Journal of Industrial Information Integration 45 (2025) 100833

Contents lists available at ScienceDirect

Journal of Industrial Information Integration

journal homepage: www.elsevier.com/locate/jii

Full Length Article

Bridging the gap between Industry 4.0 and manufacturing SMEs: A

framework for an end-to-end Total Manufacturing Quality 4.0’s
implementation and adoption
,∗
Badreddine Tanane a,c , Mohand-Lounes Bentaha a , Baudouin Dafflon b , Néjib Moalla a
a Univ Lyon, Université Lumière Lyon 2, INSA Lyon, Université Claude Bernard Lyon 1, DISP, EA4570, 69676 Bron, France
b Univ Lyon, Université Claude Bernard Lyon 1, INSA Lyon, Université Lumière Lyon 2, DISP, EA4570, 69621 Villeurbanne, France
c TARDY SAS, France

ARTICLE INFO ABSTRACT

Keywords: Manufacturing is one of the industrial sectors taking benefit from the 4th industrial revolution and bringing
Quality 4.0 existing production capacities closer to the "factory of the future". Quality, as a main concern in manufacturing,
SME is also to benefit from this change of paradigm by introducing new key enabling technologies such as Internet
Sensor network
of Things (IoT) and Artificial Intelligence (AI) into quality management, earning it the label of Quality 4.0
IoT
(Q4.0). The implementation of these paradigms is still gathering research efforts as it is arduous to design and
Artificial intelligence
Industry 4.0
realize effective end-to-end Decision Support Systems (DSSs) for Q4.0, with several dimensions to consider
Defect management when integrating digitalization with quality. This is an even more challenging task when it comes to SMEs’
Smart manufacturing efforts to implement these concepts, given the particularities of these entities. This paper presents an approach
to design a Total Manufacturing Quality 4.0 (TMQ 4.0) DSS by combining Sensor Network (SN) data and
historical data in an end-to-end framework. Furthermore, the paper presents the validation of the approach
through a case study application in a metal-cutting high-precision manufacturing SME. It shows promising
Q4.0 estimations with regular Machine Learning (ML) algorithms (kNN, Random Forest, Logistic Regression,
XGboost, feed-forward Deep Neural Network) when the steps of tending to data quality, data augmentation,
and end-to-end design and implementation are applied. By providing building blocks for an end-to-end Q4.0
DSS design and implementation in an integrated quality control application, this approach aims at supporting
end-users in the in-process quality control of their manufacturing operations.

1. Introduction • Addressing manufacturing issues at a later stage, while the wanted

state is to accompany the production process right from the
Customer confidence is the backbone of the manufacturing industry, beginning to detect defects in-process (or in-line) [5].
especially in the current digitally-enhanced and global-scale market • Approaching quality either through a process-related lens or a
where keeping a competitive edge is a must for a company’s sur- product-related one [6].
vival [1]. With quality issues (repair, scrap, . . . ) estimated to cost up
to 30% of the total income of a manufacturing enterprise [2], defect The first limitation makes those approaches unable to predict and/or
management has always been a research topic for academia and indus- prescribe course-correction actions before the defects affect the prod-
try alike to achieve continuous improvement cycles and reduce process uct/process, as per the Zero-defect-manufacturing (ZDM) paradigm [7,
variation causes [3]. This relevance has been seen through the evolu- 8]. The second limitation makes their perspective no longer relevant
tion of different quality approaches: from traditional Statistical Process in recent manufacturing where issues are found to have (multiple)
Control (SPC) (Quality 1.0), to Quality insurance (Quality 2.0) to Six entwined causes in both product and process aspects.
Sigma’s DMAIC and other Total Quality Management approaches [4]. Industry 4.0 (I4.0), an ongoing transformation of the existing indus-
These quality approaches suffer from many limitations, such as: trial systems, brought up many vital innovations through digitalization

∗ Correspondence to: 59 Avenue Galline, 69100 Villeurbanne, France.

E-mail address: [email protected] (B. Tanane).

https://doi.org/10.1016/j.jii.2025.100833
Received 27 September 2023; Received in revised form 8 December 2024; Accepted 12 March 2025
Available online 21 March 2025
2452-414X/© 2025 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

1. Quality before manufacturing: Process configuration (PC)

Acronyms
• Operations quality characterization.
AI Artificial Intelligence
• Machines maintenance-related aspects.
DMLP Deep Multilayer perception
• Parameters (more specifically cutting parameters).
DSS Decision Support System
I4.0 Industry 4.0 2. Quality while manufacturing: In-Process Quality (IPQ).
IoT Internet of Things 3. Quality after manufacturing: Quality Control (QC).
IPQ In-Process Quality
kNN K-Nearest Neighbors Through this research’s scope, the aim is to comprehensively deal
with quality in the manufacturing stage, targetting both process and
LGR Logistic Regression
product-oriented defects, focusing on real-time intervention for in-
LVHV Low-Volume High-Variety
process monitoring. An AI module design is proposed to aim for real-
ML Machine Learning world applicability and address AI limitations in Q4.0 through col-
NC Nonconformity laborative labeling with the machine operator. Consequently, high
Q4.0 Quality 4.0 contextualization for sensor data and seamless integration and inter-
QC Quality Control action between the resulting tool and manufacturing processes are
RDF Random Forest achieved through the data pipeline. Finally, the end-to-end design
SN Sensor Network (from defects to the SN covering the defects to the AI-based decision
TMQ 4.0 Total Manufacturing Quality 4.0 support for manufacturing) intends to deal with Q4.0 barriers by:
XGB XGBoost
• Encouraging Q4.0 strategies adoption on the strategic level of
companies (decision makers) through immediate added value.
• Increase adherence of shop-floor level workers: through the im-
and diverse technological tools and resources [9]. These recent ad- mediate and increasing decision support in manufacturing opera-
vances are expected to make defect detection and handling faster, tions, as well as the dynamic collaborative labeling.
easier, and affordable, through AI-driven algorithms and enhanced
These two aforementioned aspects are known, through the different
sensing. Consequently, Quality 4.0 emerged as a main prospect of I4.0
Q4.0/I4.0 maturity assessment models in literature [13–15,18–20],
applied to quality management, aiming to propose a new paradigm to
to be some of the main limiters of Q4.0/I4.0 implementation and
overcome the abovementioned limitations [10].
adoption, especially in SMEs.
Q4.0, as a subarea of I4.0, is still in its nascent stage [11], with many
In this paper, the focus will be on the theoretical foundation as
additional hurdles facing the realization of effective Q4.0-oriented
well as the end-to-end implementation of the different conceptual
decision support systems:
bricks of Q4.0. With an emphasis on cost-effectiveness and real-world
• Companies skepticism about the real added value of Q4.0 ap- implementability, a holistic approach is proposed for the design and
proaches when faced with smart manufacturing high entry bar- implementation of Quality 4.0, adaptable to all manufacturing-based
riers [12] on a: companies. Results are validated in a real-world case-study applica-
tion with promising results in said aspects representing state-of-the-art
– Material level: extensive digitalization expenditure, lack of challenges, as will be further discussed in the related work section.
IT facilities, . . . The paper is structured as follows: Section 2 will discuss literature
– Organizational/workforce level: lack of top management work related to the topic at hand, Section 3 will present the proposed
understanding and commitment, poor digital culture, need framework of this approach, Section 4 will be dedicated to its imple-
for advanced ICT skills ... mentation steps, Section 4.4 will describe a real-world case study based
on the implementation while Section 5 will conclude this paper and
• Lack of approaches for low-volume/high-variety production: also
introduce future research perspectives.
called one-lot size manufacturing, mass-customization induces
further variability and a lack of control over production processes,
2. Related works
needing agile methodologies to address quality.

Adding other SMEs’ related constraints (financial budget, opera- 2.1. Lack of SMEs’ adapted industry 4.0 implementations
tional resources, . . . ) [13,14], it explains why most I4.0 models appear
to be disconnected from the actual state of digitalization and smart Many papers approach the abovementioned issue of I4.0/Q4.0 im-
manufacturing level of SMEs [15]. plementation and adoption in manufacturing companies generally and
AI applications in current Quality 4.0 approaches add to these SMEs in particular. Enrique et al. [21] states that some of the main
barriers by making overly optimistic assumptions about the data avail- issues of proposed I4.0 adoption approaches are the ‘‘rigid technology
able on shopfloor levels. Implementations are based on quality data, roadmaps’’ not considering the nuances of different production targets
with available labels (supervised-learning based works being the most between productivity and flexibility, as well as not having a global
numerous), clear contextualization, and stationary distribution. These quality management approach for both process and product related.
assumptions are mostly inapplicable, as real-world implementations The same rigidity in approaches was identified in [22] and in [23] from
show a chronic shortage of contextualized and labeled data [16] with a design research perspective.
shifting data distributions [17], making conventional static Machine In a systematic review of papers related to smart manufacturing in
Learning hard to apply. SMEs conducted by Mittal et al. [15], it has been found that most I4.0
To deal with these barriers, this paper introduces the notion of TMQ maturity models, as well as implementation guidelines and frameworks,
4.0 to address quality in the manufacturing stage of the production start from a base level (‘‘level 1’’) which is completely disconnected
process. Consequently, quality related to other stages such as design or from the real digitization and smart manufacturing maturity level of
service is out of scope. As illustrated in Fig. 1, quality in manufacturing many SMEs. Two main issues were mainly targeted, the first being
lifecycle stages can be generally divided into three phases, and the the lack of IT integration such as the absence of modern connected
following aspects will be targetted by this paper: machines with high-level data acquisition capacities, and the second

2
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 1. Total manufacturing quality within product lifecycle modeling phases.

was adherence issues to newly implemented I4.0 initiatives due to a Amaral and Peças [13] finds similar constraints, namely opera-
more day-to-day oriented mindset and lack of visible immediate effects. tional resources and financial budget, making I4.0 notions and their
The paper goes on to propose a ‘‘level 0’’ reflecting the real entry level constituents (among which is Q4.0) intangible regarding implemen-
for SMEs but remains on a conceptual level to help reflect SMEs’ needs tations and economic benefits for SMEs. According to the authors,
into a new maturity level, concluding that SMEs need to develop their this led this category of companies to fail to grasp or trust these
own unique SM vision and roadmap, which is the same conclusion notions without real, tangible, and accessible examples, while bigger
found by the study of Masood and Sonntag [19] firms can afford to take more risks. This is even more pronounced
Neumann et al. [24] conducts a study about I4.0 methodologies and for non-repetitive production enterprises, i.e. Low-Volume High-Variety
frameworks and their actual implementation stages where it targets (LVHV) ones [27].
the lack of end-to-end solutions, stating that it is not fully clear to Schönfuß et al. [14] points to the same barriers, adding that the low-
many in both industry and research what fully realized I4.0 applications volume high-complexity products mix of SMEs heavily rely on human
might look like of how they might operate. It also indicates that the workers, leading to a variety in output products’ quality. As a solution,
lack of concern for human factors in I4.0 implementation strategies the authors suggest that approaches to use low-cost components and
results in an important research and application gap when speaking technologies as well as retrofitting existing machinery with digital
about adoption and adherence to SM initiatives, consequently leading capabilities is the way to go for bridging this gap, while also indicating
to ‘‘digital transformation and its implications on operations processes
that there is a lack of systematic methods for developing such solutions.
remaining a big black box’’. This lack of understanding on how to
A study of retrofitting an existing old manufacturing robot arm into a
implement these solutions, a majority of which are yet to be grasped
CPPS based on requirements analysis is given by Lins and Oliveira [28],
by the manufacturing industry, is pointed out in the I4.0 technologies
but is yet to englobe all aspects of the approach and be generalized for
overview and implementation levels by Frank et al. [20].
machining manufacturing. Similarly, Ottesjö et al. [29] find that em-
Raj et al. [18], Dassisti et al. [25], Ghobakhloo and Ching [26] also
phasizing technical solutions leads to a lack of end-to-end approaches
focus on the lack of digital strategy and resource scarcity being the
to connect things within the company, making SM implementation in
most prominent barriers for I4.0 implementations in manufacturing,
SMEs challenging due to missing the bigger picture. McFarlane et al.
adding that the constant working environment’s changes induced by
[30], Hawkridge et al. [31], besides pointing out the same barriers, in-
the disruptive uplift of SM lead to poor adherence in a yet to skill-up
dicate the need for low-cost solutions, called ‘‘the shoestring approach’’,
workforce. Among the rest of the barriers that can be found:
using off-the-shelf non-industrial monitoring components. The same
• Strategic barriers: Lack of clarity regarding economic benefits, need for low-cost smooth incremental implementation was presented
lack of standards/regulations/certifications, ineffective change in [32].
management. Abubakr et al. [33] also focuses on the cost-effective implementa-
• Technological barriers: Challenges in value-chain integration, low tion of I4.0 with an approach to develop a tool condition classification
maturity level of involved technologies, old infrastructure, chal- model (TCCM) using ‘‘only’’ one sensor, which proved to be highly ef-
lenges ensuring data quality. fective. Contrary to existing studies focused on increasing the accuracy
• Financial barriers: High initial investments. using multi-sensor TCCMs, which increases the cost of this technology
• Human resources barriers: Disruption to existing jobs, lack of and makes it inaccessible, especially for small and medium enterprises,
digital skills and internal culture, resistance to change. with different characteristics extended in [34].

3
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 2. Maturity levels of decision support systems.

2.2. Quality 4.0 in industry 4.0 blocks with a lack of integration. Similar integration issues of comput-
ing information within manufacturing processes in I4.0 were pointed
Quality 4.0 as defined by Fonseca et al. [35] is the combination of out by Sanchez et al. [42].
quality management with digitalization and one of the drivers of the Huang et al. [43] states that social factors have a significant impact
successful I4.0 adoption. Psarommatis et al. [36] defined quality man- on Q4.0 implementation and adoption, which is similar to general
agement as a system that can be further divided into product-oriented, Quality Management barriers, but with an even more pronounced effect
which deals with quality related to the defect in the product, while due to the technological barriers of digitalization. The authors argued
process-oriented will deal with the manufacturing equipment defect that the key to quality performance in the digital era does not lie solely
that affected the end-product. Process-oriented quality has been exten- in technical factors such as data analytics and IoT, but also in social
sively researched through advances in predictive maintenance [37,38] factors which will influence the adoption and implementation of Q4.0
such as the quality of the interaction between the operator and DSS as
with the combination of IoT capacities and AI-based analytics [39,40].
well as its adaptability.
Conversely, the framing of product-oriented quality in the data-rich
In the recent Quality Q4.0 literature review by Saihi et al. [44],
context of Industry 4.0 is still going at a slow pace. In the extensive
it has been found that current research suggests insight for potential
review of data analytics in Quality 4.0 by Bousdekis et al. [41], it
contributions but lacks further detail on the exact applications and
has been found that data management methods and tools still rely
solutions through use cases and case studies. It also highlights the
on subjective judgment and experienced intuition, and have yet to be
general need for substantial investments and hence points to the lack
adapted to cope with the data-intensive manufacturing environment of economic concern when implementing these approaches.
of I4.0, making them inefficient. The study relates process-oriented Another literature review by Liu et al. [45] further explains the
quality with process configuration, while product-oriented quality is transition from traditional quality management to quality 4.0, and
the addition of IPQ and QC as per the second level of Fig. 1. The study states that Q4.0 is the future of Total Quality Management (TQM) by
concludes that in regards to the three levels of data/AI analytics in Fig. using I4.0 technologies in real-time monitoring and AI-based insights
2: to extend the TQM scope to go from reactive/proactive to predictive
quality management activities. The study explains that one of the
• Process configuration is mostly addressed with descriptive an- main gaps of the current research on Q4.0, being still in its infancy
alytics, based on manual processes heavily relying on expert stage, is that there are not enough applications of Q4.0 being studied
knowledge due to high complexity/variety/criticality processes. in order to reconcile research with practical applications, and that
• Literature is rich regarding in-process quality with descriptive this reconciliation should not start from theoretical results but from
and predictive analytics. Prescriptive analytics is scarce and ap- real-world quality management problems.
proaches utilize domain-specific optimization.
• Quality control is governed by manual and classical quality pro- 2.3. Lack of data quality concerns in decision support systems design
cedures (Lean, 6 sigmas, . . . ) with papers using the same ex-
perimental dataset. Even quantitative methodologies such as sta- Data quality is paramount for decision support systems, especially
tistical process control (SPC) and theory of constraints are lim- in a highly complex environment such as LVHV manufacturing. These
ited when incorporating/analyzing various and numerous data systems need to have high-quality data by design, meaning data that
sources dynamically in a complex environment. will have enough information for the AI-based algorithms to extract
knowledge from.
This means that beyond the descriptive phase, literature becomes scarce The literature review by Bousdekis et al. [41] states that pure data-
and when the approaches exist profusely (like in in-process control) driven systems are limited in terms of decision-making ability, while
they have yet to be operationalized, their efficiency in real-world research works trying to incorporate expert knowledge at design time
conditions (i.e. outside of testbeds) is largely untested, and, most and combine it with sensors’ data analytics often lead to domain-
importantly, remain as single components and domain-specific building specific solutions with low flexibility and scalability. Also, the existing

4
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 3. Classification of AI/ML vibrational-based IoT systems in literature.

hybrid algorithms require the embodiment of expert knowledge and 2.4. Hurdles of applying AI/ML in quality 4.0
process specifications in a structured way, a process that cannot keep up
with the large volume of unstructured data (sensors) they need to aug- There are plenty of papers in the literature regarding AI/ML applica-
ment. The authors also pointed to a research gap in combining multiple tions for I4.0. In the metal-working industry, there is a heavy focus on
data sources (sensors, ERP, . . . ), leading to an existing research gap in vibration-based IoT systems since vibration is the most convenient and
fusing data sources and then combining it with domain knowledge to relevant physical quantity for TMQ 4.0-related issues, as it was detailed
provide optimized human-machine collaboration in quality procedures. in [50].
Yuan et al. [46] emphasizes the importance of data quality by An overview of these applications can be seen in Fig. 3, with
stating that it can cause wrong decisions, resulting in huge losses due to references in Table 1 (see Appendix). The applications’ vibrational data
defective products, equipment damage, and human injury. It also states either being:
one of the main limits of research in I4.0 by saying that the datasets
used to train intelligence algorithms are significantly different from • Directly fed to AI/ML models, generally architectures that can
the actual sensor data collected in the production factory. In practical automatically learn and extract information representation. For
applications, algorithms often fail to reach the expected effectiveness example deep and complex neural networks or state-of-the-art
level and so, before involving them in any decision process, one should attention-based neural networks.
evaluate the related risk. • Or, before ending in AI/ML models going through feature engi-
In a proposed approach to deal with cutting processes’ issues in neering and selection methods.
milling by Xi et al. [47], it is concluded that a variety of sensors and
AI/ML usage in Q4.0 can also be classified by purpose (further refer-
signals can be used to monitor tool wear with satisfactory results in
ences can be found in Table 2:
the laboratory. However, to what extent these methods can be im-
plemented at the shop-floor level without negatively affecting existing • For fault diagnosis, a multiclass classification problem in AI/ML
manufacturing processes is yet to be explored. terms.
Hinojosa-Palafox et al. [48] also advocates the need for a ref- • For anomaly detection, a binary classification problem.
erence architecture for developing software-based solutions for I4.0 • Remaining useful life (RUL) estimation, a regression problem.
applications, focusing on data quality management and integration, to
not end with I4.0 research far from industrial application. A problem With models being either supervised/semi-supervised if labels are
thoroughly discussed in [49] where the authors concede that even present, self-supervised if there is a way for the algorithm to interact
with the explosion in IoT research volume, much of it still fails to with the environment, or unsupervised (clustering) when none of the
reach any industrial application because of many barriers. Mainly: above apply.
affordable, multipurpose sensing and computing hardware embraced As stated by the comprehensive literature review of Zhang et al.
by researchers are still sometimes seen by the industry as unserious or [51], fault diagnosis supervised learning is the best choice in terms
unfamiliar to the engineers and developers who could commercialize of discriminative power. However, it is restricted mostly to laboratory
the research findings. As such it is advised that sensors and ICT tools data due to the insufficiency of labeled data in real-world applica-
need to be approached in research from the perspective of fitment to tions, where destructive experiments are costly and/or not allowed.
purpose, with familiar and already adopted popular components by Furthermore, the learned classifiers may only be sensitive to the faults
industry. included in the training set, knowing that lab data and real-world data

5
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 4. Sensor network design.

are inherently different, this can cause poor generalization capability • Quality 4.0 design: Need for a holistic approach to tackle mul-
to unseen faults, leading to low testing accuracy in real world. tiple quality aspects and orientations.
If multi-class classification is not possible due to, for example, a • AI applicability: with barriers in two main aspects:
lack of data on complex failures, anomaly detection is the second-best
choice. If supervised learning is not possible, unsupervised methods can – Data quality: for AI real-life applications, available data is
be used to generate scores representing the abnormality of a given sam- scarce and suffers from many limitations (structuring, la-
ple. But in practice, a threshold is needed to judge the anomalousness, beling, contextualization ...), making supervised and semi-
making the process complex and very much application-dependent. supervised ML/AI algorithms hardly usable.
RUL aims to learn a function that maps the condition of an item to – Algorithms benefits: Existing approaches heavily rely on IoT
its RUL estimates. It being a regression problem makes it challenging to sensor data, which lacks informativeness in itself to make use
provide accurate labels for training, even more so than classification. of the full cognitive capacity of AI.
Indeed, in prognostic applications, it is hard if not impossible to accu-
rately determine the RUL of an object at any given time. While most
research uses data from run-to-failure tests, deriving the RUL labels
from them as well as the fitness function is very difficult and extremely These limitations make current approaches hardly applicable in a
application-dependent [51]. real-world setting, which explains the lack of research solutions making
To confirm these findings, regarding the ML/AI aspect of the DSS their way into the industry.
for I4.0 applications, Baumung and Baumung [52] pointed out that ML Consequently, this paper will aim to define a holistic approach for
provides a variety of tools for classification/pattern recognition and
Q4.0 called TMQ 4.0. The design proposition and implementation take
so on, but the performance of these algorithms relies very heavily on
into account the entry barriers aforementioned to maximize adoption
training material, i.e. data quality. It is indicated that a major problem
and adherence. Data quality of sensor data is also addressed through
in the practical application of ML is that training material (labeled data)
the interaction with legacy and shop-floor data. Finally, the proposed
is expensive and time-consuming to obtain. The authors use transfer
learning (training and testing on different domains) in order to deal data pipeline allows for operator-tool interaction, providing labels and
with this aspect. an incremental learning strategy for the tool to ensure usefulness.
Another recent study [53] tackles the issue of AI-based quality
supervision challenges and the lack of labeled data through active 3. Proposed framework
learning in order to reach a better prediction accuracy with a smaller
number of training data instances. 3.1. Overview
Most research tries to tackle these problems with increasingly com-
plex models, leading to high training/testing costs as well as heavy
An overview of the design and exploitation approach, which can be
models lacking timeliness for on-the-fly prediction. Also, Labeling is-
divided into two distinct parts (the SN and the rest of the DSS), can be
sues are dealt with using bench-scale experiments attempting to simu-
late reality, but often with simplifications and strong assumptions that summarized as follows:
may not hold in reality [51,54].
1. First bloc ‘‘SN design’’ (Fig. 4):
2.5. Paper position and contributions
• Input: all defect-related data, whether process-related or
product-related, used to design a relevant SN to enhance
From the above, the literature regarding the implementation of I4.0
quality applications, especially in an SME LVHV environment, indicates data coverage for both aspects. The final retained SN will
the following scientific gaps which are still actively researched: be a fusion of both SNs concerning the barriers discussed in
Section 2.1.
• Adoption and Adherence: Lack of approaches adapted to SMEs • Output: Arbitrations are made (constraints, cost-benefit ...) as
and LVHV manufacturing. Cost-effective, low-cost starting, and explained in Section 3.2, resulting in a suitable acquisition
clear economically benefitting solutions to be integrated on an system sending IoT data to a relevant operational DB.
end-to-end basis in the company’s value-chain are needed to
refurbish old infrastructure without disrupting existing jobs. 2. Second bloc ‘‘DSS architecture’’ (Fig. 5):

6
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 5. TMQ decision support system architecture building blocks.

• Data sources: Shop-floor data (maintenance plannings, 3.3.1. Total manufacturing quality
NCs related data, workpiece details, . . . ) and machine-related Quality approaches in the industry went through different stages,
data stored for training and dashboarding. which can be summed up with the following:
• Databases (DBs) for different purposes.
• Operator interface: for operational awareness, labeling, and • Quality 1.0 and the introduction of inspection through Quality
decision-making support display. Control (QC), a product-oriented approach aiming to exclude
• AI-based module: for manufacturing monitoring and quality non-compliant products from production with techniques such
predictions. as statistical process control (SPC) and statistical quality control.
Namely, trying to keep production deviations from reaching the
client.
• Quality 2.0 with quality insurance, a process-oriented proactive
3.2. Sensor network design
methodology to standardize production. Or, in other words, to
A bottom-up starting approach is proposed to ensure the SN is keep deviations from affecting the product.
capable of sensing quality issues, as can be seen in Fig. 4. As such, from • Quality 3.0, or Total Quality Management, is a company-oriented
process-related defects and maintenance history, a suitable Proc-Q SN approach to implement strategies throughout the company to
is devised (with more detail in [34]) through a defect analysis (common heighten client satisfaction.
causes and special causes [55]).
For the product-oriented counterpart, the product-related defects While current research has yet to produce a standardized definition
and other ERP product legacy data are analyzed to characterize the for Quality 4.0, it was mainly defined at first as the addition of
NCs. This leads to the choice of the type of physical signals needed to digitalization and I4.0 capacities to Total Quality management. Recent
be monitored depending on the type of NC [56]. papers are questioning this definition due to the limitations exposed in
The chosen SN are then checked for resilience/risk assessment 1. Due to those challenges, research is asking for more synergy between
for each sensor (redundancy, signal sensitivity, data reliability, . . . ), technologies and quality management to achieve in-process monitoring
and each type of physical signal is analyzed for applicability con- instead of just digitalizing post-manufacturing quality approaches [57].
straints in real-life industrial monitoring, with extensive sensor spec- On holistic quality approaches, Liu et al. [45] focuses on design
ifications (contact type, integration possibility, cost, measuring range, and service quality instead of manufacturing quality since they can
. . . ). Further details of the SN design can be found in [50].
provide more added value to the production. This is hardly applicable
Finally, the resulting aggregated SN needs to go through arbi-
for manufacturing SMEs where the norm is LVHV production, almost
tration with, for example, trade-offs between redundancy and cost-
solely relying on mass-customization and the ability to constantly deal
effectiveness (entry barriers gap in Section 2.5), types of sensing, and
with new and complex to manufacture products. Variable and hard-to-
other constraints (sensitivity, resolution, environment interference ...).
standardize processes put a considerable strain on the manufacturing
3.3. DSS design principles stage of the Product Lifecycle Modeling (see Fig. 1) rather than Design
and Service. As the main obstacle toward mass-customization and a
The rest of the DSS design can further be decomposed into the main Q4.0 objective, ensuring Manufacturing-based quality is not only
following building blocks as per Fig. 5: necessary but can also affect the other two phases:

7
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 6. DSS pipeline.

• Service: predicting defects can improve company-client and 3.3.3. Data flow
supplier-company aspects (cost control, delay control, production To deal with the issues of data quality, the following design is
planning, . . . ) proposed in Fig. 5:
• Design: enhanced control over processes and products in manu-
3.3.3.1. Operational DB. The operational DB deals with real-time pre-
facturing can help predict bad designs and suggest early design
dictions and data analysis. As such, it receives:
changes (CAD/CAM).
• IoT Data originating from the designed SN.
Many papers dealing with Q4.0 from a manufacturing view focus on
• Shop-floor legacy data (process and product-oriented, ERP for
Product-oriented quality (IPQ and QC) or process-oriented quality (ma-
product specifications, . . . ) as well as machine meta-data (cutting
chine monitoring and predictive maintenance). The proposed approach
parameters, operation types, corresponding timestamps, . . . ) for
aims for a more holistic view by proposing a TQM model taking into
IoT data-contextualization.
account the ‘‘operations’’ and ‘‘parameters’’ aspects of the workpiece
• Operator’s annotations for training the AI-module.
manufacturing, addressing all general manufacturing steps and their
main sub-processes as can be seen in Fig. 1. This holistic viewpoint And sends:
makes the approach extensible to all manufacturing domains without
loss of generalization. • Operation information to a dashboard serving as an interface with
the operator for operational awareness enhancement
3.3.2. Data quality • Historical data to a strategic database acting as a data lake/data
Data quality is paramount when dealing with IoT-based AI algo- warehouse for Online analytical processing (OLAP) purposes
rithms. Unlike structured and annotated experimental datasets used (KPIs, strategic mid-long term management, . . . ) and data reten-
by many approaches in literature to train I4.0 supervised machine tion policies.
learning algorithms, real-world applications’ IoT-data is unstructured
(raw) and mainly unlabeled. It lacks the necessary contextualization 3.3.3.2. Operator interface. The Operator interface is also a core com-
and information to be useful ML models ready to be integrated into the ponent of the suggested DSS design. Its main purpose is displaying the
shop floor. AI module quality predictions and its labeling inquiries. Those inquiries
can be:
Generally, either the data scientist would have to annotate it man-
ually by sticking through the entire experimentation to the machine/ • Intermediate quality controls to validate a prediction.
operator, or the operator would annotate the data post-manufacturing • A labeling query to annotate significant subsets of the data, which
as an expert. This process is inefficient and far from being deployable can be chosen through Active Learning.
in real production systems. Approaches trying to circumvent the lack
of labeling through unsupervised or semi-supervised learning are inher- Those inquiries are then answered by the operator, acquired through
ently weaker in discriminating capacity, leading to significant setbacks the interface, and sent to the prediction module. This can deal with
in detection and prediction accuracy. data quality limitations by labeling and contextualizing the data. With
The same goes for data contextualization. With most actors yet to queries in varying operation conditions, it favors applicability for
achieve complete digitalization, raw sensor data with no context is changing data distribution, another limitation of AI in Q4.0 and mass-
difficult to use in machine learning since there is no indication of which customization.
parameter/operation/workpiece,. . . , is being monitored through each It also displays strategic operation-related information in real-time,
part of the signal. such as process parameters and workpiece information, gathered in

8
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 7. Operator/AI-DSS dynamical interaction for manufacturing operations.

one place from the many data sources. In doing so, the operator’s Cognitive maturity level characterization The requirements for an
adherence is maximized through immediate added value even in the AI module in quality are expressed below. In the following, numbers 1-
initial learning phases where predictions are still being built. 2-3 would correspond to each cognitive maturity level of Fig. 2. Those
By reducing the burden on operational responsibilities, the operator levels can be translated as requirements for an AI module in quality:
will be more available to interact with the interface to send feedback
• Process-oriented (Process configuration):
and annotations. This ensures a smooth and efficient human-AI interac-
tion throughout the training of the system, from detection to prediction – Operations:
to prescription. A proposition to address the human-factor-related entry
barriers for real-world implementation of Q4.0 especially in SMEs. 1. Differentiate between different operations to character-
ize well-made or faulty operations (operation discrimi-
3.3.3.3. Total manufacturing quality 4.0 AI module. nation).
Design The third component of the DSS is the AI module. Data- 2. Predict operation deviations at early stages (operation
label availability remains the biggest hurdle for real-world implemen- adjustment).
tation of supervised ML algorithms. While the design of the interface 3. Suggest optimal usage scenarios throughout the manu-
alleviates this problem, it cannot provide labels for all the data, with facturing (operation copiloting).
an acquisition system in real-world conditions monitoring the machine
– Machines: detect machine deviations, both from special and
for extended durations (2 × 8 or 3 × 8 shifts every day), with different
common causes (bearing faults, spindle issues, kinematic
workpieces and machining programs. The operator cannot be con-
drifts ...):
stantly interacting with the interface, since his main job is to operate
the CNC machine and make sure the manufacturing goes well, among 1. Optimized reactive maintenance.
other operations in hidden time (switching raws, machine preparation 2. Predictive maintenance.
...). 3. Prescriptive maintenance: scenarios for machine-life en-
The AI module is able to: hancement (load adjustment, operations adaptation/
restriction, . . . )
• Detect bad operation quality characteristics and help optimize
– Parameters: differentiate between different sets of cutting
cutting parameters and workpiece disposition
parameters (tool feed, cutting depth, rotation speed, clamp-
• Detect machine deviations and tool defects
ing, . . . ) by:
• Detect and prevent workpiece quality issues either in-process or,
if the prediction confidence level is not enough, recommending 1. Characterizing if parameters are appropriate for an op-
post-manufacturing QC checks if the workpiece is suspected to be eration or not.
damaged. 2. Predicting unadapted parameters at an early stage.
3. Early-stage suggesting of optimized parameters.
The design needs to address the aforementioned limitations of AI
models, without needing to use overly complex algorithms to compen-
sate for poor data quality, an application barrier as per literature. This • Product-oriented:
is done through the data pipeline offering data contextualization, with
process and product-related data provided as input during the training – In-process quality:
to enhance prediction ability.
1. Detect and characterize NCs.
For labeling queries, an Active Learning (AL) approach is cho-
2. Predict manufacturing deviations at an early stage and
sen on top of the classic ML algorithms for fault diagnosis/anomaly
suggest intermediate controls.
detection/RUL. While AL in I4.0/Q4.0 is still a newly researched sub-
3. Suggest early-stage prescriptive adjustments (change of
ject [53], it is a growth-type method in which the algorithm can select
operation, parameters, . . . ).
a subset of examples it considers to be the most informative from
the pool of unlabeled data and query the user interactively (here the – Quality control: detection and prediction that a workpiece
operator interface). This method allows for continuous learning, with might need a final post-manufacturing quality check. Can
gradually higher levels of accuracy and cognition with smaller training be a first step before achieving sufficiently efficient and
labels while keeping training/testing costs low, achieving the so-called confident IPQ.
human-in-the-loop paradigm.
To summarize, the DSS can be roughly schematized as a 3 macro-
components system: data, AI, and operator interaction. By putting more Other maturity indicators can also be considered for the DSS such
emphasis on the first and second components, the aforementioned lim- as speed of intervention, or explainability of detection/prediction/
itations can be addressed to achieve an applicable end-to-end decision prescription in line of works where black boxes are structurally impos-
support system for manufacturing quality. sible.

9
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 8. DSS deployment.

3.4. Discussion control over the highly variable production processes while keeping
in mind the inherent conditions and resources (old legacy machine,
Considering the different barriers and research gaps raised in Sec- technological investment limitations, . . . ).
tion 2.5, the presented approach:

• Proposes a solution to help industry actors and SMEs achieve 4.2. Total manufacturing quality
Q4.0 by minimizing entry barriers: Cost-effectiveness, low-cost
starting, and an immediate business-added value. The continuous To develop a classification of the primary causes of manufacturing
learning offers an evolving system able to adapt to the actual dig-
deviations, a comprehensive analysis was carried out, both on a:
italization levels in the said industry without disrupting existing
jobs.
• Quantitative level: with QC data, NCs reports, business process
• Covers all quality aspects in manufacturing.
bottlenecks, cost prices, customers feedback, . . .
• Focuses on data quality and data integration issues for AI in Q4.0
• Qualitative level: field surveys of different stakeholders including
(labeling, varying operating conditions, contextualization).
machine operators.
• Targets shopfloor operators’ adherence and cooperation by design
through the modeled collaboration.
As a result, and following wide-spread failure mode and effect analysis
The next section will present a partial implementation in a real- (FMEA) principles [58], the following classification was achieved:
world scenario of the different principles and components of the sys-
tem. • Tool issues: accounting for roughly 49% of the critical deviations.
• Material deformation: accounting for 27% of the critical devia-
4. Implementation tions.

4.1. Overview
4.2.1. Tool-related quality issues
Using the right tool for the right operation can be very challenging
The proposed approach in this paper was implemented in Tardy
SAS, an SME operating in the subtractive metallurgy industry, which for make-to-order production where every workpiece is unique in terms
provides a comprehensive range of services, from conceptualization to of specifications, machining parameters, materials, and so on. Tool
the production of various metal-based workpieces and special-purpose issues can encompass:
machines. As a rank 1 subcontractor for critical sectors like armament
and aerospace, operating on a make-to-order basis, Tardy faces an in- • Process-oriented quality issue: lack of control over processes (tool
creasing demand for technically complex, low-volume, and high-variety choice, cutting parameters, machine deviations) leads to spindle
requests. issues. The lack of visibility on tool-life management causes cut-
As such, implementing a Q4.0 approach is of the utmost importance ting tool problems (wear, breakage, heating, deviation ...), and
to maintaining client satisfaction and the company’s reputation in an ultimately machine downtime.
increasingly competitive market. In order to do so, Tardy needs an effi- • Product-quality issue: out-of-tolerance dimensions, bad surface
cient uplift toward this concept to enhance the reactivity, security, and finishing, . . . .

10
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

• Cost-effectiveness for a near industrial-grade data acquisition

quality.
• The open-source software with access to all its component cus-
tomization through Python or C scripts available on the construc-
tor’s GitHub repository. As such, the user can seamlessly integrate
it into the DSS deployment app as well as provide immediate-use
added value through available libraries and functions.

Other aspects are natively lacking compared with costlier industry-

target solutions (IP protection, shockproof, . . . ), but most of these can
be circumvented in a non-redhibitory way (for example custom cases).
Regarding the needed requirements, this DAQ is a good alternative to
Fig. 9. DMG Mori 210p.
industrial-specific DAQs which are generally very costly, mostly run
on proprietary software, and need a lot of effort for customization and
integration.

4.4. DSS application

4.4.1. Design
While the detailed implementation of AI module will not be dis-
cussed in the scope of this paper, the experimental pipeline and its
general rationale are given in Fig. 6, depicting the implementation of
the framework’s building blocks (see Fig. 5), where:

• The h.file (machining code for Heidenhain-based CNC) is going

Fig. 10. DMU 210p machine-related NCs.
to be parsed from the CNC machine to get machine and op-
erations metadata (no need for new machines with enhanced
connectivity).
4.2.2. Material deformation quality issues • The operational database used is TimescaleDB, which is a popular
Manufacturing transformation processes involve material removal extension over PostgreSQL for time-series handling. Running on
from raw parts. Depending on many specifications, raw parts need to be
open source software and plenty of examples and community
adequately held in place using various angles, positions, and intensities.
support, it also has a very meager learning curve for it is an
The lack of control of make-to-order production in this aspect can lead
SQL database and coincidently would integrate easily with legacy
to:
databases based on tables and SQL.
• Process-oriented NCs: excessive clamping can cause tool break- • Legacy data is connected to the system through the ERP
age, while insufficient one can generate spindle deformation. • The User dashboard is for monitoring and displaying data and
• Product-oriented NCs: excessive clamping can lead to raw defor- insights depending on the maturity level of the system, as well as
mation, and an insufficient one can induce bad surface finishing providing contextualization and labeling depending on the needs
or out-of-tolerance workpiece ... of the Active Learning ML algorithm.

4.3. Sensor network design A partial zoom-in for the operator and AI-module interaction is
proposed in the dynamical diagram shown in Fig. 7. A first-level inter-
A minimally-covering SN is proposed for the PC/IPQ/QC-related
action provides the operator with descriptive analytics of the manufac-
issues of a DMG Mori DMU 210p 9, an industrial-grade 5-axis universal
turing order (summary, statistical indicators, unsupervised ML) before
machining CNC.
The Product-related and process-related NCs made in Section 4.2 starting manufacturing. This part extends the core principles of SMEs-
brought out a suitable SN distribution for each NC. Also, a specific oriented dashboard design outlined in [60] to promote an effective
machine-related NCs for the DMU gave the coverage diagram in Fig. shop-floor activity. The second level iteratively produces a back-and-
10: forth of AI-based recommendations from the DSS and labels/evalua-
The chosen minimal SN is a vibration-based one, which seemed to tions for those decisions and other operations by the operator, guid-
be a good trade-off following the barriers in Section 2.5. The chosen ing him through the manufacturing operations while enhancing the
sensors were industrial-grade piezoelectric IEPE accelerometers with learning capacity of the AI algorithms.
characteristics that can be found below, and a discussion in [50] as
to why they are relevant for vibration monitoring in this case:
4.4.2. Deployment
• Measurement range: 0–10 000 Hz For deployment purposes in a real-world application, a system
• Sensitivity: average of 100 mV/g resilient enough to resist a lot of data per day is needed. Indeed, before
• Temperature working range: −54 to +124 ◦ C
data retention and aggregation if needed, and at 2000 Hz, a suitable
• IP protection index: IP68
scan rate, a day generates 2000 ∗ 3600 ∗ 24 = 172.8 million rows.
The data acquisition system is an MCC 172 DAQ HAT for sound and As such, a suitable deployment also bridges the gap toward a real
vibration measurements on top of a Raspberry Pi 4 (8go RAM). With application with MLOps principles, since the model is made to be up
the possibility of stacking up to 2 sensors per card and 8 cards per Rpi and running while AI models are constantly being enhanced/changed
and a maximum scan rate of 51.2 kHz [59]. Its main advantages are: as can be seen in the deployment Fig. 8.

11
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

• Mills: [Roughing 63 mm 3-teeth carbide mill, finishing 16 mm

2-teeth carbide mill]

Each file corresponded to a sequence of multiple combinations of

these parameters and, for each file, 4 sequences were used (3 for
training and one for testing). Different standard [61] time-domain,
frequency-domain, and time-frequency domain features were calcu-
lated based on windows of different sizes. 4 classic off-the-shelf ML
algorithms and one deep learning (DL) algorithm were applied: K-
Nearest Neighbors (kNN) - Random Forest (RDF) - Logistic Regression
(LGR) - XGBoost (XGB) - Deep Multilayer perception (DMLP). The
algorithms were used to predict the TMQ 4.0 indicators, with grid
search for best parameters estimation. The result of a 10-run average
result for the accuracy, precision, as well as the F1-score can be seen
in Fig. 13. further details can be found in Appendix Table 3.
The preliminary results are quite satisfactory for real-world data.
Fig. 11. Faulty operation detection. The best models for precision are the XGB and RDF with respectively
(72%, 66%) for window sizes of (1500,1700). The same goes for accu-
racy with a score of respectively (64%, 70%), joined by kNN (63%), for
windows sizes of respectively (1000,1500,1700). The F1-score was also
satisfactory with 71% for XGB, 68% for DMLP, 66% for the RDF and
LGR, and 65% for the kNN for windows of (1500,100,200,2000,1700).
Optimum results overall in terms of scores and consistency are given
by the XGB and the RDF, with optimum windows being in the narrow
ones (100–200) and wide ones (1500,1700).

4.4.4. Integration
Initially, the DSS is integrated into the manufacturing processes as
Fig. 12. AI module for TMQ.
a visualization tool. Through inquiries, the AI module improves its
prediction ability regarding the targeted NCs and provides inferences
that are then confirmed or denied by the planned operator checks. With
4.4.3. Exploitation and validation continuous learning, it provides more and more accurate predictions to
In this paragraph, the exploitation results of the framework’s build- detect and predict defects.
ing blocks’ implementation are given, with a special emphasis on the On the shopfloor level, the DSS has been well received by the
Data/Machine/AI interaction validation. The general principles of the operators since it enhances their operational activities by overcoming
operator-related interactions with the DSS are given in Section 4.4.4. the limits of legacy machines and difficult mass-customization-related
An initial analysis of the time series data indicates that the end- quality issues. The usage is light so as to not overburden their daily
to-end proposed solution offers the ability to sense and monitor the operations.
machining operations in terms of precision (up to 10−6 g) and variabil- Finally, the DSS uses a micro-services architecture with APIs to
ity of data. The SN meets the different requirements for the targeted connect the chosen components. Mostly, those components (ERP, qual-
phenomenon by finely representing the different machining sequences ity databases ...) are common in manufacturing industries. The addi-
and parameters (process). tional ones, i.e the Sensor Network and the dashboard interface, can
Fig. 11 provides an example of a detected faulty operation. Both easily be developed through the design guidelines. This ensures the
blue and red figures indicate exactly the same workpiece (material, generalization and cost-effectiveness capacity of the approach.
dimensions, operation sequences, . . . ) but with different cutting pa-
rameters and clamping. The system detected that the first part of the 5. Conclusion and future works
milling (along the X-axis) yielded way more vibrational energy, leading
to bad surface finishing roughness after final controls. The rest of Industry 4.0 has introduced several innovative aspects, one of which
the monitoring shows no difference, meaning that it can characterize is Quality 4.0. This concept strives to offer a comprehensive approach
operation quality depending on the workpiece/sequence/parameters to enhancing production quality through ICT technologies. However,
combinations. implementing Quality 4.0 poses numerous challenges, particularly for
To further confirm the ‘‘data-quality by design’’ of the DSS, several mass-customization. These barriers, ranging from organizational ones
ML algorithms were trained for TMQ 4.0. Cutting parameters and to AI applicability ones, through the Q4.0 approaches paradigm itself,
tool wear were considered for ProcQ and surface finishing roughness severely limit the implementation of Q4.0 for manufacturing.
for ProdQ. The process is described by Fig. 12. IoT data from real- To address these limits, this work presents:
world multi-mining operations of metal workpieces were applied. The
acquisition protocol, as described in previous work [50], consists of • An end-to-end design and implementation of a decision-support
multiple mining sequences with several parameters: system for Quality 4.0: from defect characterization to Sensor
Network design, followed by an AI module and data orchestration
• Feed rate: [140%, 120%, 100%, 75%, 50%] pipelines.
• Rotation: [120%, 100%, 75%, 50%] • A Total Manufacturing Quality 4.0 concept, aiming to target all
• Lubricant: with/without manufacturing stages to provide a holistic quality approach.
• Material: mild steel/aluminum • Data interaction and integration to offer a pipeline to address
• Cutting depth: [0.5 mm, 1 mm, 1.5 mm, 4.5 mm] data-quality issues such as contextualization, structuring, and
• Axis direction: [X,Y] informativeness.

12
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Fig. 13. Precision, accuracy and F1-score of optimal instances of the trained models.

Table 1
AI application methods in vibrational-based I4.0 IoT systems.
Method References
Raw data directly into AI models Wang et al. [62], Fang et al. [63], Chen et al. [64] and Xu et al.
[65]
Image-based methods Yan et al. [66], Garcia et al. [67], Nam and Kang [68] and Kou
et al. [69]
Feature engineering and selection
Modal decomposition methods Yesilli et al. [70], Mezni et al. [71], Zhao et al. [72], Han et al.
[73], Minhas et al. [74] and Dovedi and Upadhyay [75]
Entropy-based methods Zheng et al. [76], Wang et al. [77], Han et al. [78] and Zheng
et al. [79]
Time-frequency methods Heistracher et al. [80], Puchalski and Komorska [81], Liu et al.
[82] Jin and Chen [83], Zhu et al. [84], Chen et al. [85] and
Shakya et al. [86]

Table 2 • The preliminary results of the AI module are satisfactory, but they
AI use-cases categories in vibrational-based I4.0 IoT systems.
can be further developed to include other NCs.
Method References
• The implementation of an adapted active learning approach in
Fault diagnosis Chen et al. [87], Jia et al. [88], Yang the AI module is also considered. While all data contextualization
et al. [89], Ahmad et al. [90], Zhu et al.
and quality insurance is automated, labeling is still systematic and
[91], Roy et al. [92], Vos et al. [93]
and Li et al. [94] hence ‘‘unintelligent’’. The operator’s burden in the loop needs
Anomaly detection Zonzini et al. [95], Patra et al. [96], to be further lessened for better adherence. An adapted Active
Chen et al. [97], Miki and Demachi Learning is still being researched, as the domain is in its nascent
[98], Wang et al. [99], Saucedo-Dorantes
as an AI semi-supervised technique for I4.0 environments.
et al. [100], Radhakrishnan et al. [101],
Tang et al. [102] and Ma et al. [103]
• Use-case scenarios, with validation feedback loops and resilience
Remaining useful life (RUL) Wu et al. [104], Liu et al. [105], Ben Ali metrics, need to be developed to monitor the AI module.
et al. [106] and Soualhi et al. [107] • The exploitation phase of the DSS needs to be further developed
(from detection to prediction to prescription), to further display
how the AI predictions can be made into actionable integrated
• An AI module enhanced by an interface dashboard, striving to decision-making.
achieve continuous-learning through operator interaction. This
methodology allows for targeted labeling of data and adaptation CRediT authorship contribution statement
to data distribution shifts which are some of the main limitations
of AI applicability in Q4.0. Badreddine Tanane: Writing – review & editing, Writing – origi-
nal draft, Methodology, Investigation, Formal analysis, Data curation,
To address the adoption and adherence barriers faced by Quality 4.0 Conceptualization. Mohand-Lounes Bentaha: Writing – review & edit-
initiatives in the industry, this proposition emphasizes cost-effectiv ing, Validation, Project administration, Investigation, Formal analysis,
eness, generalization, and seamless integration into manufacturing op- Conceptualization. Baudouin Dafflon: Writing – review & editing,
erations, Validation, Supervision. Néjib Moalla: Writing – review & editing, Val-
For future work, the aim is to further expand the exploitation of the idation, Supervision, Resources, Project administration, Methodology,
created DSS by tending to the following limits of the proposed research: Funding acquisition, Formal analysis, Conceptualization.

13
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

Table 3
AI models best parameters (best windows are for the F1-score).
Model Best parameters References
Random Forest {max_feature = sqrt, n_estimators = 150, Yesilli et al. [70], Yesilli et al. [108],
window_size = 200} Han et al. [78], Xu et al. [65], Hu et al.
[109], Yesilli et al. [110] and Sadhu
et al. [111]
kNN {n_neighbors = 11, p = 2, weights = Khan and Kim [112], Mezni et al. [71],
uniform, window_size = 700} Yesilli et al. [70], Maliuk et al. [113],
Hu et al. [109], Zhang et al. [114],
Puchalski and Komorska [81], Zhang
et al. [115], Chen et al. [85], Islam
et al. [116], Ma et al. [117] and
Georgoulas et al. [118]
Feedforward DNN {activation = tanh, hidden_layer_sizes = Zhang et al. [119], Wang et al. [120],
(DMLP) (100,50), solver = adam, window_size = Prosvirin et al. [121], Han et al. [73],
100} Han et al. [122], Puchalski and
Komorska [81], Zhang et al. [123] and
Zhang et al. [115]
Logistic Regression {C = 0.1, penalty = L1, solver = Yesilli et al. [70], Yesilli et al. [108],
liblinear, window_size = 2000} Yesilli et al. [110] and Aydin et al.
[124]
XGboost {learning_rate = 0.01, max_depth = 7, Yesilli et al. [70], Yesilli et al. [108]
n_estimators = 150, window_size = and Yesilli et al. [110]
1500}

Declaration of competing interest [8] F. Psarommatis, G. May, A practical guide for implementing zero defect
manufacturing in new or existing manufacturing systems, Procedia Comput.
Sci. 217 (2023) 82–90.
The authors declare that they have no known competing finan-
[9] H. Lasi, P. Fettke, H.-G. Kemper, T. Feld, M. Hoffmann, Industry 4.0, Bus. Inf.
cial interests or personal relationships that could have appeared to Syst. Eng. 6 (4) (2014) 239–242.
influence the work reported in this paper. [10] C.A. Escobar, D. Macias-Arregoyta, R. Morales-Menendez, The decay of six
sigma and the rise of quality 4.0 in manufacturing innovation, Qual. Eng. 36
Acknowledgments (2) (2024) 316–335, Jv1z6 Times Cited:0 Cited References Count:105.
[11] R.R. Kumar, L. Ganesh, C. Rajendran, Quality 4.0 – a review of and framework
for quality management in the digital era, Int. J. Qual. Reliab. Manag. 39 (6)
This paper presents some results that are developed in collabora- (2021) 1385–1411.
tion between the TARDY SAS company and the University Lumiere [12] A. Majumdar, H. Garg, R. Jain, Managing the barriers of industry 4.0 adoption
Lyon 2, DISP Lab. This research is established under a CIFRE contract and implementation in textile and clothing industry: Interpretive structural
model and triple helix framework, Comput. Ind. 125 (2021) 103372.
(2020/1663). The content of this paper reflects an R&D initiative
[13] A. Amaral, P. Peças, SMEs and industry 4.0: Two case studies of digitalization
promoted by TARDY SAS. Responsibility for the information and views for a smoother integration, Comput. Ind. 125 (2021) 103333.
expressed in this paper lies entirely with the authors. [14] B. Schönfuß, D. McFarlane, G. Hawkridge, L. Salter, N. Athanassopoulou, L.
de Silva, A catalogue of digital solution areas for prioritising the needs of
Appendix manufacturing SMEs, Comput. Ind. 133 (2021) 103532.
[15] S. Mittal, M.A. Khan, D. Romero, T. Wuest, A critical review of smart
manufacturing & industry 4.0 maturity models: Implications for small and
See Tables 1–3. medium-sized enterprises (SMEs), J. Manuf. Syst. 49 (2018) 194–214.
[16] T.T. Chen, V. Sampath, M.C. May, S. Shan, O.J. Jorg, J.J.A. Martin, F. Stamer,
Data availability G. Fantoni, G. Tosello, M. Calaon, Machine learning in manufacturing towards
industry 4.0: From ’for now’ to ’four-know’, Appl. Sci.- Basel 13 (3) (2023).
[17] D. Sarkar, M. Ramezankhani, A. Narayan, A.S. Milani, Non data hungry
Data will be made available on request. smart composite manufacturing using active transfer learning with sigma point
sampling (SPSATL), Comput. Ind. 151 (2023).
[18] A. Raj, G. Dwivedi, A. Sharma, A.B.L. de Sousa Jabbour, S. Rajak, Barriers
References to the adoption of industry 4.0 technologies in the manufacturing sector: An
inter-country comparative perspective, Int. J. Prod. Econ. 224 (2020) 107546.
[1] N. Gramegna, F. Greggio, F. Bonollo, Smart factory competitiveness based [19] T. Masood, P. Sonntag, Industry 4.0: Adoption challenges and benefits for SMEs,
on real time monitoring and quality predictive model applied to multi-stages Comput. Ind. 121 (2020) 103261.
production lines, in: Advances in Production Management Systems. Towards [20] A.G. Frank, L.S. Dalenogare, N.F. Ayala, Industry 4.0 technologies: Implemen-
Smart and Digital Manufacturing, Springer International Publishing, 2020, pp. tation patterns in manufacturing companies, Int. J. Prod. Econ. 210 (2019)
185–196. 15–26.
[2] W. Sundblad, What’s at stake in the race to industry 4.0? 2018. [21] D.V. Enrique, G.A. Marodin, F.B.C. Santos, A.G. Frank, Implementing indus-
[3] J. de Mast, J. Lokkerbol, An analysis of the six sigma DMAIC method from the try 4.0 for flexibility, quality, and productivity improvement: Technology
perspective of problem solving, Int. J. Prod. Econ. 139 (2) (2012) 604–614. arrangements for different purposes, Int. J. Prod. Res. (2022) 1–26.
[4] A. Chiarini, M. Kumar, What is quality 4.0? An exploratory sequential mixed [22] H. Erbay, N. Yıldırım, Combined technology selection model for digital transfor-
methods study of Italian manufacturing companies, Int. J. Prod. Res. 60 (16) mation in manufacturing: A case study from the automotive supplier industry,
(2022) 4890–4910. Int. J. Innov. Technol. Manag. 19 (07) (2022).
[5] R. Dastoorian, L. Wells, Gauge capability studies for high-density data: SPC [23] S. Perera, X. Jin, P. Das, K. Gunasekara, M. Samaratunga, A strategic framework
phase 0, in: 48th Sme North American Manufacturing Research Conference, for digital maturity of design and construction through a systematic review and
vol. 48, Namrc, 2020, pp. 105–113. application, J. Ind. Inf. Integr. 31 (2023) 100413.
[6] F. Psarommatis, G. May, V. Azamfirei, F. Konstantinidis, Optimizing efficiency [24] W.P. Neumann, S. Winkelhaus, E.H. Grosse, C.H. Glock, Industry 4.0 and the
and zero-defect manufacturing with in-process inspection: Challenges, benefits, human factor – a systems framework and analysis methodology for successful
and aerospace application, Procedia Comput. Sci. 232 (2024) 2857–2866. development, Int. J. Prod. Econ. 233 (2021) 107992.
[7] F. Psarommatis, S. Prouvost, G. May, D. Kiritsis, Product quality improvement [25] M. Dassisti, A. Giovannini, P. Merla, M. Chimienti, H. Panetto, An approach
policies in industry 4.0: Characteristics, enabling factors, barriers, and evolution to support industry 4.0 adoption in SMEs using a core-metamodel, Annu. Rev.
toward zero defect manufacturing, Front. Comput. Sci. 2 (2020). Control. 47 (2019) 266–274.

14
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

[26] M. Ghobakhloo, N.T. Ching, Adoption of digital technologies of smart [51] L.W. Zhang, J. Lin, B. Liu, Z.C. Zhang, X.H. Yan, M.H. Wei, A review on
manufacturing in SMEs, J. Ind. Inf. Integr. 16 (2019) 100107. deep learning applications in prognostics and health management, Ieee Access
[27] M. Galetto, E. Verna, G. Genta, F. Franceschini, Uncertainty evaluation in the 7 (2019) 162415–162438.
prediction of defects and costs for quality inspection planning in low-volume [52] W. Baumung, V. Baumung, Application of machine learning and vision for
productions, Int. J. Adv. Manuf. Technol. 108 (11–12) (2020) 3793–3805. real-time condition monitoring and acceleration of product development cycles,
[28] T. Lins, R.A.R. Oliveira, Cyber-physical production systems retrofitting in Procedia Manuf. 52 (2020) 61–66.
context of industry 4.0, Comput. Ind. Eng. 139 (2020) 106193. [53] J.M. Rožanec, L. Bizjak, E. Trajkova, P. Zajec, J. Keizer, B. Fortuna, D. Mladenić,
[29] B. Ottesjö, S. Nyström, D. Nåfors, J. Berglund, B. Johansson, P. Gullander, A Active learning and novel model calibration measurements for automated visual
tool for holistic assessment of digitalization capabilities in manufacturing SMEs, inspection in manufacturing, J. Intell. Manuf. (2023).
Procedia CIRP 93 (2020) 676–681. [54] S. Zhang, F. Ye, B.N. Wang, T.G. Habetler, Semi-supervised bearing fault
[30] D. McFarlane, S. Ratchev, L. de Silva, G. Hawkridge, B. Schönfu, G.T. Angulo, diagnosis and classification using variational autoencoder-based deep generative
Digitalisation for SME manufacturers: A framework and a low-cost approach, models, Ieee Sens. J. 21 (5) (2021) 6476–6486.
IFAC- Pap. 55 (2) (2022) 414–419. [55] M. Ben-Daya, U. Kumar, D. Murthy, Introduction to Maintenance Engineering,
[31] G. Hawkridge, A. Mukherjee, D. McFarlane, Y. Tlegenov, A.K. Parlikad, N.J. Wiley, Hoboken, NJ, États-Unis, 2016.
Reyner, A. Thorne, Monitoring on a shoestring: Low cost solutions for digital [56] ISO-standard, ISO 9001:2015(en) quality management systems — requirements,
manufacturing, Annu. Rev. Control. 51 (2021) 374–391. 2015.
[32] F. Doyle, J. Cosgrove, Steps towards digitization of manufacturing in an SME [57] F. Psarommatis, Zero defect manufacturing (ZDM): The future method for
environment, Procedia Manuf. 38 (2019) 540–547. quality assurance in the era of industry 4.0/5.0, in: Encyclopedia of Sustainable
[33] M. Abubakr, M.A. Hassan, G.M. Krolczyk, N. Khanna, H. Hegab, Sensors Technologies, Elsevier, 2024, pp. 615–632.
selection for tool failure detection during machining processes: A simple [58] P. Klaput, R. Hercik, Z. Machacek, D. Noskievicova, V. Dostal, D. Vykydal,
accurate classification model, CIRP J. Manuf. Sci. Technol. 32 (2021) 108–119. Mutual combination of selected principles and technologies of industry 4.0 and
[34] B. Tanane, M.L. Bentaha, B. Dafflon, V. Ferreiro, N. Moalla, Toward an adap- quality management methods-case study, Qual. Eng. (2023).
tive approach to implement predictive maintenance in metallurgical industry [59] M.C.C. (MCC), MCC 172 IEPE measurement DAQ HAT for raspberry pi, 2022.
SMEs through IoT and AI, in: Product Lifecycle Management. Green and [60] S. Vilarinho, I. Lopes, S. Sousa, Developing dashboards for SMEs to improve
Blue Technologies To Support Smart and Sustainable Organizations, Springer performance of productive equipment and processes, J. Ind. Inf. Integr. 12
International Publishing, Cham, 2022, pp. 537–547. (2018) 13–22.
[61] S.K. Wan, X.H. Li, Y.J. Yin, J. Hong, Milling chatter detection by multi-feature
[35] L. Fonseca, A. Amaral, J. Oliveira, Quality 4.0: The EFQM 2020 model and
fusion and adaboost- SVM, Mech. Syst. Signal Process. 156 (2021).
industry 4.0 relationships and implications, Sustainability 13 (6) (2021) 3107.
[62] B. Wang, Y. Lei, N. Li, T. Yan, Deep separable convolutional network for
[36] F. Psarommatis, G. May, P.-A. Dreyfus, D. Kiritsis, Zero defect manufacturing:
remaining useful life prediction of machinery, Mech. Syst. Signal Process. 134
State-of-the-art review, shortcomings and future directions in research, Int. J.
(2019) 106330.
Prod. Res. 58 (1) (2019) 1–17.
[63] H. Fang, J. Deng, B. Zhao, Y. Shi, J. Zhou, S. Shao, LEFE-Net: A lightweight
[37] T. Zonta, C.A. da Costa, R. da Rosa Righi, M.J. de Lima, E.S. da Trindade, G.P.
efficient feature extraction network with strong robustness for bearing fault
Li, Predictive maintenance in the industry 4.0: A systematic literature review,
diagnosis, IEEE Trans. Instrum. Meas. 70 (2021) 1–11.
Comput. Ind. Eng. 150 (2020) 106889.
[64] Q.P. Chen, Q.S. Xie, Q.N. Yuan, H.S. Huang, Y.T. Li, Research on a real-time
[38] S. Werbińska-Wojciechowska, K. Winiarska, Maintenance performance in the
monitoring method for the wear state of a tool based on a convolutional
age of industry 4.0: A bibliometric performance analysis and a systematic
bidirectional LSTM model, Symmetry- Basel 11 (10) (2019).
literature review, Sensors 23 (3) (2023) 1409.
[65] H.L. Xu, R.Z. Ma, L. Yan, Z.M. Ma, Two-stage prediction of machinery fault
[39] S.K. Jagatheesaperumal, M. Rahouti, K. Ahmad, A. Al-Fuqaha, M. Guizani,
trend based on deep learning for time series analysis, Digit. Signal Process.
The duo of artificial intelligence and big data for industry 4.0: Applications,
117 (2021).
techniques, challenges, and future research directions, IEEE Internet Things J.
[66] J.L. Yan, J.M. Kan, H.F. Luo, Rolling bearing fault diagnosis based on Markov
9 (15) (2022) 12861–12885.
transition field and residual network, Sensors 22 (10) (2022).
[40] M. Javaid, A. Haleem, R.P. Singh, R. Suman, Artificial intelligence applications
[67] G.R. Garcia, G. Michau, M. Ducoffe, J.S. Gupta, O. Fink, Temporal signals to
for industry 4.0: A literature-based study, J. Ind. Integr. Manag. 07 (01) (2021)
images: Monitoring the condition of industrial assets with deep learning image
83–111.
processing algorithms, Proc. Inst. Mech. Eng. O- J. Risk Reliab. 236 (4) (2022)
[41] A. Bousdekis, K. Lepenioti, D. Apostolou, G. Mentzas, Data analytics in quality 617–627.
4.0: Literature review and future research directions, Int. J. Comput. Integr. [68] J. Nam, J. Kang, Classification of chaotic signals of the recurrence matrix
Manuf. 36 (5) (2022) 678–701. using a convolutional neural network and verification through the Lyapunov
[42] M. Sanchez, E. Exposito, J. Aguilar, Autonomic computing in manufacturing exponent, Appl. Sci.- Basel 11 (1) (2021).
process coordination in industry 4.0 context, J. Ind. Inf. Integr. 19 (2020) [69] R. Kou, S.W. Lian, N. Xie, B.E. Lu, X.M. Liu, Image-based tool condition
100159. monitoring based on convolution neural network in turning process, Int. J.
[43] Z. Huang, A. Shahzadi, Y.D. Khan, Unfolding the impact of quality 4.0 practices Adv. Manuf. Technol. 119 (5–6) (2022) 3279–3291.
on industry 4.0 and circular economy practices: A hybrid SEM-ANN approach, [70] M.C. Yesilli, F.A. Khasawneh, B.P. Mann, Transfer learning for autonomous
Sustainability 14 (23) (2022) 15495. chatter detection in machining, J. Manuf. Process. 80 (2022) 1–27.
[44] A. Saihi, M. Awad, M. Ben-Daya, Quality 4.0: Leveraging industry 4.0 tech- [71] Z. Mezni, C. Delpha, D. Diallo, A. Braham, Performance of bearing ball defect
nologies to improve quality management practices - a systematic review, Int. classification based on the fusion of selected statistical features, Entropy 24 (9)
J. Qual. Reliab. Manag. 40 (2) (2023) 628–650. (2022).
[45] H.C. Liu, R. Liu, X.Z. Gu, M.Y. Yang, From total quality management to quality [72] X.Z. Zhao, J.S. Cheng, P. Wang, Y. Yang, A rotating machinery fault diagnosis
4.0: A systematic literature review and future research agenda, Front. Eng. method using composite multiscale fuzzy distribution entropy and minimal error
Manag. (2023). of convex hull approximation, Meas. Sci. Technol. 32 (2) (2021).
[46] T. Yuan, K.H. Adjallah, A. Sava, H. Wang, L. Liu, Issues of intelligent data [73] D.Y. Han, K. Liang, P.M. Shi, Intelligent fault diagnosis of rotating machinery
acquisition and quality for manufacturing decision-support in an industry based on deep learning with feature selection, J. Low Freq. Noise Vib. Act.
4.0 context, in: 2021 11th IEEE International Conference on Intelligent Data Control. 39 (4) (2020) 939–953.
Acquisition and Advanced Computing Systems: Technology and Applications, [74] A.S. Minhas, N. Sharma, G. Singh, P.K. Kankar, S. Singh, Improvement in
IDAACS, IEEE, 2021. classification accuracy and computational speed in bearing fault diagnosis using
[47] T. Xi, I.M. Benincá, S. Kehne, M. Fey, C. Brecher, Tool wear monitoring in multiscale fuzzy entropy, J. Braz. Soc. Mech. Sci. Eng. 42 (11) (2020).
roughing and finishing processes based on machine internal data, Int. J. Adv. [75] T. Dovedi, R. Upadhyay, Diagnosis of ball bearing faults using double
Manuf. Technol. 113 (11–12) (2021) 3543–3554. decomposition technique, Int. J. Acoust. Vib. 25 (3) (2020) 327–340.
[48] E.A. Hinojosa-Palafox, O.M. Rodriguez-Elias, J.A. Hoyo-Montano, J.H. Pacheco- [76] J.D. Zheng, H.Y. Pan, J.S. Cheng, Rolling bearing fault detection and diagnosis
Ramirez, Towards an architectural design framework for data management in based on composite multiscale fuzzy entropy and ensemble support vector
industry 4.0, in: 2019 7th International Conference in Software Engineering machines, Mech. Syst. Signal Process. 85 (2017) 746–759.
Research and Innovation, CONISOFT, IEEE, 2019. [77] J.Y. Wang, F.G. Li, W.H. Du, Z.J. Wang, Rolling bearing diagnosis method
[49] R.K. Waddell, T.W. Fry, The affordably connected factory: A brief evaluation based on improved standardized variable distance fusion hierarchical state space
of sensors and hardware deployed in industrial applications, Procedia Manuf. correlation entropy, Meas. Sci. Technol. 33 (1) (2022).
53 (2021) 741–747. [78] T. Han, J.C. Gong, X.Q. Yang, L.Z. An, Fault diagnosis of rolling bearings using
[50] B. Tanane, B. Dafflon, M.L. Bentaha, N. Moalla, V. Ferreiro, Toward a col- dual-tree complex wavelet packet transform and time-shifted multiscale range
laborative sensor network integration for smes’ zero-defect manufacturing, in: entropy, Ieee Access 10 (2022) 59308–59326.
L.M. Camarinha-Matos, A. Ortiz, X. Boucher, A.L. Osório (Eds.), Collaborative [79] L.K. Zheng, Y. He, X.A. Chen, Research on a fault diagnosis method for rolling
Networks in Digitalization and Society 5.0, Springer International Publishing, bearing based on improved multiscale range entropy and hierarchical prototype,
Cham, 2022, pp. 31–43. Meas. Sci. Technol. 32 (9) (2021).

15
B. Tanane et al. Journal of Industrial Information Integration 45 (2025) 100833

[80] C. Heistracher, A. Jalali, I. Strobl, A. Suendermann, S. Meixner, S. Holly, [104] B. Wu, W. Li, M.Q. Qiu, Remaining useful life prediction of bearing with
D. Schall, B. Haslhofer, J. Kemnitz, Transfer learning strategies for anomaly vibration signals based on a novel indicator, Shock. Vib. 2017 (2017).
detection in IoT vibration data, in: Iecon 2021 - 47th Annual Conference of [105] Y.B. Liu, B. He, F. Liu, S.L. Lu, Y.L. Zhao, J.W. Zhao, Remaining useful life
the Ieee Industrial Electronics Society, 2021. prediction of rolling bearings using PSR, JADE, and extreme learning machine,
[81] A. Puchalski, I. Komorska, Data-driven monitoring of the gearbox using multi- Math. Probl. Eng. 2016 (2016).
fractal analysis and machine learning methods, in: III International Conference [106] J. Ben Ali, B. Chebel-Morello, L. Saidi, S. Malinowski, F. Fnaiech, Accurate
of Computational Methods Engineering Science (Cmes 18), vol. 252, 2019. bearing remaining useful life prediction based on Weibull distribution and
[82] Y. Liu, J.H. Zhang, K.J. Qin, Y.Y. Xu, Diesel engine fault diagnosis using intrinsic artificial neural network, Mech. Syst. Signal Process. 56–57 (2015) 150–172.
time-scale decomposition and multistage adaboost relevance vector machine, [107] A. Soualhi, K. Medjaher, N. Zerhouni, Bearing health monitoring based on
Proc. Inst. Mech. Eng. C- J. Mech. Eng. Sci. 232 (5) (2018) 881–894. Hilbert-huang transform, support vector machine, and regression, Ieee Trans.
[83] C.C. Jin, X. Chen, An end-to-end framework combining time-frequency ex- Instrum. Meas. 64 (1) (2015) 52–62.
pert knowledge and modified transformer networks for vibration signal [108] M.C. Yesilli, F.A. Khasawneh, A. Otto, Topological feature vectors for chatter
classification, Expert Syst. Appl. 171 (2021). detection in turning processes, Int. J. Adv. Manuf. Technol. 119 (9–10) (2022)
[84] K.H. Zhu, X.C. Yue, D.J. Sun, S.C. Xiao, X. Hu, Rolling bearing fault feature 5687–5713.
extraction using local maximum synchrosqueezing transform and global fuzzy
[109] Q. Hu, Q. Zhang, X.S. Si, G.X. Sun, A.S. Qin, Intelligent fault diagnosis
entropy, Int. J. Acoust. Vib. 27 (1) (2022) 37–44.
approach based on composite multi-scale dimensionless indicators and affinity
[85] Q.T. Chen, Y.J. Huang, Y.R. Song, The research of object tracking filter fault
propagation clustering, Ieee Sens. J. 20 (19) (2020) 11439–11453.
diagnosis method based on reconstruction of feature vector in fractional Fourier
[110] M.C. Yesilli, F.A. Khasawneh, A. Otto, On transfer learning for chatter detection
transform domain, Adv. Mech. Eng. 10 (10) (2018).
[86] P. Shakya, A.K. Darpe, M.S. Kulkarni, Bearing damage classification using in turning using wavelet packet transform and ensemble empirical mode
instantaneous energy density, J. Vib. Control 23 (16) (2017) 2578–2618. decomposition, Cirp J. Manuf. Sci. Technol. 28 (2020) 118–135.
[87] Z. Chen, A. Mauricio, W. Li, K. Gryllias, A deep learning method for bearing [111] A. Sadhu, G. Prakash, S. Narasimhan, A hybrid hidden Markov model to-
fault diagnosis based on cyclic spectral coherence and convolutional neural wards fault detection of rotating components, J. Vib. Control 23 (19) (2017)
networks, Mech. Syst. Signal Process. 140 (2020) 106683. 3175–3195.
[88] F. Jia, Y. Lei, J. Lin, X. Zhou, N. Lu, Deep neural networks: A promising tool [112] S.A. Khan, J.-M. Kim, Rotational speed invariant fault diagnosis in bearings
for fault characteristic mining and intelligent diagnosis of rotating machinery using vibration signal imaging and local binary patterns, J. Acoust. Soc. Am.
with massive data, Mech. Syst. Signal Process. 72–73 (2016) 303–315. 139 (4) (2016) EL100–EL104.
[89] B. Yang, Y. Lei, F. Jia, S. Xing, An intelligent fault diagnosis approach based on [113] A.S. Maliuk, A.E. Prosvirin, Z. Ahmad, C.H. Kim, J.M. Kim, Novel bearing fault
transfer learning from laboratory bearings to locomotive bearings, Mech. Syst. diagnosis using Gaussian mixture model-based fault band selection, Sensors 21
Signal Process. 122 (2019) 692–706. (19) (2021).
[90] S. Ahmad, Z. Ahmad, J.M. Kim, A centrifugal pump fault diagnosis framework [114] Y.L. Zhang, L.X. Duan, M.H. Duan, A new feature extraction approach using im-
based on supervised contrastive learning, Sensors 22 (17) (2022). proved symbolic aggregate approximation for machinery intelligent diagnosis,
[91] R. Zhu, M.X. Wang, S.Y. Xu, K. Li, Q.P. Han, X. Tong, K.Y. He, Fault diagnosis
Measurement 133 (2019) 468–478.
of rolling bearing based on singular spectrum analysis and wide convolution
[115] Z. Zhang, Y. Qin, L.M. Jia, X.A. Chen, Visibility graph feature model of vibration
kernel neural network, J. Low Freq. Noise Vib. Act. Control. (2022).
signals: A novel bearing fault diagnosis approach, Materials 11 (11) (2018).
[92] S.S. Roy, S. Chatterjee, S. Roy, P. Bamane, A. Paramane, U.M. Rao, M.T. Nazir,
[116] M. Islam, A. Prosvirin, J.M. Kim, Intelligent rub-impact fault diagnosis based
Accurate detection of bearing faults using difference visibility graph and bi-
directional long short-term memory network classifier, Ieee Trans. Ind. Appl. on genetic algorithm-based IMF selection in ensemble empirical mode de-
58 (4) (2022) 4542–4551. composition and diverse features models, in: Intelligent Data Engineering and
[93] K. Vos, Z.X. Peng, C. Jenkins, M.R. Shahriar, P. Borghesani, W.Y. Wang, Automated Learning Ideal 2018, Pt I, vol. 11314, 2018, pp. 147–155.
Vibration-based anomaly detection using LSTM/SVM approaches, Mech. Syst. [117] J.X. Ma, F.Y. Xu, K. Huang, R. Huang, GNAR-GARCH model and its application
Signal Process. 169 (2022). in feature extraction for rolling bearing fault diagnosis, Mech. Syst. Signal
[94] J. Li, Y.B. Liu, Q.J. Li, Generative adversarial network and transfer-learning- Process. 93 (2017) 175–203.
based fault detection for rotating machinery with imbalanced data condition, [118] G. Georgoulas, P. Karvelis, T. Loutas, C.D. Stylios, Rolling element bearings
Meas. Sci. Technol. 33 (4) (2022). diagnostics using the symbolic aggregate approximation, Mech. Syst. Signal
[95] F. Zonzini, A. Carbone, F. Romano, M. Zauli, L. De Marchi, Machine learning Process. 60–61 (2015) 229–242.
meets compressed sensing in vibration-based monitoring, Sensors 22 (6) (2022). [119] Y.H. Zhang, T.T. Zhou, X.F. Huang, L.C. Cao, Q. Zhou, Fault diagnosis of
[96] K. Patra, R.N. Sethi, D.K. Behera, Anomaly detection in rotating machinery rotating machinery based on recurrent neural networks, Measurement 171
using autoencoders based onbidirectional LSTM and GRU neural networks, (2021).
Turk. J. Electr. Eng. Comput. Sci. 30 (4) (2022) 1637–1653. [120] A.N. Wang, Y.S. Li, Z. Yao, C.Q. Zhong, B. Xue, Z.N. Guo, A novel hybrid
[97] L.J. Chen, J.Y. Fu, Y.H. Mei, D. Huang, C.T. Ng, H.D. Yao, Effects of operational
model for the prediction and classification of rolling bearing condition, Appl.
variability and damage on structural response signals: A method based on LMS
Sci.- Basel 12 (8) (2022).
radar image and residual-permutation entropy, Eng. Struct. 265 (2022).
[121] A.E. Prosvirin, Z. Ahmad, J.M. Kim, Global and local feature extraction using
[98] D. Miki, K. Demachi, Bearing fault diagnosis using weakly supervised long
a convolutional autoencoder and neural networks for diagnosing centrifugal
short-term memory, J. Nucl. Sci. Technol. 57 (9) (2020) 1091–1100.
pump mechanical faults, Ieee Access 9 (2021) 65838–65854.
[99] X.B. Wang, X.Y. Zhang, Z. Li, J. Wu, Ensemble extreme learning machines
for compound-fault diagnosis of rotating machinery, Knowl.-Based Syst. 188 [122] L. Han, C.C. Yu, C.L. Liu, Y. Qin, S.J. Cui, Fault diagnosis of rolling bearings
(2020). in rail train based on exponential smoothing predictive segmentation and
[100] J.J. Saucedo-Dorantes, M. Delgado-Prieto, R.D. Romero-Troncoso, R.A. Osornio- improved ensemble learning algorithm, Appl. Sci.- Basel 9 (15) (2019).
Rios, Multiple-fault detection and identification scheme based on hierarchical [123] Y.L. Zhang, Y.S. Zhou, M.L. Duan, L.X. Duan, X. Zhang, L.Y. Jiang, A novel
self-organizing maps applied to an electric machine, Appl. Soft Comput. 81 feature extraction algorithm for bearing fault diagnosis based on enhanced
(2019). symbolic aggregate approximation, J. Intell. Fuzzy Systems 36 (6) (2019)
[101] S. Radhakrishnan, Y.T.T. Lee, S. Rachuri, S. Kamarthi, Complexity and entropy 5369–5381.
representation for machine component diagnostics, PLoS One 14 (7) (2019). [124] I. Aydin, M. Karakose, E. Akin, Combined intelligent methods based on wireless
[102] Z.Y. Tang, Z.C. Chen, Y.Q. Bao, H. Li, Convolutional neural network-based sensor networks for condition monitoring and fault diagnosis, J. Intell. Manuf.
data anomaly detection method using multiple information for structural health 26 (4) (2015) 717–729.
monitoring, Struct. Control. Heal. Monit. 26 (1) (2019).
[103] P. Ma, H.L. Zhang, W.H. Fan, C. Wang, Novel bearing fault diagnosis model
integrated with dual-tree complex wavelet transform, permutation entropy and
optimized FCM, J. Vibroeng. 20 (2) (2018) 891–908.

16