Design and Implementation of Real-time Anomaly Detection System
based on YOLOv4
DOOHWAN KIM1,2, YO-HAN HAN1,2, JONGPIL JEONG1,*
1Department of Smart Factory Convergence, Sungkyunkwan University
2066, Seobu-ro, Jangan-gu, Suwon-si, Gyeonggi-do, 16419,
REPUBLIC OF KOREA
2AI Machine Vision Smart Factory Lab, Dev
296, Sandan-ro, Danwon-gu, Ansan-si, Gyeonggi-do, 15433,
REPUBLIC OF KOREA
Abstract: - To solve the problem of high-wage employment and unemployment that is constantly occurring in
industrial sites, we designed a real-time anomaly detection system based on YOLOv4 to automate the detection
of defective products at actual manufacturing sites. This contributes to reducing labor costs and increasing work
efficiency in the field. It also contributes to manufacturing data collection and smart factory system construction
by utilizing the established system
Key-Words: - AI Deep Learning, Anomaly Detection, Smart Factory, Supervised Learning, YOLOv4, Edge
Computing, Manufacturing Data Platform
Received: July 22, 2021. Revised: October 11, 2022. Accepted: November 14, 2022. Published: December 29, 2022.
1 Introduction
With the advent of Industry 4.0, new changes are
coming to the manufacturing industry as IT
technology develops. The application of smart
factories based on innovative technologies for
manufacturing competitiveness cannot be delayed
any longer, [1], [2]. Smart factories are meaningful
in that they meet various customer requirements and
improve production efficiency by collecting and
analyzing production information in real-time from
the manufacturing site by breaking away from the
existing mass production method.
After coronavirus Disease-2019 ( COVID-19),
the preference for non-face-to-face work and the rise
in labor costs have made it difficult to secure
manufacturing manpower at the manufacturing site.
Recently, to solve this problem, there is a demand to
entrust the product quality inspection task to an AI-
based vision inspection system, [3]. As it is
necessary to perform the task of determining
whether the product is defective or not Must have a
high level of accuracy Since the appropriate cost has
to be calculated at a level that can be introduced
even in small and medium-sized enterprises, a high-
quality, low-cost and efficient system is required.
This study is YOLOv4, [4], based on the anomaly
detection, [5], inspection method, [6], it inspects
objects moving on the conveyor belt in real-time
and aims to determine whether the object is good or
bad, and to classify products. The proposed system
meets the level of product defect inspection required
by actual manufacturing sites. It can be expected to
have the effect of improving productivity and the
operating cost of the company, [7].
In addition, it aims to link the system to the
Manufacturing Execution System (MES) and
Enterprise Resource Planning (ERP) systems used
in the factory by building the system on an edge
computing device.
The established system supports the monitoring
system to adjust factory production schedules or
respond to emergencies using the collected data
while collecting manufacturing data. Through this, it
contributes to building a smart factory that can be
controlled remotely, and a deep learning trained
model can be applied in real-time in a cloud
environment. And uploading the collected data
contributes to building a manufacturing data
platform so that manufacturing data can be used
more innovatively, [8].
The structure of this paper is as follows. Section 2
describes the related research, and Section 3
describes the design of the real-time anomaly
detection system applied in this paper. Section 4
shows the experimental results of this system, and
Section 5 concludes with a conclusion.
2 Related Work
This section outlines the concepts or techniques
WSEAS TRANSACTIONS on ELECTRONICS
DOI: 10.37394/232017.2022.13.17
Doohwan Kim, Yo-Han Han, Jongpil Jeong
E-ISSN: 2415-1513
130
Volume 13, 2022
used in papers such as anomaly detection or
YOLOv4
2.1 Anomaly Detection
Anomaly detection refers to the problem of
distinguishing between normal samples and
abnormal samples. technology being applied. When
both normal and abnormal samples exist by labeling
in a given training data set, it is called Supervised
Learning, [9].
When a model trained in this way identifies an
object, it is called supervised anomaly detection.
This way high It is mainly used when accuracy is
required, The more diverse the normal and abnormal
samples are, the higher the performance can be
derived, [10].
In the case of semi-supervised learning, it means
that only normal classes are defined and learned.
This can be useful when you want to detect a class
other than the normal class. In the case of
unsupervised learning, no class is defined and only
images are learned to induce the deep learning
model to make its judgment. As a separate class is
not defined, the learning image aims to use only the
normal class or train only the bad class. However, it
is recommended to train only one class of images to
be learned, [11].
In this paper, supervised learning is applied to
the
above concepts. There are many cases where only
one type of defect class is not defined, and each
defect type has a different method to be taken on-
site. For example, if a product is not cut correctly, it
may be necessary to correct the cut. Likewise, if the
color of the product has changed, the material of the
product may require checking the injection
temperature.In addition, statistics on what kind of
defects should occur a lot should be available, and
since high accuracy is required, supervised learning
was selected over semi-supervised or unsupervised
learning even if there were difficulties in learning.
2.2 YOLOv4
You only look once (YOLO) is a state-of-the-art,
realtime object detection system. R-CNN, [12],
which is representative object detection, generates
and learns about 2,000 Region Proposals when
learning images. This consumes a lot of time and
resources for learning speed. To compensate for
these shortcomings, the YOLOv4 model appeared,
and real-time, Highquality and reliable object
detection results can be obtained. Fig. 1 shows the
labeling of predefined objects in images using the
YOLOv4 Model.
Fig. 1: Detected Object by YOLOv4
The image acquired on the conveyor belt must be
discriminated in real-time, and the discriminating
result must be delivered to the discharge unit. In
addition, a long learning rate may interfere with the
input of the field that changes flexibly. Therefore, in
this paper, the YOLOv4 model with fast learning
speed and discrimination speed was adopted.
2.3 Edge Computing
Edge computing, [13] refers to performing
computing at or near the physical location of a user
or data source. By processing computing services at
a location close to the user’s end device, users will
receive faster and more reliable services, and
businesses will benefit from flexible hybrid cloud
computing. Edge computing is one way for
enterprises to distribute data computation and
processing in multiple locations using a common
pool of resources, [14].
Image learning can be performed in the cloud or
on a remote server. However, if image identification
is performed through a remote server or a cloud
server, product identification and communication
with the programmable logic controller (PLC) may
not be performed smoothly due to network speed
and various variables. Accordingly, in this paper, an
edge computing device was built next to a conveyor
belt where products are produced. In addition, it
may be difficult to prepare a separate learning server
in a small-scale factory. Accordingly, a device
capable of learning images was prepared.
Additionally, the Monitoring User Interface (UI)
Program was prepared to be executed so that the
user can check the determined image in real-time.
The device is connected to the Internet and is ready
to be linked to the factory’s MES, ERP system at
any time. In addition, a system that can remotely
check for product defects at the site and instruct
production, suspension, and change of products in
real-time has been linked.
WSEAS TRANSACTIONS on ELECTRONICS
DOI: 10.37394/232017.2022.13.17
Doohwan Kim, Yo-Han Han, Jongpil Jeong
E-ISSN: 2415-1513
131
Volume 13, 2022
2.4 Precision-Recall
The Confusion Matrix is the basis for determining
how accurate the learning of the AI model is, [15],
[16]. To understand the Confusion Matrix, you need
to understand 4 concepts first. True Positive (TP) :
Predicts that an answer that is actually True is True
(correct answer), False Positive (FP) : Actual False
Predict that the correct answer is True (incorrect
answer), False Negative (FN) : Predict the correct
answer that is True as False (Answer incorrect),
True Negative (TN) : Predict the correct answer that
is actually False as False (correct answer) It shows
intuitively, [16]. Recall or Sensitivity (as it is called
in Psychology) is the proportion of Real Positive
cases that are correctly Predicted Positive. This
measures the Coverage of the Real Positive cases by
the +P (Predicted Positive) rule. Its desirable feature
is that it reflects how many of the relevant cases the
+P rule picks up. Conversely, Precision or
Confidence (as it is called in Data Mining) denotes
the proportion of Predicted Positive cases that are
correctly Real Positives. This is what Machine
Learning, Data Mining and Information Retrieval
focus on, but it is totally ignored in ROC analysis. It
can however analogously be called True Positive
Accuracy (TPA), being a measure of the accuracy of
Predicted Positives in contrast with the rate of
discovery of Real Positives, [17].
In this paper, we will apply the concepts of
precision and Recall to determine whether products
identified by deep learning training have been
accurately classified. For accurate learning and
classification of deep learning models, the objective
model evaluation must be performed by grafting the
corresponding concept.
2.5 Manufacturing Data platform
While global manufacturing is becoming more
competitive due to variety of customer demand,
increase in production cost and uncertainty in
resource availability, the future ability of
manufacturing industries depends upon the
implementation of Smart Factory. With the
convergence of new information and
communication technology, Smart Factory enables
manufacturers to respond quickly to customer
demand and minimize resource usage while
maximizing productivity performance, [8].
Fig. 2: A big data analytics platform architecture
for manufacturing systems, [8].
In addition, there are also cases where new smart
factories have been established based on
manufacturing data in introducing companies that
focus on the personalized cosmetics manufacturing
industry, [18]. Manufacturing data is now attracting
attention as useful data that cannot be discarded in
the 4th industrial era, and it is important to secure
data collected in all production processes. Fig. 2
shows the smart factory platform architecture of the
big data-based manufacturing site system.
It collects good and bad production of products
through edge computing equipment. It also collects
product production information and set information.
It can be used to build a manufacturing data
platform using the collected information and apply
it to build an artificial intelligence-based smart
factory system.
3 YOLOv4-Based Real-Time Anomaly
Detection System
In this paper, the composition of an optical system
that can determine whether a sticker attached to the
bottom of a cosmetic container is defective among
plastic injection (molded) products and image
reading through deep learning training was studied.
Shows the conceptual diagram of the system
proposed in this paper. Due to the material
characteristics of the inspected product Since diffuse
reflection occurs due to the reflection of light, it is
configured as an indirect lighting type that
minimizes reflection of light for optimal
configuration of the optical system. Similarly, to
avoid diffuse reflection of the product, the camera
angle was corrected by 2 to 5 degrees so that it was
WSEAS TRANSACTIONS on ELECTRONICS
DOI: 10.37394/232017.2022.13.17
Doohwan Kim, Yo-Han Han, Jongpil Jeong
E-ISSN: 2415-1513
132
Volume 13, 2022
not perpendicular to the product, [19].
When the proposed system is run through an edge
computer, the installed camera continuously
acquires images with softwareGrab, [20], in the
acquired area (FOV, [21]) Check that the product is
detected using a Hough Circle Transform, [22].
When a product is detected, the coordinates of the
photographed product are calculated to check
whether the product passes through a discernable
location. If the product passes through a
discriminable location, the learned YOLOv4 Model
determines whether the product is defective. The
determined image and result are output from the
separately implemented monitoring system The
determination result is transmitted to the discharge
cylinder. Since the discharge cylinder only moves
the cylinder according to the electrical signal of the
program, After storing the determination result in
the edge computer’s determination program, when
the determined product passes in front of the
cylinder, the product determination result is
transmitted. As a result of the transfer, the defective
product is pushed through the cylinder to be moved
to a different conveyor belt line than the normal
product. To prevent further damage to other normal
or defective products in the process of pushing the
product, a sponge is attached to the cylinder to
alleviate the impact when the cylinder and product
collide. Fig. 3 shows the concept described above as
a figure.
Fig. 3: System Design Concept.
4 Experimental Results
4.1 Experiment Environment
Table 1 describes the list of Hardware used to
design the system. An acA1300-60gm mono camera
sold by Balser was used as the camera, and a 16mm
C-mount lens sold by Tamron was used as the
camera lens. The monitoring system and the edge
computer that performs deep learning and image
discrimination were prepared with the same
specifications as above. For optical system lighting,
two LED bar lights were used.
Table 1. Real-Time Anomaly Detection System
Hardware Spec.
LENS
Manufacturer
Part Number
Tamron M118FM16
Focal Length
16mm
Lens Mount
C-mount
Optical Format
1/1.8”
CAMERA
NAME
acA1300-60gm
Sensor Format
1/1.8”
Sensor Type
CMOS
Mono or Color
Mono
IPC
Processor
AMD RyZen 5 5600X 6-Core
Processor 3.70 GHz
RAM
32 GB
GPU
NVIDIA GeForce RTX 2080 Ti
Optical
Lighting
Product
/Model Name
LED T5/HF-T5030
Rated voltage
220V 60Hz 5W
Spec
300 x 22.5 x 33 (mm)
The test site was conducted on a conveyor belt
where actual products are produced. The
specifications and test environment of the products
used are introduced in Table 2, The learned labeling
class shown in Fig. 5 can be checked.
As a special feature, since we plan to read two
products in one image, we defined the distance
between the products and set the working distance
between the product and the camera a little further
than when shooting one product. In addition, if the
product moves too fast, it is difficult to acquire
images, so the movement speed is specified.
However, if the moving speed needs to be increased
for the production speed of the product, it can help
to acquire accurate images by adjusting the settings
of the camera and lens.
Table 2. Cosmetic Container and Spec Experimental
Environment Setting Value.
73.8mm
44.4mm
25.2mm
6∼10mm
450mm
50mm/s
As explained in Section 3, since the inspection
material of the product is a sticker, reflection occurs
when light shines directly on the product, so the
image is not accurately acquired. Therefore, by
correcting the angle of incidence of the light
illuminating the product, the product is induced to
WSEAS TRANSACTIONS on ELECTRONICS
DOI: 10.37394/232017.2022.13.17
Doohwan Kim, Yo-Han Han, Jongpil Jeong
E-ISSN: 2415-1513
133
Volume 13, 2022
be illuminated with indirect light. In addition, when
the camera shoots a product vertically, light
reflection may also occur, so adjust the angle of the
camera so that light reflection does not occur. Fig. 4
is a picture prepared to understand the above
explanation.
Fig. 4: Products with light reflection (left) and
products without reflection (right)
In the case of the product on the left, a reflection
of light occurs, so the left and bottom sides of the
sticker, which must be identified, are not visible. In
addition, reflection occurs in the center of the
sticker, making it difficult to accurately detect the
defective class. On the other hand, in the case of the
product on the right, there is almost no light
reflection, and especially the area of the sticker that
detects the defective class is accurately illuminated.
Fig. 5: Normal determined by the inspection system
The cosmetic container, which is the subject of this
study, undergoes the sticker attachment process
after defective product. the injection process. If the
position of the attached sticker is not correct or the
sticker surface is damaged and an image different
from the normal product is taken, the product must
be passed through the discharge cylinder to a
discharge part different from the normal product.
Fig. 6 is a picture of the set-up system
discriminating between a defective product and a
normal product based on the above description. The
normal class was not learned separately but was
learned in three classes: sticker peeled off (’Peeled
off’), wrinkled sticker (’Wrinkle’), and out of
position (’Escape’). The reason for not learning the
normal class is that the product’s bad class features
often include the characteristics of the normal class,
so good performance was not obtained as a result of
learning the normal class. Therefore, only images of
defective products were defined and trained without
learning the normal class. Of course, even if normal
class and defective class were detected together, it
could be classified through programming, but as a
result of the experiment, the accuracy of the system
was measured to be higher when only the defective
class was learned.
Based on the defined class, Deep Learning Train
was performed on the YOLOv4 model. 1 week from
initial data learning to final inspection took time,
and A total of 5 inspection courses were conducted
and Before the experiment, it was trained using
about 3,000. The reason for shortening the learning
period was to examine whether accurate
performance could be achieved even in a short
learning period. As the production process focused
on small-lot production of various products is highly
demanded, it may be difficult to apply the
requirements of various fields if a lot of learning
time is required. Accordingly, to achieve maximum
performance with minimal learning, the learning
image and learning time were set to be small.
Fig. 6: Discharge of the identified defective product
WSEAS TRANSACTIONS on ELECTRONICS
DOI: 10.37394/232017.2022.13.17
Doohwan Kim, Yo-Han Han, Jongpil Jeong
E-ISSN: 2415-1513
134
Volume 13, 2022
4.2 Results
In the case of early learning, due to the difference
between abnormal sample data and actual test data,
a normal product was judged as defective (Over
Detection) or a defective product was judged as
normal into the cylinder. (Miss Detection).
However, as the amount of learning of bad sample
data and normal sample data gradually increased, it
finally showed Precision 1, Recall 0.996. Table 3 is
a tabular summary of the experimental results.
Table 3. Experimental Results.
Total
Test
Count
Real
Bad
Product
Detected
Bad
Product
(FN)Over-
Detection
(FP) Miss-
Detection
Precision Recall
1st
Inspection
130 85 52 0 33 1.00 0.611
2nd
Inspection
832 640 601 10 49 0.98 0.92
3rd
Inspection
386 232 348 132 16 0.62 0.93
4th
Inspection
152 106 105 0 1 1.00 0.99
Final
Inspection
440 263 262 0 1 1.00 0.996
Since the bad class was learned in this
experiment, FP corresponds to detecting the bad
class in a good product. Conversely, FN is a case
where a defective product cannot be detected as a
defective class and is judged to be a normal product.
You should look at the table with this in mind. For
example, in the first inspection, 130 products were
prepared, and 85 defective products were included.
Here we detected 52 defective products. After
checking the products that were identified, 33
defective products were identified among the
products that were determined to be good. So there
are 33 FNs. Conversely, among the products
identified as defective products, there were 0 normal
products. So FP will be zero. Now, according to the
calculation formula, Precision = TP / (TP + FP), so
Precision = 52 / (52 + 0 ) = 1 is derived. For Recall,
Recall = TP / (TP + FN ), which leads to Recall =
52 / ( 52 + 33) = 0.61. Viewing the table in this way
will help you understand.
5 Conclusion
By using the designed real-time anomaly detection
system, it was possible to reduce the number of
people put into the inspection department at the site,
and it was possible to obtain the result of accepting
the fatigue of the existing workers. In addition, it
has great significance in that it built a model that
can be applied immediately in the field at a low
cost.
Since this system is a model created based on
Supervised Learning, collection of bad sample data
is essential, and as shown in Table 1 of 4.2, if
various data cannot be secured, good performance
cannot be achieved. Accordingly, the goal is to
review the GAN, [23] model and study an anomaly
detection model, [24] that can be satisfied in the
field even with Unsupervised Learning so that the
data is produced by artificial intelligence and used
as learning data.
This paper does not give specific suggestions
on how to utilize the collected manufacturing
data. There are infinite ways to utilize
manufacturing data, and as a representative
example, research to implement a predictive
maintenance system using manufacturing data in
the smart factory industry can be confirmed, [25].
As the number of products produced increases,
the amount of data collected is vast, so big data-
based platform research is also being prepared,
[8], and it is attracting attention as data that can
be used in various other systems. If raw data
collected in small factories can be
transmitted to Cloud, MES, or ERP servers using a
simple system, it is judged that it will be helpful for
applied research using this.
References:
[1]
Son Young-Jin and Choi Hwan Young . “A Study
on Smart Factory Introduction Cases and
Sustainable Effect”, Journal of Practical
Engineering Education (JPEE) , 2022, vol.14,
no.1, pp. 127-136
[2]
HOZDIĆ Elvis, “Smart factory for industry 4.0: A
review”, International Journal of Modern
Manufacturing Technologies, 2015, 7.1: 28-35.
[3]
PAGNON William, “The 4th industrial
revolution–A smart factory implementation
guide”, International Journal of Advanced
Robotics and Automation (IJARA), 2017, 2.2: 1-5.
WSEAS TRANSACTIONS on ELECTRONICS
DOI: 10.37394/232017.2022.13.17
Doohwan Kim, Yo-Han Han, Jongpil Jeong
E-ISSN: 2415-1513
135
Volume 13, 2022
[4]
BOCHKOVSKIY Alexey and WANG Chien-Yao
and LIAO Hong-Yuan Mark, “Yolov4: Optimal
speed and accuracy of object detection”, arXiv
preprint arXiv:2004.10934, 2020.
[5]
CHALAPATHY Raghavendra and CHAWLA
Sanjay, “Deep learning for anomaly detection: A
survey”, arXiv preprint arXiv:1901.03407, 2019.
[6]
KIM Chang-Min and CHUNG Kyungyong and
PARK Roy C, “Anomaly detection model of
mammography using YOLOv4-based histogram”,
Personal and Ubiquitous Computing, 2021, 1-12.
[7]
Tushar Jain and Meenu A, “Integration of
Industries Through Computer Vision systems.”,
International Journal of Information and
Computation Technology, 2013 , ISSN 0974-2239
Volume 3, Number 9 , pp. 963-970
[8]
SHIN Seung-Jun and WOO Jungyub and SEO
Wonchul, “Developing a big data analytics
platform architecture for smart factory”, Journal of
Korea Multimedia Society, 2016, 19.8:
1516-1529.
[4]
[5]
[6]
[7]
[8]
[9]
CUNNINGHAM and Pádraig CORD and Matthieu
DELANY and Sarah Jane m , “Supervised
learning”. In: Machine learning techniques for
multimedia , Springer, Berlin, Heidelberg, 2008. p.
21-49.
[10]
Guansong Pang and Chunhua Shen and Longbing
Cao and Anton Van Den Hengel , “Deep learning
for anomaly detection: A review”, ACM
Computing Surveys (CSUR), 2021, 54.2: 1-38.
[11]
OMAR Salima and NGADI Asri and JEBUR
Hamid H, “Machine learning techniques for
anomaly detection: an overview”, International
Journal of Computer Applications, 2013, 79.2.
[12]
Ross Girshick and Jeff Donahue and Trevor
Darrell and Jitendra Malik, “Rich feature
hierarchies for accurate object detection and
semantic segmentation”, In: Proceedings of the
IEEE conference on computer vision and pattern
recognition, 2014, p. 580-587.
[13]
Weisong Shi and Jie Cao and Quan Zhang and
Youhuizi Li and Lanyu Xu, “Edge computing:
Vision and challenges”, IEEE internet of things
journal, 2016, 3.5: 637-646.
[14]
Blesson Varghese and Nan Wang and Sakil
Barbhuiya and Peter Kilpatrick and Dimitrios S.
Nikolopoulos, “Challenges and opportunities in
edge computing”, In: 2016 IEEE International
Conference on Smart Cloud (SmartCloud), IEEE,
2016. p. 20-26.
[15]
DAVIS Jesse and GOADRICH Markm “The
relationship between Precision-Recall and ROC
curves”, In: Proceedings of the 23rd international
conference on Machine learning, 2006, p. 233-240.
[16]
Justin M. Johnson and Taghi M. Khoshgoftaar,
“Survey on deep learning with class imbalance”,
Journal of Big Data, 2019, 6.1: 1-54.
[17]
[18]
[19]
[20]
[21]
[22]
[23]
Antonia Creswell and Tom White and Vincent
Dumoulin and Kai Arulkumaran and Biswa
Sengupta and Anil A. Bharath, “Generative
adversarial networks: An overview”, IEEE signal
processing magazine, 2018, 35.1: 53-65.
[24]
Subutai Ahmad and Alexander Lavin and Scott
Purdy and Zuha Agha, “Unsupervised real-time
anomaly detection for streaming data”,
Neurocomputing, 2017, 262: 134-147.
[25]
CHUN Seung-Man and SUK Soo-Young,
“Development and implementation of smart
manufacturing big-data platform using opensource
for failure prognostics and diagnosis technology of
industrial robot”, IEMEK Journal of Embedded
Systems and Applications, 2019, 14.4: 187-195.
WSEAS TRANSACTIONS on ELECTRONICS
DOI: 10.37394/232017.2022.13.17
Doohwan Kim, Yo-Han Han, Jongpil Jeong
E-ISSN: 2415-1513
136
Volume 13, 2022
[17]
David M. W. Powers, “Evaluation: from precision,
recall and F-measure to ROC, informedness,
markedness and correlation”, arXiv preprint
arXiv:2010.16061, 2020.
[18]
Oh Sewon and Jeong Jongpil and Park Jungsoo,
“Design and implementation of smart factory
system based on manufacturing data for cosmetic
industry”, The Journal of the Institute of Internet,
Broadcasting and Communication, 2021, 21.1:
149-162.
[19]
Kim, Yu-Sin and An-Seop Choi, “Analysis of the
change of intensity distribution and luminous
environment by reflector shape in indirect
reflected LED luminaires”, Journal of the Korean
Institute of Illuminating and Electrical Installation
Engineers 25.1 ,2011, p.: 9-17.
[20]
Ryan Sun and Matthew B. Bouchard and Elizabeth
M. C. Hillman, “SPLASSH: Open source software
for camera-based high-speed, multispectral in-vivo
optical image acquisition”, Biomedical optics
express, 2010, 1.2: 385-397.
[21]
RHODY Harvey, “Lecture 10: Hough circle
transform”, Chester F. Carlson Center for Imaging
Science, Rochester Institute of Technology, 2005.
[22]
BSubutai Ahmad and Alexander Lavin and
Scott Purdy and Zuha Agha, “Unsupervised
real-time anomaly detection for streaming
data”, Neurocomputing, 2017, 262: 134-147.
Conflicts of Interest
The author(s) declare no potential conflicts of
interest concerning the research, authorship, or
publication of this article.
Contribution of individual authors to
the creation of a scientific article
(ghostwriting policy)
The author(s) contributed in the present
research, at all stages from the formulation
of the problem to the final findings
and solution.
Sources of funding for research
presented in a scientific article or
scientific article itself
This research was supported by the
MSIT(Ministry of Science and ICT), Korea,
under the ITRC (Information Technology
Research Center) support program
(IITP-2022-2018-0-01417) supervised by the
IITP (Institute for Information &
Communications Technology Planning &
Evaluation), and the National Research
Foundation of Korea (NRF) grant funded by the
Korea government (MSIT) (No.
2021R1F1A1060054).
Creative Commons Attribution
License 4.0 (Attribution 4.0
International , CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US