Accident Prevention in Industry 4.0 Using Retroﬁt: A Proposal

Paulo Henrique Mariano

, Bruno Pinho Campos

, Frederico Augusto Cardozo Diniz

Carlos Frederico Cavalcanti

and Ricardo Augusto Rabelo Oliveira

Computing Department (DECOM), Federal University of Ouro Preto (UFOP), Ouro Preto 35400-000, Brazil

Keywords:

Safe, Artiﬁcial Intelligence, Pattern Recognition, Machine Learning, Retroﬁt, Industry 4.0.

Abstract:

In this work, we present an Industry 4.0 retroﬁt solution to prevent accidents in industrial environments, specif-

ically focusing on the operation of bandsaw machines. It examines a real-world scenario where a company

aims to enhance worker safety by implementing an integrated solution. The proposed solution involves a

pattern recognition system that monitors the work area and sends commands to stop the machine in case of

dangerous movements near the bandsaw. This system adheres to Industry 4.0 principles, demonstrating how

this methodology can create a safer industrial environment to connect information technology (IT) and opera-

tional technology (OT).

1 INTRODUCTION

Despite all the technology involved and discussed,

factories still have a lot of manual labor, and some

are dangerous tasks that can cause accidents. In 2020,

the Ministry of Labor data showed 465.772 work-

place accidents in Brazil. Of these, 47.293 involved

wrist, hand, and ﬁnger injuries, representing 10.15

percent of total accidents, making them the most com-

mon type of workplace injury. These injuries include

cuts, fractures, and amputations(Social and Security,

2023).

Serious workplace accidents have signiﬁcant im-

pacts on workers. These impacts include physical in-

juries, psychological distress, economic hardship, and

social vulnerability. Workers often face long recovery

periods, permanent disabilities, and associated psy-

chological trauma. The ﬁnancial impacts are severe,

especially for informal workers who lack social pro-

tection and beneﬁts. The burden on families is sub-

stantial, as they often need to provide care without

adequate resources, exacerbating social and economic

vulnerabilities(Menezes and Magro, 2023).

In recent times, Industry 4.0 concepts have be-

come a focal point for many companies striving to

https://orcid.org/0009-0001-0505-3879

https://orcid.org/0009-0006-8124-0998

https://orcid.org/0009-0007-6977-2411

https://orcid.org/0000-0002-8522-6345

https://orcid.org/0000-0001-5167-1523

Figure 1: Day of Work in an Band Saw Machine.

remain competitive in the modern market. Industry

4.0, often referred to as the Fourth Industrial Rev-

olution, encompasses the integration of Information

Technology (IT) and Operational Technology (OT) to

create smart factories(Dalenogare et al., 2018). This

Mariano, P. H., Campos, B. P., Diniz, F. A. C., Cavalcanti, C. F. and Oliveira, R. A. R.

Accident Prevention in Industry 4.0 Using Retroﬁt: A Proposal.

DOI: 10.5220/0013473900003929

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 2, pages 649-656

ISBN: 978-989-758-749-8; ISSN: 2184-4992

649

fusion aims to improve productivity, optimize pro-

cesses, and signiﬁcantly reduce waste of materials

and money. Industry 4.0 leverages advanced tech-

nologies such as the Internet of Things (IoT), artiﬁ-

cial intelligence (AI), machine learning, big data an-

alytics, and blockchain to create interconnected sys-

tems. These technologies facilitate real-time moni-

toring, predictive maintenance, and seamless commu-

nication at all levels of production(Dalenogare et al.,

2018). Beyond these capabilities, this technology can

also improve the safety and security environment of

industrial workers and employees.

This work proposes a novel Industry 4.0 retroﬁt

solution speciﬁcally designed to prevent accidents in

industrial environments by focusing on the opera-

tion of band saw machines. Unlike traditional safety

mechanisms, which often rely solely on reactive mea-

sures, this solution introduces an integration of ma-

chine learning-based pattern recognition to monitor

and enhance worker safety proactively. The machine

learning system continuously monitors the work area

for dangerous movements, sending immediate com-

mands to stop the machine in case of a potential haz-

ard, as depicted in Figure 9.

A key scientiﬁc contribution of this work is the

proposal of integrating technology into mechanical

machines to enhance not only productivity but also

safety.

This combination of real-time hazard detection

represents a signiﬁcant advancement in industrial

safety systems, ensuring immediate and long-term

protection in high-risk environments. The proposed

methodology contributes to the ﬁeld by providing a

scalable solution that can be adapted to various in-

dustrial processes, establishing a new standard for ac-

cident prevention and safety management in Industry

4.0 environments.

2 RELATED WORK

This section presents articles about using pattern

recognition in an industry 4.0 context. The paper

(Serey et al., 2023) highlights the growing impor-

tance of PR and DL due to the increased volume of

data generated and the need for efﬁcient management

of this data in asymmetric structures. Examines re-

cent PR and DL applications, evaluating their ability

to process large volumes of data and identify signiﬁ-

cant patterns. The review includes an in-depth study

of 186 references, selecting 120 for detailed analy-

sis. PR and DL are integrated with various computa-

tional techniques, such as data mining, data science,

and cognitive systems, to improve automation and

reduce human intervention in decision-making pro-

cesses. Challenges include the need for large datasets

for practical training, the high cost of computational,

and the difﬁculty in interpreting deeply nested mod-

els.

The article (Zouhal et al., 2021) Provides a com-

prehensive view of how industrial inspection can be

transformed within the context of Industry 4.0. It fo-

cuses on adapting industrial inspection to new tech-

nologies that allow for an agile and personalized

conﬁguration and integrating cameras embedded in

robots to automate visual inspection. Discusses the

state of the art in visual computing applied to in-

dustrial inspection, highlighting how techniques Ad-

vanced image processing and machine learning are

transforming inspection equipment and quality con-

trol. The case study shows how cameras embedded in

industrial robots can perform image acquisitions and

processing to detect wear on tools and other anoma-

lies in a production environment. The suggestion

is that data collected during automated inspections

could be used to train predictive maintenance algo-

rithms.

(Oborski and Wysocki, 2022) explores the use

of convolutional neural networks (CNNs), to im-

prove visual quality for control systems in Industry

4.0. Several AI algorithms were evaluated to identify

those most suitable for visual quality control tasks in

a real manufacturing environment. CNNs were cho-

sen for further study because of their effectiveness

in processing images and performing complex clas-

siﬁcation and detection tasks. CNN algorithms were

tested with production line datasets and photos. In

which he showed efﬁciency.

The paper (Sharma and Sethi, 2024)show high-

lights the importance of agriculture in the economy

and how plant diseases can signiﬁcantly affect agri-

cultural production. Recognizes the need for effec-

tive methods to detect diseases early to reduce losses.

Utilizes deep learning techniques, speciﬁcally CNNs,

which are well suited to data processing tasks im-

age processing due to its ability to extract essential

features from large datasets of images. The model

allows for a more precise identiﬁcation of areas af-

fected by wheat leaf diseases, separating them pixel

by pixel. The article presents a signiﬁcant improve-

ment in classifying diseases in wheat leaves, achiev-

ing a 99.43 percent accuracy using Point Rend seg-

mentation, compared to 96.08 percent without seg-

mentation.

The article (Li et al., 2023) presents an improved

YOLOv8-based algorithm for detecting glove use

in industrial environments, addressing the speciﬁc

challenges of detecting small objects like hands and

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

650

Figure 2: Industry 4.0 Retroﬁt proposal.

gloves in complex factory settings. The study in-

troduces YOLOv8-AFPN-M-C2f, a modiﬁed version

of YOLOv8 that replaces the standard PAFPN with

the Asymptotic Feature Pyramid Network (AFPN)

to improve the integration of non-adjacent feature

layers and enhance the detection of smaller objects.

The paper also incorporates a new shallow feature

layer to increase sensitivity to detailed visual ele-

ments, such as gloves. Using a custom dataset of

2,695 high-resolution images from industrial environ-

ments, the study demonstrates a 2.6 percent increase

in mAP@50 percent, a 63.8 percent improvement in

FPS, and a 13 percent reduction in parameters com-

pared to the original YOLOv8 model. These improve-

ments highlight the potential for more accurate and

efﬁcient detection of protective gloves in industrial

contexts, contributing to enhanced worker safety and

automated monitoring of personal protective equip-

ment (PPE) compliance.

Although these articles have focused on product

quality and how various techniques can enhance it,

this is well-established in the industry. The empha-

sis on applying these principles to human safety at

work is sometimes neglected. Our study approaches

this topic by examining the use of technology from a

security perspective with integration in industrial ma-

chines.

2.1 Machine Learning

The ﬁrst proposal, as described before, is a computer

vision system based on machine learning to improve

the recognition of hands with graphene gloves. This

study presents three types of machine learning com-

puter vision technologies(MediaPipe, Detectron2 and

Yolov8) and compares them to identify which is better

for this type of project.

The recognition time is crucial because it deﬁnes

the speed of the signal sending. The initial algorithm

is calculated based on this. The time of response

of the algorithm when some point crosses the line is

given by:

= T

− T

(1)

Where:

• T

is the detection time,

• T

is the ﬁnal time (the moment when the point

crosses the line),

• T

is the initial time (the moment when the image

processing starts).

3 DEVELOPMENT

A Proof-of-Concept (POC) was developed, consist-

ing of a computer vision system and a communica-

tion interface with the machine, to validate the solu-

tion and integrate various technologies, as described

below. OpenCV (Open Source Computer Vision Li-

brary) is a highly optimized open-source library that

implements real-time computer vision and machine

learning algorithms . It is used to capture video in

real-time, process the images, and convert them into

different color formats to facilitate the recognition

(Bradski and Kaehler, 2000). In the project, it is used

to video capture and draw the ”dangerous area”.

3.1 MediaPipe

MediaPipe, developed by Google, is a framework for

building multimodal (e.g., video, audio, and sensor

data) machine learning pipelines. It provides ready-

to-use solutions for detecting and tracking body parts.

In the algorithm, it is used to identify the hands and

ﬁngers. (Research, 2023)

Accident Prevention in Industry 4.0 Using Retroﬁt: A Proposal

651

Figure 3: Hand Recognition With MediaPipe.

The algorithm works as follows:

• Image Capture: OpenCV captures real-time video

from the camera. The captured images are ﬂipped

for a selﬁe view and converted from BGR to RGB

format, which is the format expected by Medi-

aPipe.

• Hand Detection and Tracking: MediaPipe is con-

ﬁgured to detect hands with a minimum detec-

tion conﬁdence threshold. MediaPipe processes

the image for each captured frame to detect hands

and identify speciﬁc landmarks on each hand.

• Safety Veriﬁcation: The image includes a safety

line. Each hand landmark is checked against this

safety line, which covers an area of 8 centime-

ters and needs to be placed physically at the cor-

rect distance from the band saw. If any landmark

crosses the safety line, a signal is sent to analogic

relay thats stop the machine.

The MediaPipe function works well with bare

hands, creating landmarks in points of hands, but it

has recognition problems with gloves, resulting in

intermittent performance. This lack of recognition

poses a dangerous problem in this context. MediaPipe

uses a dataset that does not include the speciﬁc gloves

used in this work and in the factory, which may cause

these recognition issues. We particularly notice the

intermittence in areas where the gloves have a contrast

of gray and black, like can be observe in the Figure 6.

3.2 Detectron2

To enhance recognition capability, we chose Detec-

tron2, a library developed by Facebook AI Research

(FAIR) for object detection, segmentation, and other

visual recognition tasks (Wu et al., 2019). It’s created

and used our own dataset of 1200 images of hands

with gloves (specially graphene gloves). The chose

Detectron2 because of its robust resources for iden-

tifying and segmenting images in complex scenarios.

Figure 4: Hand Recognition With MediaPipe, issues in use

of gloves.

However, this power requires a machine with high-

level hardware.

In our project, after creating the dataset, its pre-

pared the environment and trained the model based on

our dataset. Its chose the Mask R-CNN-50 algorithm

for training, aiming to segment the gloves in the im-

ages, based on this related good accuracy (Tahir et al.,

2021). Image segmentation is necessary to recognize

only the gloves and not other parts manipulated by the

operator.

Figure 5: Hand Recognition( in Red ) throw by segmenta-

tion with Detectron2.

The recognition results with Detectron2 were

good, as it was able to recognize hands with different

types of gloves (and most important to the project the

graphene gloves). When an image is segmented and

recognized, it is painted in red. We combined Detec-

tron2 with OpenCV to provide the solution. OpenCV

was used to capture real-time images and draw the

”dangerous area,” a rectangle to simulate the place of

the saw. When the segmentation area crosses the dan-

gerous area, the rectangle is also painted in red.

However, the performance of the solution was

poor. The frames per second (fps) were less than

one, compromising the effectiveness of the solution

in real time. Other algorithms based on Detectron2

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

652

were tested, but we could not achieve success in im-

proving performance. We reduced the size of the im-

ages to 640x480, but the fps did not improve. This

indicates that more computing power is needed to use

this solution effectively.

Figure 6: The output during the execution of Detectron2

showed intermittent FPS, highlighting the performance is-

sues.

3.3 YOLOv8

Figure 7: Result the training of model with YOLOv8.

Then we tried other techniques and decided to use

YOLOv8 to recognize the hands. The approach with

YOLOv8 was chosen due to its capability to clas-

sify and segment images, especially in real-time, with

good FPS and accuracy (Jocher et al., 2023).

We started with training using a dataset of 1320

images of gloves in special graphene gloves in differ-

ent positions, qualities, and lighting conditions. The

algorithm performed well, with good accuracy and

FPS, which led us to invest in this solution. The pa-

rameters to the training is describe in the table 1:

The graphs in the Figure 9, shows the results of

metrics and losses during the training and validation

of a model conﬁgured with 50 epochs, for the task of

object detection and/or segmentation of one class (in

this case, gloves).

Columns: Each column represents different met-

rics or types of loss. Rows (Train/Val):

Figure 8: Solution Developed with YOLOv8.

Table 1: Training Parameters.

Parameter Value

Model yolov8n.pt

Epochs 50

Image Size 320

Batch Size 16

Workers 8

The top row refers to the training results. The bot-

tom row refers to the validation results.

- The train/box-loss and val/box-loss represents

the loss related to bounding boxes. Loss should de-

crease over epochs, indicating the model is learning

to predict boxes correctly and int the model there is

consistent reduction, which is a positive sign.

- The train/seg-loss and val/seg-loss: Related

to segmentation loss (e.g., segmentation masks for

gloves). The expected behavior is gradual decrease

over epochs.Despite some initial ﬂuctuations in val-

idation loss, it shows convergence, suggesting im-

provement.

- The train/cls-loss and val/cls-loss refers to clas-

siﬁcation loss, indicating how well the model identi-

ﬁes objects as ”glove.”. The loss decreases, but more

pronounced peaks in validation suggest potential data

inconsistencies or overﬁtting.

- The train/dﬂ-loss and val/dﬂ-loss related to the

prediction of distribution locations (distillation loss).

The expected behavior is a continuous decrease.

While loss decreases overall, peaks in validation in-

dicate challenges with certain instances.

Accident Prevention in Industry 4.0 Using Retroﬁt: A Proposal

653

Figure 9: Results of training with YOlOv8.

- The metrics/precision(B) and metrics/

precision(M): Precision measures the proportion

of correct detections relative to all detections. B

and M might represent different conﬁgurations (e.g.,

bounding boxes vs. masks). The precision increases

rapidly and stabilizes above 0.8, which is excellent.

- The metrics/recall(B) and metrics/recall(M): Re-

call measures the proportion of correctly identiﬁed

objects relative to all existing objects. In th e training

the recall improves and stabilizes above 0.8, indicat-

ing good coverage.

- metrics/mAP50(B) and metrics/mAP50(M):

mAP50 (mean Average Precision) evaluates average

precision with an IoU threshold of 50The results sta-

bilize above 0.8, showing strong glove detection per-

formance.

- The metrics/mAP50-95(B) and metrics/mAP50-

95(M): mAP50-95 is a stricter metric considering

multiple IoU thresholds (from 50 percent to 95 per-

cent). While more challenging, values stabilize above

0.6, which is good for complex object detection tasks.

The results obtained over 50 epochs demonstrate

consistent model performance. The training and val-

idation losses (box-loss, cls-loss, seg-loss, and dﬂ-

loss) gradually decreased, indicating effective learn-

ing. Despite some peaks in validation losses, the over-

all trend shows convergence, suggesting good gener-

alization.

Performance metrics, such as precision and re-

call, stabilized above 0.8, indicating high reliability

in glove detection. The mAP50 metric showed strong

results ( 0.8), while the more stringent mAP50-95 re-

mained around 0.6, reﬂecting the inherent challenges

of the task.

The validation loss, although following a trajec-

tory similar to the training loss, shows some spikes

(particularly in val/cls-loss and val/dﬂ-loss), which

may be related to outliers or speciﬁc characteristics of

the validation set that make it difﬁcult for the model to

generalize. These spikes are expected in complex ob-

ject detection problems and do not compromise over-

all performance, given that the global trend of the

losses is towards convergence. In practice the results

were observer, with a good balance between accuracy

and performance.

After training the model, it was combined with

OpenCV to capture and analyze video in real time for

detection. OpenCV was also used to draw a ”dan-

gerous area,” a rectangle around the saw of the band

saw machine. When any part of the glove’s mask seg-

mentation enters the rectangle, the rectangle changes

color to red, an alert sound is triggered, and a signal

is sent to an analog relay. The Python library Tkin-

ter was used to create controls for adjusting the width

and height of the ”dangerous area,” allowing occupa-

tional safety engineers to deﬁne the dimensions of this

safety zone. Additionally, the recognition box was re-

moved, leaving only the mask segmentation to avoid

cluttering the screen and prevent false alerts.

To ensure that the system runs smoothly without

interruptions or delays, the sound signal processing

was handled in a separate, independent process from

the recognition process. This means that while one

process is dedicated to detecting and segmenting the

gloves in real time, another process is responsible for

managing the sound alert system and sending the sig-

nal to the relay. This separation helps prevent any

potential bottlenecks or slowdowns in the recognition

process, ensuring that the emergency stop system can

respond quickly and reliably.

The analog relay, activated by the signal from

the independent process, triggers the existing physi-

cal emergency stop system of the machine.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

654

4 RESULTS

The Proof-of-Concept (POC) assessed the system’s

ability to recognize hands and promptly send a stop

signal when any landmark crossed the predeﬁned

safety line. The algorithms tested showed that the

best combination of performance and accuracy was

achieved using YOLOv8.

The YOLOv8 model demonstrated high reliability

in detecting gloves, with performance metrics such

as precision and recall stabilizing above 0.8. The

mAP50 metric showed strong results ( 0.8 ), while

the more stringent mAP50-95 remained around 0.6,

reﬂecting the inherent challenges of the task.

The system was able to process real-time video

and accurately detect when a hand with graphene

gloves entered the dangerous area. The integration

with OpenCV allowed for real-time video capture and

the drawing of the ”dangerous area,” while the Python

library Tkinter provided controls for adjusting the di-

mensions of this safety zone.

The separation of the sound signal processing

into an independent process ensured that the system

ran smoothly without interruptions or delays. This

setup allowed the emergency stop system to respond

quickly and reliably, enhancing the overall safety of

the workplace.

In conclusion, the POC validated the effectiveness

of the proposed solution, demonstrating that it is pos-

sible to integrate advanced computer vision and ma-

chine learning technologies into mechanical machines

to enhance safety.

4.1 Experimental Setup

This initial phase validates the project’s foundational

concepts and technologies. These tests were not per-

formed in the actual deployment scenario but in a con-

trolled laboratory setting, denoted as the POC phase.

A full HD 1080p webcam was used to connect to a

PC via a USB 2.0 port in this setup. The application

and the Modbus simulator ran on the same computer,

which was equipped with an Intel 11th generation i5

processor at 3.2 GHz, 32 GB of RAM, an NVMe SSD

of 512 GB, and a GeForce 4060 GPU with 8 GB. The

solution was also tested with videos from the indus-

trial process.

The project was further tested on a Raspberry Pi 5

(8 GB RAM, Quad-Core Arm Cortex-A76 processor

at 2.4 GHz) with an AI Kit, and it demonstrated good

performance as well.

4.2 Restrictions

The dataset for training the machine learning models

was time-consuming and required capturing a wide

variety of images of gloves in different positions,

lighting conditions, and qualities, especially graphene

gloves, the increase of the dataset can improve the ac-

curacy with other types of gloves. Industrial environ-

ments are typically harsh on devices. The next step of

the project is to implement the solution in the factory

to observe its behavior, particularly with variations in

lighting throughout the day and the presence of dirt.

This aggressive environment can be a challenge for

small devices like Raspberry Pi.

It is essential to raise awareness among workers

about safety. The system does not work alone; adher-

ence to safety rules is necessary to ensure the success

of the project.

5 CONCLUSION

Retroﬁt is an effective solution for adapting existing

equipment to meet the demands of modern industries.

The project applied contemporary concepts such as

machine learning to a standard industrial band saw

machine. The initial test successfully integrated these

technologies into the Industry 4.0 framework. In con-

clusion, the POC shows that this method can be ex-

tended to other machines and industrial processes,

particularly to enhance the safety and security of em-

ployees in dangerous operations.

6 FUTURE WORKS

The work is currently in development, and the results

from the Proof-of-Concept (POC) tests are promis-

ing, indicating a positive direction. The initial proof-

of-concept phase involved controlled experiments in

real-time recognition and videos of the workplace,

which were sufﬁcient to validate the system’s core

functionality and response times. Future work will in-

clude scaling these tests to real-world industrial envi-

ronments and conditions to further reﬁne the system’s

robustness. This includes integrating the application

into small devices and observing the algorithm’s be-

havior, speed, and accuracy in recognizing hands in a

workplace environment. Additionally, an integration

with a private blockchain solution will be developed

to create immutable records of unsafe behaviors.

Accident Prevention in Industry 4.0 Using Retroﬁt: A Proposal

655

REFERENCES

Bradski, G. and Kaehler, A. (2000). Opencv: Open source

computer vision library. https://opencv.org/. Ac-

cessed: 2025-01-19.

Dalenogare, L. S., Benitez, G. B., Ayala, N. F., and Frank,

A. G. (2018). The expected contribution of industry

4.0 technologies for industrial performance. Interna-

tional Journal of Production Economics.

Jocher, G., Chaurasia, A., and Qiu, J. (2023). Ultralytics

yolov8.

Li, S., Huang, H., Meng, X., Wang, M., Li, Y., and Xie,

L. (2023). A glove-wearing detection algorithm based

on improved yolov8. Sensors, 23(24).

Menezes, M. N. and Magro, M. L. P. D. (2023). Psychoso-

cial impacts of serious accidents at work: A look at the

workers monitored by the worker’s health reference

center. Revista Jur

ıdica Trabalho e Desenvolvimento

Humano, 6:1–30.

Oborski, P. and Wysocki, P. (2022). Intelligent visual qual-

ity control system based on convolutional neural net-

works for holonic shop ﬂoor control of industry 4.0

manufacturing systems. Advances in Science and

Technology. Research Journal, 16(2):89–98.

Research, G. (2023). Mediapipe: Cross-platform frame-

work for building perception pipelines. https://

ai.google.dev/edge/mediapipe/solutions/guide. Ac-

cessed: 2025-01-19.

Serey, J., Alfaro, M., Fuertes, G., Vargas, M., Dur

an, C.,

Ternero, R., Rivera, R., and Sabattin, J. (2023). Pat-

tern recognition and deep learning technologies, en-

ablers of industry 4.0, and their role in engineering

research. Symmetry, 15(2):535.

Sharma, T. and Sethi, G. K. (2024). Improving wheat leaf

disease image classiﬁcation with point rend segmen-

tation technique. SN Computer Science, 5:244.

Social and Security (2023). Work acci-

dent and disability. https://www.gov.br/

previdencia/pt-br/assuntos/previdencia-social/

saude-e-seguranca-do-trabalhador/acidente trabalho

incapacidade. Accessed: June 23, 2024.

Tahir, H., Khan, M. S., and Tariq, M. O. (2021). Perfor-

mance analysis and comparison of faster r-cnn, mask

r-cnn and resnet50 for the detection and counting of

vehicles. 2021 International Conference on Comput-

ing, Communication, and Intelligent Systems (ICC-

CIS), pages 587–594.

Wu, Y., Kirillov, A., Massa, F., Lo, W.-Y., and Gir-

shick, R. (2019). Detectron2. https://github.com/

facebookresearch/detectron2.

Zouhal, Z. et al. (2021). Approach for industrial inspection

in the context of industry 4.0. In 2021 International

Conference on Electrical, Computer, Communications

and Mechatronics Engineering (ICECCME), pages 1–

5, Mauritius, Mauritius.

ICEIS 2025 - 27th International Conference on Enterprise Information Systems

656