific operation the CNC machines performed over
straints were superimposed to guarantee the correct func- the acquisition time.
tioning of the model: within the 2-min aggregation inter- Therefore, it is of utmost importance to compare, time
val, there should be no more than 5 rapid movements and by time, if the gathered values are comparable or not with
2 tool-changing operations. These specific thresholds for those present in the training dataset as this should allow reli-
rapid movements and tool-changing operations have been able information for maintenance decision-making: if the
fixed after numerous tests and considering the manufactur- operations are comparable, information on how the machine
ing technology knowledge background. behaves could be extracted in order to make the health

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4503

Fig. 5  Distribution of values for three relevant variables (same legend for all three plots, see bottom part)

assessment; otherwise, it is known only that the current opera- two-interval data point (x-axis). Instead, on the y-axis of the
tion is different from historical ones, whereas it is not possible bottom plots, there are operation IDs for each two-interval
to draw any conclusion on the health state of the machine. data point (x-axis). Therefore, it is possible to read, for each
Therefore, building on the general idea already presented operation, the related Affinity Index, thus shedding light on
in Fig. 2, the Affinity Index has been implemented through the comparability of the operations.
the following steps: Indeed, it is possible to see that:

• A statistical model of the possible states that the CNC • Operation _N_N93Z99_S01_MPF is present both in the
machine goes through during the training period is cre- training and in the test dataset and the Affinity Index is
ated from the data. high; then, it is a known operation already experienced
• Each state is represented as a multivariate Gaussian in the past.
in the space of the variables describing the system • Operation _N_N93Z99_S11_MPF is not present in the
(current, drive load, speed, etc.), with mean μ i and training dataset and the Affinity Index is indeed very low
covariance matrix Σi. (around 0.6) in the test dataset; then, it is considered a
• The model has learnt automatically from the training set different operation, with respect to the past experience,
by using a Gaussian mixture model. The number of states and is not comparable.
K (i.e., the number of components of the mixture model) • Operation _N_OP006569_S02_MPF is not present in the
is inferred from the data by using the Bayesian informa- training, but the Affinity Index is high as the related vari-
tion criterion (BIC). able values are similar to those already seen in the train-
• For each new test point x, the probability of belonging to ing dataset; even if not experienced, it appears similar,
the state i that the CNC machine has visited during the i.e. statistically comparable, with past operations.
training process is evaluated as in Eq. 5:
Therefore, if operations are similar (such as the Operation
1(
[ )T )]
pi (x) = exp − x − 𝜇i Σ−1 (5) _N_N93Z99_S01_MPF and the Operation _N_OP006569_
(
i
x − 𝜇i
2
S02_MPF), then the ND model results can be trusted and, in
case of novelty thresholds overshoot, an incipient functional
• The Affinity Index of the point x is defined as the highest failure is foreseeable. Otherwise, if operations are not simi-
probability amongst the states composing the model as lar (e.g. the Operation _N_N93Z99_S11_MPF), then the
defined in Eq. 6 (where A.I. stands for Affinity Index): only information retrievable refers to the machine perform-
A.I.(x) = max pi (x) (6) ing operations never seen and not comparable with respect to
i=1..K
those observed during the period used to train the algorithm.
The application of the Affinity Index enables discerning The scaled-up evaluation of the Affinity Index involved
between already known operations, different operations, the adoption of the equal error rate (EER) assessment to
and similar operations, as noticeable in Fig. 6, based on the identify the best threshold (THaffinity). Specifically:
statistical comparison of collected variables. On the y-axis
of the upper plots, there is the Affinity Index value for each 1. The available dataset of operations was divided into
training set and test set.

13
4504 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

Fig. 6  Affinity Index as a function of the performed operations

2. The operations consistency checking was carried out as correspondent score (see Eq. 2) exceeds a predefined
a function of THaffinity; those operations with Affinity threshold, then an anomaly has arisen. In particular, GBR
Index higher than THaffinity were consistent, and were not is the ND model (as in subsection 3.3) and is used at the
otherwise. spindle and axis level so as to be able not only to notify an
3. The confusion matrix was calculated considering the anomaly but also to locate the anomaly within the machine
ground truth given by the labelling of the operations. tool. As a matter of fact, a CNC machine tool with 3 axes,
4. The false acceptance rate (FAR) and false rejection rate has 4 GBR models, one for the spindle, one for the x-axis,
(FRR) were evaluated and the EER showed that 0.75 as one for the y-axis and one for the z-axis. Table 3 sum-
THaffinity was the best option as visible in Fig. 7. marises the predictors for each type of model (spindle or
axis) as well as the response. Predictors and responses
5.3.2 Definition of the fault detection score have been identified starting from the variables/features
identified as a result of data manipulation.
Building on the approach by Pimentel et al. [26], the dif- Each model is trained on a 2-month period and tested on
ference between the predicted value due the ND model a 1-month period. The dataset is selected based on feedbacks
and the current measured value is computed and, if the from the company about the continuous work of the machine

Fig. 7  EER for THaffinity defini-

tion (example for the 4-axis and
5-axis CNC machines tools)

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4505

Table 3  Predictors and response for the spindle and axis models Using the box plots, and assuming the healthiness of the
Predictors Response
machine over the monitoring period during which the train-
ing and testing have been performed in this case, the reader
Spindle model Drive load Actual current may be sensitive to the improvement: leaving out the vari-
Actual rotary speed
ability induced by the different/not comparable operations,
Working state percentage
the box plot should be closer to the one from the training
Axis model Drive load of all axis except the Actual current
one under analysis phase. In Fig. 7, this happens for all axes.
Working state percentage Below are the comparison plots with the results for the
Absolute velocity of all axes 4-axis CNC machine tools (Fig. 9), the 5-axis CNC machine
except the one under analysis
tools (Fig. 10) and the 6-axis machine tools (Fig. 11). On the
y-axis of each figure, there is the average score, while on the
x-axis is the specific axis or spindle to which the score refers to.
along the considered monitored period. Also, the dates are It is worth seeing that the model performance (test after
compared with the maintenance interventions historian to AI versus test) for the 6-axis CNC machine tools, even
avoid considering weeks in which maintenance revisions acceptable, is overall worse than the 4- and 5-axis ones. As
took place. Figure 8 shows the results for a 4-axis CNC a reasonable hypothesis, this evidence could be motivated
machine tool; specifically, the value of the score is expressed by the fact that the formers perform more complex and het-
for the training dataset (named “train”), the testing dataset erogenous operations compared with the latters.
without any filter (named “test”), and the testing dataset after Overall, the results obtained by the application of the
filtering based on Affinity Index (named “test filtered”). GBR models confirmed first of all the goodness of the model

Fig. 8  Scores for the 4-axis CNC machine tool

Fig. 9  Model performance for the 4-axis CNC machine tools

13
4506 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

Fig. 10  Model performance for 5-axis CNC machine tools

Fig. 11  Model performance for the 6-axis CNC machine tools

in predicting the value of the current given the historical behaviour of the other HIs. However, to provide mean-
trend within the healthy state and, also, the capability of the ingful information to the maintenance decision-maker,
Affinity Index to improve model performance by avoiding the health state should be determined. Hence, two
the evaluation for those operations different from the ones thresholds should be defined, one that sets the limit
in the training dataset. between the healthy and the abnormal states, and the
second one that separates the abnormal state from the
5.4 Health assessment faulty state.
The decision is to adopt the same two thresholds for all
Running the GBR models allows evaluating the health indi- machines of the same type. This is established so as to have
cators HI as expressed in subsection 3.4: the health indicator a reference baseline that the decision-maker could use to
is the moving average of daily alert frequency. Figure 12 compare the HI of more machines. For the 4-axis CNC
shows the HI of the spindle for five 4-axis CNC machine machine tools, the two thresholds are set at 0.10 and 0.12,
tools, named 144 (blue line), 145 (orange line), 242 (green which respectively, means that 10% and 12% of the calcu-
line), 243 (red line) and 244 (purple line) according to the lated scores over the day bring an alert. Figure 13 shows
company’s internal identification number. the HI of the five monitored machines with highlighted the
It is possible to observe a constant trend for machine two thresholds that define the health states (healthy is green,
145, whose HI clearly deviates from the almost stable abnormal is yellow and faulty is red).

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4507

Fig. 12  Health indicators of the spindle for the 4-axis CNC machine tools

Fig. 13  Health indicators of the spindle for the 4-axis CNC machine tools with highlighted thresholds

As Fig. 13 shows, machine 145 is experiencing abnormal spindles and axes. Therefore, along a few months, the two
behaviour bringing it to a faulty state. In order to offer a solutions—the one proposed in this work and the condition
shred of consistency to the results and given that, according monitoring leading to vibration analysis—worked concur-
to the company’s experience, CNC machine tools are very rently, assuming the traditional and reliable vibration analy-
reliable, hence, no failures are expected at their components, sis as the state-of-the-art and ground truth. Indeed, vibration
an alternative way to validate the proposed method was analysis as a meaningful technique to understand the health
defined. Specifically, an independent maintenance provider state comes from the experts’ knowledge and it is supported
(IMP) expert in vibration analysis was asked to install its by previous scientific works as shown in section 2. It is
condition monitoring solution to measure anomalies of the important to point out that the installed condition monitoring

13
4508 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

solution worked under the same condition of no failure data per machine per month, scaling up to 9 TB of required
as the proposed operations-aware ND framework. storage per month for the 16 CNC machine tools. The
For the specific monitoring period reported in Fig. 12, the cost-benefit analysis was, therefore, core in defining the
time domain and frequency domain analyses from the accel- final architecture with on-edge processing with 2-min
erometers by the IMP performed on a daily basis, confirmed aggregation for cloud storage and subsequent elabora-
that the spindle of the CNC machine tool 145 was experi- tion.
encing an important degradation trend as highlighted by the
proposed solution. The observation of the physical state of
the spindle confirmed that it was close to sudden breakdown. 2. Data manipulation:
Hence, the double-check from the IMP and the company’s
maintenance technicians confirmed that the proposed solu-
tion correctly represented the health state of the monitored (a) The selection of the relevant features used to train and
spindle of the CNC machine tools. If already in place, the run the models has required various interactions with
proposed operations-aware ND solution could have been the project partners, especially with the personnel
able to foresee the spindle failure 10 days in advance with no working daily on the CNC machine tools; this enabled
additional sensor installation within the related architecture. the exploitation of domain knowledge in terms of man-
ufacturing technology as well as industrial automation,
thus reducing the “time-to-deploy” solution.
6 Discussion and lessons learnt

This section 6 aims at reflecting on the developed opera- 3. State detection:

tions-aware ND framework and the obtained results, in order
to discuss the practical relevance and the managerial impli-
cations of the work. (a) The choice not to go for artificial intelligence algo-
The results obtained have been reached always consider- rithms, e.g. neural network or deep learning, was based
ing the need to achieve a good balance between the model on the request of having interpretable model outputs
performance, the industrial viability and the cost-effective by maintenance technicians and managers. In manag-
applicability and usability of the solution. In particular, it ers’ minds, statistical modelling of the behaviour of the
is worth summing up some considerations regarding the machines could have promoted adequate trustworthi-
requirements and constraints requested for the development ness by the future users of the solution, while artificial
of the framework, before analysing the contributions this intelligence could be included in a roadmap of next
research work aims at bringing to the scientific literature. project developments.
Lessons learnt gained during the development and (b) The selected statistical model for fault detection, i.e.
deployment of the proposed framework could be indeed GBR, was identified based on (i) the previous experi-
summarised as follows, retracing the PHM model: ence the company in charge of algorithm development
had, and (ii) the results from a trade-off between the
1. Data acquisition: scalability options and easiness in model maintenance
in the long-term.

(a) The identification of items towards which modelling

effort is worth being directed, comes from the results of 4. Health assessment:
the FMECA as well as the availability of already-in-place
monitoring solutions, as a kind of digital readiness of the
machines. Consequently, the spindle and the axes were (a) The selection of the health indicator was based on the
identified as those of high criticality and already con- trade-off between the model’s capability to identify a
nected to CNC and PLC; vice versa, high-critical items deviation and the “suitable” notification frequency so
like swarf removal and refrigeration units were excluded as not to overload the maintenance process. Therefore,
due to the need for important retrofitting activities, in the best health indicator was built based on the daily
spite of their relevance in terms of impacts on downtime. frequency of alerts, not the point alerts.
They may be the target of the following projects.
(b) The sampling frequency for data collected from CNC The final result is a cloud-based solution that informs
and PLC is influenced by constraints on cloud storage. the maintenance decision-maker about the health states of
Setting a sampling time equal to 10 ms requires 560 GB the monitored subsystems of the CNC machine tools on the

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4509

manufacturing shop floor. The availability of an alternative to other technicians (i.e., experts in specific techniques such
CBM solution based on vibrations signals allows the valida- as vibration analysis). This cost-effective solution provides
tion of the results obtained with the operations-aware ND reliable information on the ongoing degradation risks during
models and health assessment. This proves it effectively the machine runtime while optimising the usage of experts
detects anomalies in the spindle and potentially in the CNC when required in further condition investigation.
machine tool axes, even though one failure only has been Overall, the project demonstrates that developing, deploy-
recorded during the monitored period. The information is ing and releasing a stable CBM solution for industrial
accessible to multiple users with different privileges and machines like CNC machine tools requires looking for a
can be integrated with the company’s CMMS for better good balance between model performance, industrial viabil-
decision-making in preventive and condition-based main- ity, cost-effective applicability and usability.
tenance actions.
Stemming from these results, it is worth digressing over
the adaptation of the proposed framework to other CNC 7 Conclusions and outlooks for future
machine tools. From the technological point of view, the research
framework is able to adapt to multi-axis CNC machine tools;
indeed, it was already applied to 4-, 5- and 6-axis CNC The research work aims at developing a framework for oper-
machine tools. Furthermore, if the CNC machine tool under ations-aware ND for CNC machine tools with available data
study does perform always the same set of part programmes from already-in-place CNC and PLC. As the scientific state-
and operations, then the Affinity Index could be skipped as of-the-art confirms, most of the developed ND solutions rely
there is no need to dynamically adapt to ever-changing and on vibration or temperature signals so as to perform FDD
unpredictable part programmes dictated by varying produc- of subsystems of machine tools. However, the availability of
tion requirements. If run-to-failure data are available, the such data does often require an important retrofitting activ-
evaluation of the score as novelty detection could be skipped ity on the machines under study, while CNC and PLC data
as it may be effective to move towards a complete supervised could be easier to gather as they are already installed. There-
learning. In all other cases, the suggested framework may fore, the research tackles this gap by proposing an ND solu-
find applicability as it targets the most critical step in PHM tion based on CNC and PLC data, by assuming no vibration
that is state detection in the particular case of limited avail- monitoring system available; indeed, the developed solution
ability of failure data and limited knowledge of upcoming is able to cope with varying operations the machine tools
operations resulting from changing production requirements. could realise to enable reliable and usable information for
Discussing the advancements with respect to the scien- maintenance decision-making.
tific state-of-the-art, this research work proposes an ND The proposed framework is based on the PHM process
framework based on the concurrent use of two statistical guidelines and extends the SD level by adding the so-
models, each for a specific goal: the first one looks for auto- called “operations consistency checking” step, towards the
matic operations consistency checking, and does represent Machine Tool 4.0 paradigm. This additional step is in charge
the main novelty that this work brings about as it enables of guaranteeing that the data-driven models for ND are run-
industrial viability in cases of different working conditions ning under similar conditions between the runtime data gath-
of the operating machines; the second one is focused on ered from the machines and the training dataset. Specifically,
anomaly detection, once similar working conditions are an index called Affinity Index, is evaluated, which compares
identified. The joint evaluation of both in terms of Affinity the current data with the historical one on which the models
Index and fault detection score allows robust information were trained, and measures the degree of similarity of the
to be transferred to the decision-makers. In particular, it is operations. If the Affinity Index is high enough, then the
worth remarking that the experimental observations (see model response could be used to judge the health state of
section 5.3.2) confirmed that the effect of variability due the CNC machine tools, triggering eventual maintenance
to the changing operations, was absorbed, finally leading actions. Otherwise, the maintenance decision-makers are
to a reliable fault detection score and, subsequently, health informed about the different operations the machine is per-
indicator to support decisions. forming with respect to its previous history and, through
Considering the decision-makers’ perspective, the solu- their expertise, a decision could be eventually judged.
tion offers usability with limited retrofitting. It utilizes vari- The proposed solution has been developed and deployed
ables measured by the existing controllers for the machining during a 3-year industrial project, which allowed testing
process, allowing the maintenance decision-maker to make of the operations-aware ND framework over several CNC
informed decisions by employing the ND and health assess- machine tools. After proper modelling of a GBR algorithm
ment on a regular basis, i.e. while the machining operations for each of the interested subsystems, namely the machin-
are running. Further condition investigation can be delegated ing unit that includes the spindle and the axes, the solution

13
4510 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

effectively identified an abnormal state of the spindle of one represent per se a future endeavour. Furthermore, other
of the CNC machine tools. If already operating, the model operational contexts resulting from different production
could have identified a real failure 10 days in advance. Also, requirements, could be helpful to extend the validation of
this solution has been tested and compared via traditional the automatic operations consistency checking and of its
monitoring solutions based on vibrations, which are rec- experimental proofs as herein reported (in subsection 5.3.1
ognised as the state-of-the-art in research and industry for and, concerning the effect on fault detection, in subsec-
this kind of machine, showing comparable health assess- tion 5.3.2). Overall, this may inform future experimental
ment performance. Overall, the obtained results improved works both in industrial settings and in laboratory settings.
the information available to maintenance technicians and Due to controlled environment, laboratory tests are helpful
managers as the cloud-based system is accessible by multi- to extend the experimentation to explore different influ-
ple users even remotely, highlighting the health state of the ent factors, while variations that could result in specific
monitored subsystems of the CNC machine tools. industrial settings, are also relevant in order to align to
Concerning the scientific state-of-the-art, the proposed challenges in real settings as well as needs and require-
operations-aware ND framework overcomes current limita- ments from the decision-makers and other stakeholders
tions related to the need for RTF data to implement an effec- involved in the industrial development.
tive condition-based maintenance solution. Also, the claim Furthermore, additional future research works could first test
this research brings about is the possibility to establish a well- the application of other algorithms for ND, searching for AI
performing ND solution even in case of no condition moni- solutions to improve prediction performance. Also, the Affinity
toring data, e.g. vibrations or temperatures, but relying only Index is currently based on a statistical comparison between his-
on CNC and PLC already-in-place. Finally, besides extensive torical operations the machine performed and the current ones;
data manipulation to enable data quality for the next PHM lev- also for this purpose, AI may be tested to verify any potential
els, this research claims that CNC/PLC data are a key element advantage in classifying similar and different operations. Sec-
to finally achieve reliable information, as they form the basis ondly, the proposed solution does not consider any feedback
for adapting to the changing working conditions as dictated to be included automatically to improve model performance;
by new part programmes introduced to satisfy continuously hence, an effort to establish a comeback of information from
evolving production requirements; this is an essential concept experts to be included is envisioned as a potential breakthrough
when multiple operations, determining varying working con- to optimise the performance of the overall solution. Finally,
ditions, impact on the health state evolution of the machine. integration with other data may be useful to extend diagnos-
From an industrial point of view, the proposed frame- tics and prognostics capabilities. This may come out in cloud-
work could be adapted, even made simple, to various situ- edge-based solutions capable to combine the computational
ations. As described in section 6, according to the vari- potentialities at different levels of the cyber-physical produc-
ability in part programmes and operations and/or in the tion system built as a result of digitisation. Indeed, the proposed
availability of failure data, some step of the framework operations-aware ND framework is conceived as a core part of
could be skipped (e.g. Affinity Index) or modified (e.g. an industrial information system stack able to inform multiple
complete supervised learning), respectively. Hence, the actors for integrated decision-making on the shop floor.
research could be of interest to any company willing to
Author contribution All authors contributed to the study’s concep-
implement a condition-based maintenance solution with- tion and design. The literature review was conducted by Adalberto
out important retrofitting activity. Indeed, this may repre- Polenghi, Laura Cattaneo and Marco Macchi. The data collection and
sent the first implementation of Industry 4.0 technologies preparation were carried out by Danilo Bruno, Sofia Mosci, Fabio Fer-
that could be then pushed forward by leveraging also on riani and Roberto Caneva. The deployment of the solution was sup-
ported by all authors. The first draft of the manuscript was written by
condition monitoring data and integration with informa- Adalberto Polenghi and all authors commented on previous versions
tion systems, such as CMMS, to label events thanks to the of the manuscript. All authors read and approved the final manuscript.
register of maintenance work orders.
Current and main limitation of this work relates to the Funding Open access funding provided by Politecnico di Milano
within the CRUI-CARE Agreement. This work was supported by the
assessment of the solution with one failure only available. Ansaldo Energia Lighthouse Plant (“FCS Accordo Quadro”, project
This has been counteracted by the concurrent evaluation named “Lighthouse Plant” led by Ansaldo Energia SpA through the
of health by means of vibration analysis, somehow cer- Programme Agreement entered into on 14 February 2018 between the
tifying the correct health state definition for the sixteen Italian Ministry of Economic Development and the Liguria Region).
machines explored in the study. However, the capability
of the framework and related algorithms to identify evolv-
Declarations
ing failures should be better assessed, beyond the only Conflict of interest The authors declare no competing interests.
one failure occurred during the project timeline, which

13
The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512 4511

Open Access This article is licensed under a Creative Commons Attri- 14. Cattaneo L, Polenghi A, Macchi M (2021) A framework to integrate
bution 4.0 International License, which permits use, sharing, adapta- novelty detection and remaining useful life prediction in Industry
tion, distribution and reproduction in any medium or format, as long 4.0-based manufacturing systems. Int J Comput Integr Manuf
as you give appropriate credit to the original author(s) and the source, 35:388–408. https://​doi.​org/​10.​1080/​09511​92X.​2021.​18850​62
provide a link to the Creative Commons licence, and indicate if changes 15. de Jonge B, Teunter R, Tinga T (2017) The influence of practical
were made. The images or other third party material in this article are factors on the benefits of condition-based maintenance over time-
included in the article's Creative Commons licence, unless indicated based maintenance. Reliab Eng Syst Saf 158:21–30. https://​doi.​
otherwise in a credit line to the material. If material is not included in org/​10.​1016/j.​ress.​2016.​10.​002
the article's Creative Commons licence and your intended use is not 16. Saari J, Odelius J (2018) Detecting operation regimes using
permitted by statutory regulation or exceeds the permitted use, you will unsupervised clustering with infected group labelling to
need to obtain permission directly from the copyright holder. To view a improve machine diagnostics and prognostics. Oper Res Per-
copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/. spect 5:232–244. https://​doi.​org/​10.​1016/j.​orp.​2018.​08.​002
17. Luo W, Hu T, Ye Y et al (2020) A hybrid predictive mainte-
nance approach for CNC machine tool driven by Digital Twin.
Robot Comput Integr Manuf 65:101974. https://​d oi.​o rg/​1 0.​
References 1016/j.​rcim.​2020.​101974
18. He X (2016) Recent development in reliability analysis of NC
1. Cohen Y, Faccio M, Pilati F, Yao X (2019) Design and manage- machine tools. Int J Adv Manuf Technol 85:115–131. https://​
ment of digital manufacturing and assembly systems in the Indus- doi.​org/​10.​1007/​s00170-​015-​7926-0
try 4.0 era. Int J Adv Manuf Technol 105:3565–3577. https://​doi.​ 19. Liu W, Kong C, Niu Q et al (2020) A method of NC machine
org/​10.​1007/​s00170-​019-​04595-0 tools intelligent monitoring system in smart factories. Robot
2. Ibarra D, Ganzarain J, Igartua JI (2018) Business model inno- Comput Integr Manuf 61:101842. https://​d oi.​o rg/​1 0.​1 016/j.​
vation through Industry 4.0: a review. Procedia Manuf 22:4–10. rcim.​2019.​101842
https://​doi.​org/​10.​1016/j.​promfg.​2018.​03.​002 20. Filev DP, Chinnam RB, Tseng F, Baruah P (2010) An industrial
3. Lee J, Davari H, Singh J, Pandhare V (2018) Industrial artificial strength novelty detection framework for autonomous equip-
intelligence for Industry 4.0-based manufacturing systems. Manuf ment monitoring and diagnostics. IEEE Trans Industr Inform
Lett 18:20–23. https://​doi.​org/​10.​1016/j.​mfglet.​2018.​09.​002 6:767–779. https://​doi.​org/​10.​1109/​TII.​2010.​20607​32
4. Negri E, Pandhare V, Cattaneo L et al (2021) Field-synchronized 21. Xu X (2017) Machine Tool 4.0 for the new era of manufactur-
Digital Twin framework for production scheduling with uncer- ing. Int J Adv Manuf Technol 92:1893–1900. https://​doi.​org/​10.​
tainty. J Intell Manuf 32:1207–1228. https://​doi.​org/​10.​1007/​ 1007/​s00170-​017-​0300-7
s10845-​020-​01685-9 22. Lee J, Ni J, Djurdjanovic D et al (2006) Intelligent prognostics
5. Ahmad R, Kamaruddin S (2012) An overview of time-based and tools and e-maintenance. Comput Ind 57:476–489. https://​doi.​
condition-based maintenance in industrial application. Comput org/​10.​1016/J.​COMPI​ND.​2006.​02.​014
Ind Eng 63:135–149. https://​doi.​org/​10.​1016/j.​cie.​2012.​02.​002 23. Jardine AKS, Lin D, Banjevic D (2006) A review on machin-
6. Zonta T, da Costa CA, da Rosa RR et al (2020) Predictive mainte- ery diagnostics and prognostics implementing condition-based
nance in the Industry 4.0: a systematic literature review. Comput maintenance. Mech Syst Signal Process 20:1483–1510. https://​
Ind Eng 150:106889. https://​doi.​org/​10.​1016/j.​cie.​2020.​106889 doi.​org/​10.​1016/j.​ymssp.​2005.​09.​012
7. Li K, Deng Q, Zhang L et al (2021) An effective MCTS-based 24. Kan MS, Tan ACC, Mathew J (2015) A review on prognostic
algorithm for minimizing makespan in dynamic flexible job shop techniques for non-stationary and non-linear rotating systems.
scheduling problem. Comput Ind Eng 155:107211. https://d​ oi.o​ rg/​ Mech Syst Signal Process 62–63:1–20. https://​d oi.​o rg/​1 0.​
10.​1016/j.​cie.​2021.​107211 1016/j.​ymssp.​2015.​02.​016
8. Polenghi A, Roda I, Macchi M, Pozzetti A (2021) Ontology-aug- 25. Xia T, Dong Y, Xiao L et al (2018) Recent advances in prog-
mented Prognostics and Health Management for shopfloor-synchro- nostics and health management for advanced manufacturing
nised joint maintenance and production management decisions. J Ind paradigms. Reliab Eng Syst Saf 178:255–268. https://​doi.​org/​
Inf Integr 100286. https://​doi.​org/​10.​1016/j.​jii.​2021.​100286 10.​1016/j.​ress.​2018.​06.​021
9. Guillén AJ, Crespo A, Macchi M, Gómez J (2016) On the role of 26. Pimentel MAF, Clifton DA, Clifton L, Tarassenko L (2014) A
Prognostics and Health Management in advanced maintenance review of novelty detection. Signal Process 99:215–249. https://​
systems. Prod Plan Control 27:991–1004. https://d​ oi.o​ rg/1​ 0.1​ 080/​ doi.​org/​10.​1016/j.​sigpro.​2013.​12.​026
09537​287.​2016.​11719​20 27. Galar D, Thaduri A, Catelani M, Ciani L (2015) Context aware-
10. ISO 13374-1 (2003) Condition monitoring and diagnostics of ness for maintenance decision making: a diagnosis and prog-
machines — data processing, communication and presentation— nosis approach. Measurement 67:137–150. https://​doi.​org/​10.​
Part 1: General guidelines. British Standards Institution 3:31 1016/J.​MEASU​REMENT.​2015.​01.​015
11. Lei Y, Li N, Guo L et al (2018) Machinery health prognostics: 28. Li R, Verhagen WJC, Curran R (2020) A systematic methodol-
a systematic review from data acquisition to RUL prediction. ogy for Prognostic and Health Management system architecture
Mech Syst Signal Process 104:799–834. https://d​ oi.o​ rg/1​ 0.1​ 016/j.​ definition. Reliab Eng Syst Saf 193:106598. https://​doi.​org/​10.​
ymssp.​2017.​11.​016 1016/j.​ress.​2019.​106598
12. Yucesan YA, Dourado A, Viana FAC (2021) A survey of mod- 29. Peng Y, Dong M, Zuo MJ (2010) Current status of machine
eling for prognosis and health management of industrial equip- prognostics in condition-based maintenance: a review. Int J
ment. Adv Eng Inform 50:101404. https://​doi.​org/​10.​1016/j.​aei.​ Adv Manuf Technol 50:297–313. https://​d oi.​o rg/​1 0.​1 007/​
2021.​101404 s00170-​009-​2482-0
13. Gittler T, Glasder M, Öztürk E et al (2021) International Confer- 30. Schlagenhauf T, Burghardt N (2021) Intelligent vision based wear
ence on Advanced and Competitive Manufacturing Technologies forecasting on surfaces of machine tool elements. SN Applied
milling tool wear prediction using unsupervised machine learning. Sciences 3:858. https://​doi.​org/​10.​1007/​s42452-​021-​04839-3
Int J Adv Manuf Technol 117:2213–2226. https://d​ oi.o​ rg/1​ 0.1​ 007/​ 31. Denkena B, Dittrich M-A, Noske H et al (2021) Data-based
s00170-​021-​07281-2 ensemble approach for semi-supervised anomaly detection in

13
4512 The International Journal of Advanced Manufacturing Technology (2023) 128:4491–4512

machine tool condition monitoring. CIRP J Manuf Sci Technol 43. Zhang Y, Mu L, Shen G et al (2019) Fault diagnosis strategy of
35:795–802. https://​doi.​org/​10.​1016/j.​cirpj.​2021.​09.​003 CNC machine tools based on cascading failure. J Intell Manuf
32. Wu Y, Liu L, Qian S (2021) A small sample bearing fault diagno- 30:2193–2202. https://​doi.​org/​10.​1007/​s10845-​017-​1382-7
sis method based on variational mode decomposition, autocorrela- 44. Cheng Y, Zhu H, Hu K et al (2019) Multisensory data-driven
tion function, and convolutional neural network. Int J Adv Manuf health degradation monitoring of machining tools by generalized
Technol. https://​doi.​org/​10.​1007/​s00170-​021-​08126-8 multiclass support vector machine. IEEE Access 7:47102–47113.
33. Goyal D, Dhami SS, Pabla BS (2021) Vibration response-based https://​doi.​org/​10.​1109/​ACCESS.​2019.​29088​52
intelligent non-contact fault diagnosis of bearings. J Nondestruct 45. Aghazadeh F, Tahan A, Thomas M (2018) Tool condition moni-
Eval Diagn Progn Eng Syst 4. https://​doi.​org/​10.​1115/1.​40493​71 toring using spectral subtraction and convolutional neural net-
34. Bernini L, Waltz D, Albertelli P, Monno M (2021) A novel prognostics works in milling process. Int J Adv Manuf Technol 98:3217–3227.
solution for machine tool sub-units: the hydraulic case. Proc Inst Mech https://​doi.​org/​10.​1007/​s00170-​018-​2420-0
Eng B J Eng Manuf. https://​doi.​org/​10.​1177/​09544​05421​10646​82 46. Herranz G, Antolínez A, Escartín J et al (2019) Machine tools
35. Du X, Gai J, Chen C (2020) Condition-based maintenance opti- anomaly detection through nearly real-time data analysis. J Manuf
mization for motorized spindles integrating proportional hazard Mater Process 3. https://​doi.​org/​10.​3390/​jmmp3​040097
model with SPC charts. Math Probl Eng 2020:7618376. https://​ 47. Huang X, Zhao F, Sun Z et al (2020) A novel condition monitor-
doi.​org/​10.​1155/​2020/​76183​76 ing signal analysis method of numerical control machine tools in
36. Fan J, Xue L, Liu Y, Li W (2021) Reliability analysis of spindle varying duty operation. IEEE Access 8:72577–72584. https://​doi.​
system of CNC grinder based on fault data. Int J Adv Manuf Tech- org/​10.​1109/​ACCESS.​2020.​29880​28
nol 117:3169–3183. https://​doi.​org/​10.​1007/​s00170-​021-​07552-y 48. Li W, Liu T (2019) Time varying and condition adaptive hidden
37. Cao H, Zhang X, Chen X (2017) The concept and progress of Markov model for tool wear state estimation and remaining use-
intelligent spindles: a review. Int J Mach Tools Manuf 112:21–52. ful life prediction in micro-milling. Mech Syst Signal Process
https://​doi.​org/​10.​1016/j.​ijmac​htools.​2016.​10.​005 131:689–702. https://​doi.​org/​10.​1016/j.​ymssp.​2019.​06.​021
38. Rastegari A, Archenti A, Mobin M (2017) Condition based main- 49. Friedman JH (2001) Greedy function approximation: a gradient
tenance of machine tools: vibration monitoring of spindle units. boosting machine. Ann Stat 29:1189–1232. https://​doi.​org/​10.​
In: 2017 Annual Reliability and Maintainability Symposium. 1214/​aos/​10132​03451
RAMS, pp 1–6 50. Friedman JH (2002) Stochastic gradient boosting. Comput Stat
39. Soori M, Arezoo B, Dastres R (2023) Machine learning and arti- Data Anal 38:367–378. https://​doi.o​ rg/​10.​1016/​S0167-​9473(01)​
ficial intelligence in CNC machine tools, a review. Sustain Manuf 00065-2
Service Econ 100009. https://d​ oi.o​ rg/1​ 0.1​ 016/j.s​ mse.2​ 023.1​ 00009 51. Grubbs FE (1969) Procedures for detecting outlying observations
40. Kuntoğlu M, Salur E, Gupta MK et al (2021) A state-of-the-art in samples. Technometrics 11:1–21
review on sensors and signal processing systems in mechanical 52. Fumagalli L, Cattaneo L, Roda I et al (2019) Data-driven CBM
machining processes. Int J Adv Manuf Technol 116:2711–2735. tool for risk-informed decision-making in an electric arc fur-
https://​doi.​org/​10.​1007/​s00170-​021-​07425-4 nace. Int J Adv Manuf Technol 1–14. https://​doi.​org/​10.​1007/​
41. Denkena B, Bergmann B, Stiehl TH (2021) Transfer of process s00170-​019-​04189-w
references between machine tools for online tool condition moni-
toring. Machines 9. https://​doi.​org/​10.​3390/​machi​nes91​10282 Publisher’s note Springer Nature remains neutral with regard to
42. Baur M, Albertelli P, Monno M (2020) A review of prognostics jurisdictional claims in published maps and institutional affiliations.
and health management of machine tools. Int J Adv Manuf Tech-
nol 107:2843–2863. https://​doi.​org/​10.​1007/​s00170-​020-​05202-3

13