INDUSTRY 4.0’s smart industry uses information

and communication technology to improve produc-

tivity [1]. One of them is the Internet of Things (IoT)

system, which allows monitoring and control in various

areas due to advances in hardware and low-power com-

munication technology [2]. Industrial IoT (IIoT) is a sig-

niﬁcant study [3, 4]. IIoT uses a variety of sensors to

monitor the performance of the production process [5].

High-performance cloud computing (CC) is attracting

attention to processing and analyzing much of IIoT’s

data. CC’s ﬂexibility and scalability of its resources have

made its technology a huge success. However, CC’s speed

and latency do not meet service quality requirements [6],

but edge computing (EC) is close to IIoT and can meet

the requirements for speed and latency [7]. EC provides

computing, storage, and network resources in close prox-

imity to IoT devices. The distance between Edge and IoT

is close, so the delay time is reduced and the transmission

time is shorter than that of CC [8].

But we can’t handle everything with EC in the im-

mediate area. It is true that EC is inferior to CC in per-

formance to CC. And there may be latency and latency

because you need to connect to the server to save it in

DB. Distributed processing is also diﬃcult in terms of

management due to increased management points. Con-

versely, from a centralized point of view, there are few

points of management that make it easier to manage.

Therefore, you need to deﬁne an advantageous role in

Edge to maximize the beneﬁts of distributed process-

ing and ensure smooth production at the manufacturing

site. In addition, a study was conducted that as net-

work technology develops through distributed process-

ing through the edge of the network, the use of the edge

has advantages due to pretreatment of the network [9].

Through this study, it was designed to reduce the load

of the server by pre-processing manufacturing data and

parsing and pre-processing IoT data and manufacturing

data by adopting the distributed processing method of

the edge. There was a study of deep learning (DL) and

machine learning (ML) at the edge [10]. There are various

characteristics of data in manufacturing. ML is a simple

analysis of one layer, so it is suﬃciently possible in EC,

but DL is an analysis of several layers, so it is necessary

to distinguish the areas that can be done. So, we ana-

lyzed the small-scale and pre-trained DL and designed it

to be able to respond to the facility immediately.

In our study, IoT Edge was designed using the devel-

opment of EC technology in order to process the large

number of data at the manufacturing site in real time.

IoT Edge was used to allow distributed processing, away

from the existing centralized processing method on the

server. There are three types of distributed treatment.

First, to prevent the number of network connections from

increasing by connecting directly to the facilities from the

server, we decided to make a hub-type connection using

IoT Edge. Second, data from the facility was distributed

across each node of the IoT Edge to reduce the load on

the server. Third, Anomaly Detection was designed using

Real-Time Control AE-TadGAN Model in IoT Edges for Smart

Manufacturing

SANGHOON DO, JONGPIL JEONG

Department of Smart Factory Convergence, Sungkyunkwan University

2066 Seobu-ro, Jangan-gu, Suwon 16419

REPUBLIC OF KOREA

Abstract- With the development of the Internet of Things (IoT), real-time processing of data has

become an important key as various and many data have been generated at the manufacturing site. The

development of IoT has brought Cloud Computing (CC) to attention. However, it has the problem of

latency and delay, and traditional centralized data processing can violate real-time processing, drawing

attention to distributed processing technology. Edge Compression (EC) technology is a technology

that distributes a variety of data at the manufacturing site and enables real-time processing. Distribute

the various processes of traditional servers and use a near-field network to compensate for latency and

latency problems. In this study, we propose an architecture that allows EC to perform the pre-

processing, small-scale analysis, and connection for facility control, which are the processes performed

on the server with EC development.

Keywords- IoT, Edge, Smart Manufacturing, MES

Received: April 18, 2021. Revised: June 19, 2022. Accepted: July 21, 2022. Published: September 13, 2022.

1. Introduction

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.13

Sanghoon Do, Jongpil Jeong

E-ISSN: 2415-1521

Volume 10, 2022

Autoencoder (AE) and Generative Adversarial Networks

(GAN). The IoT Edge processes the pre-trained model

on the server so that it can respond to the facility in real

time. These three objectives are designed to ensure real-

time performance, which is the most important factor in

the manufacturing ﬁeld. The ﬁnal experimental results

of Figure 1 show that the average time of the analysis

model using AE-TadGAN is 33.1% faster to transmit on

IoT Edge than on Cloud.

Fig. 1: Comparison of Processing Time vetween IoT Edge

(Left) and Cloud (Right)

This study consists of: Section II. describes the rel-

evant operation. Section III. describes an architecture

using IoT Edge. Section IV. conﬁrms the results of the

analysis on IoT Edge through experiments. Finally, Sec-

tion V. discusses conclusions and future research plans.

The concept of IoT ﬁrst appeared in ”The Road

Ahead,” published in 1995. [11, 12] However, it was not

noticed due to problems with communication, hardware,

and sensor technology. With the recent rapid develop-

ment of RFID technology, the technology of IoT is draw-

ing attention. This is because IoT is a promising tech-

nology because it can build an industrial system by in-

creasing utilization through the development of commu-

nication, hardware, and sensors [13]. IoT technology is

a technology that allows real-time operation by connect-

ing physical and virtual objects, and it is an applica-

tion that can be used in various ﬁelds other than man-

ufacturing [14, 15]. EC is being studied a lot to han-

dle computational intensive tasks of IIoT. In EC sce-

narios, dynamic computational oﬄoading techniques are

proposed for use with IIoT as a technique that mini-

mizes energy consumption [16]. Supports detachable de-

lay and accuracy-aware video analysis in the cloud edge

IoT framework [17], and EC shares pipelines to reduce

unnecessary resource costs between edges.

The anomaly detection method has received a lot of

attention over the past few years. Typically, you can

classify them into statistical methods, neighborhood-

based methods, and dimension-reduction-based meth-

ods. AE is an unsupervised critical feedforward artiﬁ-

cial neural network architecture (NN) and is a data-

dimensional reduction technology consisting of encoders

and decoders[18, 19] With the improvement of artiﬁcial

intelligence, we overcome the shortcomings of intelligent

diagnostic methods such as NN and Support Vector Ma-

chine (SVM), which require manual design, many pre-

analysis, and comparison processes[20]. AE provides an

eﬀective way to learn representative features. Sakurada

proposed AE-based anomaly detection method [21]. AE

is highly detectable because it can capture nonlinear

correlation as well as linear correlation. However, the

use of AE for image anomaly detection does not always

show good results. This is because a single AE does not

fully capture the correlation between features in a high-

dimensional dataset, resulting in poor detection accu-

racy. That’s how an ensemble-based AE was born [22].

Generative Adversarial Networks can successfully

perform many image-related tasks, including image gen-

eration [23], image translation [24], video prediction [25],

and researchers have demonstrated the eﬀectiveness of

GAN for anomaly detection in recent images [26, 27].

Previous GAN-related operations contained little time

series data. This is because complex time correlations

within this type of data pose signiﬁcant challenges to

generative modeling. Three works released in 2019 are

drawing attention. First, Li et al. to use GAN for

anomaly detection in time series. [28] suggests using the

vanilla GAN model to capture the distribution of multi-

variate time series and to use Critical to detect anoma-

lies. Another approach to this line is BeatGAN [29], an

encoder and decoder GAN architecture that can use re-

construction errors to detect abnormalities in heartbeat

signals. More recently, Yun et al. [30]We propose a time

series GAN that adopts the same idea but introduces

time embeddings to support network training. However,

their work is designed to learn time series representa-

tion instead of anomaly detection.We present TadGAN,

a new framework that allows time series reconstruction

and eﬀective anomaly detection, to show how GAN can

be eﬀectively used for anomaly detection in time series

data [31].

In our study, IoT Edge using Edge was designed for

real-time data processing in manufacturing sites where a

lot of data is generated.

Figure 2 is the architecture we studied. The focus

of this architecture is the introduction of IoT Edge on

shopﬂoor to reduce the load on servers in centralized

MES. It is designed to allow small-scale AI analysis with

analysis models such as AE-TadGAN and distributed

processing in IoT Edge with enhanced hardware perfor-

mance. IoT Edges and Assets, except MES server, have

Assets on the shop ﬂoor. In addition, the IoT Edge we

designed will exist as much as we can process Asset’s

data in real time on a Node-by-Node basis. Therefore,

as the number of Asset increases or the amount of data

increases, the number of nodes in IoT Edges increases.

2. Related Work

2.1 Smart Manufacturing

2.2 AE

2.3 GAN

3. Real-Time Control AE-Tad

GAN Model

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.13

Sanghoon Do, Jongpil Jeong

E-ISSN: 2415-1521

100

Volume 10, 2022

Fig. 2: Small Scale Analysis Model for IoT Edge Using

AE-TadGAN

The asset creates data for the shop ﬂoor’s digitized

facilities. Data from Asset will be sent to IoT Edge for

collection. The collected data will be preprocessed. The

reason for preprocessing is that when the original data

is put into the server, the data parsing operation is cen-

tralized on the server. In order to reduce the load on

the server, it is handled by nodes of each connected IoT

Edge as a concept of distributed processing in IoT Edge.

In addition, data on facilities that require data analysis

can be preprocessed to enable the analysis process to run

immediately.

In data analysis, ML or pre-training small-scale AI

analysis is possible. The reason why this analysis is pos-

sible in real time is that the development of hardware

has improved the performance of PCs. In this study, we

predict defects with the pre-trained models of AE and

TadGAN in MES Server. If raw data measured by a vi-

bration sensor is used as it is, it is diﬃcult to analyze

in the form of frequency, and there is a high possibil-

ity of false alarm. Therefore, it is important to improve

the quality of input data through frequency conversion of

CPU RAM GPU

IoT Edge

AMD Ryzen 5 3600

6-Core Processor

3.60GHz

16.0 GB NVIDIA GeForce

GTX 1660 SUPER

Cloud Intel Xeon E5-2686 v4

vCPU 8 2.30GHz 61.0 GB NVIDIA Tesla V100

Table 1: Test Environment

raw data. Among the various frequency conversion meth-

ods, DWT (discrete wavelet transform) with high time

resolution in high frequency areas and high frequency

resolution in low frequency areas was used. The above

features are important because the fast-changing high

frequency has a more important time resolution to de-

termine the position of the point of change, and the fre-

quency resolution to determine the period of change at

the slow-changing low frequency. In the case of AGM

Framework, among the various types of detection algo-

rithms, models based on AE (Autoencoder) and Gen-

erative Adversarial Networks (GANs) were used. In our

work, we use the Timeseries Anomaly Detection GAN

(TadGAN), a GAN model optimized for time-series data

anomaly detection. TadGAN performs better than other

anomaly detection models and is recognized in various

ﬁelds.

Data that has been analyzed is controlled by send-

ing a signal to the asset if a defect is predicted. It is

possible to simply do a line stop, and it is possible

to perform feedback control to change the recipe by

checking whether there is a problem with the recipe.

It is an architecture that can reduce the load of MES

Server and increase real-time performance through dis-

tributed processing through IoT Edgs. In addition, the

analyzed results are transmitted to the MES server to

store the results and processing status on the server.

Control through small-scale analysis, facility conﬁgura-

tion changes due to operator, and IoT Edge conﬁgura-

tion changes are transmitted from MES server to IoT

Edge and processed. Asset control is enabled only on

IoT Edge and does not communicate directly with MES

server. The connection between MES Server and Asset

reduces the number of connections by connecting to the

IoT Edge, a recruitment group of Asset, because the con-

nection pool increases.

We experimented with an architecture designed in

Figure 2. The experimental environment is the same as

Table 1. In order to pre-training analysis models using

AE-TadGAN and check the real-time speed diﬀerence

between IoT Edge and server, AWS Cloud was used as

the role of the server, and PC with improved hardware

performance was used. The data transfer from the asset

to the IoT Edge used commercialized OPC-UA.

The main purpose of this experiment is to measure

and compare the time when sensor data is transmitted in

real time and Anomaly Detection is performed through

analysis on IoT Edge and servers, respectively, to the

workplace.

4. Experiment and Results

4.1 Experiment Environments

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.13

Sanghoon Do, Jongpil Jeong

E-ISSN: 2415-1521

101

Volume 10, 2022

For anomaly detection, the ”PHM IEEE 2012 Data

Challenge” dataset, which is a dataset for vibration val-

ues of rotating bearings, was used. The dataset consists

of 7172877 vibration values collected from rotating bear-

ings in an environment of 1800 rpm and 4000 N. The

sampling frequency is 25.6 kHz, and data is recorded ev-

ery 1/10 second the bearing rotates.

The results of pretreatment using DWT showing ex-

cellent performance at high and low frequencies are

shown in Figure 3.

Fig. 3: Bearing Vibration Data after DWT Conversion

AE-based anomaly detection uses reconstruction er-

rors that occur in the process of compressing and restor-

ing data to perform anomaly detection. Anomaly score is

a score that indicates how close each input data point is

to an outlier, and reconstruction error is used as anomaly

score. After the process of selecting the optimal threshold

based on the derived anomaly score, points with scores

higher than the threshold are judged as abnormal values.

The threshold is selected by considering the reference val-

ues of the fault frequency of ISO-10816, abnormalities

at the site, and failure experience values. This experi-

ment used USAD (Unsupervised Anomaly Detection on

Multivariable Time Series), a model that performs an

anonymous detection task with unsupervised setting in

multivariate time series. The anomaly score measured in

the process of compressing and restoring data converted

through DWT using USAD is the same as the top of

Figure 4. After setting the threshold based on the mea-

sured anomaly score, the section in which the red line

rises at the bottom of Figure 4 is determined to be an

abnormal value as a result of anomaly detection based

on this value.

Fig. 4: Anomaly Detection Using USAD

Anomaly detection using GANs is not much diﬀerent

from AE-based anomaly detection, but there is a diﬀer-

ence in obtaining an anomaly score by using the output

of the dispenser together with the restore error of the

generator. The result value of the TadGAN model for

the data converted through DWT is the same as the blue

color of Figure 5. The result of anomaly detection based

on the result value of the TadaGAN model is judged to

be an outlier from the point that occurs in the red normal

value guideline in Figure 5.

Fig. 5: Anomaly Detection Using TadGAN

To compare processing time on IoT Edge with pro-

cessing time on server, DWT Transformation and AE

and GAN-based anomaly detection were measured on

server and processing time on IoT Edge, respectively.

The expression of each processing time in time series is

as shown in Figure 6.

Fig. 6: Required Processing Time to Process Real-time

Data in IoT Edge (Below) and Cloud (Above)

At this time, the average processing time required for

DWT Transformation and AE and GAN-based anomaly

detection was similar in the two environments of IoT

Edge and server. This is because the analysis of the

small-scale data used in the experiment does not require

much computing power, so there is no diﬀerence in the

processing speed between the PC-class IoT Edge and the

server. However, in the case of servers, it takes addi-

tional transmission time to upload data from IoT Edge to

the server and download analysis results, which increases

processing time compared to IoT Edge Computing. From

the results of this experiment, it is more eﬃcient in terms

of real-time performance to process analysis tasks for

small-scale data on IoT Edge. This takes additional time

to upload and download data to the cloud, so analysis of

small-scale data is recommended for IoT Edge. In addi-

tion, when data resources (the physical number of assets

and objects) increase, the use of IoT Edge is more im-

portant for distributed processing to increase real-time,

as network resources and server throughput increase due

to upload/download to the server.

In this paper, we conducted an experiment that dis-

tributed processing using IoT Edge can improve real-

time performance rather than centralized processing that

adds load to servers due to the development of EC. IoT

4.2 Data Preprocessing

4.2 Results

5. Conclusions

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.13

Sanghoon Do, Jongpil Jeong

E-ISSN: 2415-1521

102

Volume 10, 2022

Edge designed a simple transmission program that was

only sent to the server for distributed processing, and

IoT Edge handled the pre-processing process that was

loaded on the server to reduce the load on the server,

and the small AI analysis model was tested on IoT Edge.

Therefore, the server load is increasing day by day due to

big data analysis, so we distributed the analysis on the

server to IoT Edge in order to ﬁll resources or reduce

the impact on other functions of the server. This reduces

the load and has less impact on real-time processing, and

it can be determined and controlled immediately by IoT

Edge near the shop ﬂoor to ensure real-time performance

as much as the data transmission time. In real life, un-

like experiments, there are many assets, so centralized

network traﬃc and centralized delays in processing will

be a problem as more IoT data is collected in the future.

The solution is to design distributed processing using IoT

Edge.

In this study, small-scale AI analysis was selected as

AE-TadGAN, but it was not possible to measure how

large an analysis would be possible from Edge comput-

ing. More experiments are needed and research is needed

on how to calculate AI analysis that can be analyzed on

IoT Edge. You also need to study how much you can

connect to Asset. We will focus on research on the avail-

ability of IoT Edge. And to reduce the load on MES, we

also need to conduct research on distributed processing

on servers.

This research was supported by the SungKyunKwan

University and the BK21 FOUR (Graduate School Inno-

vation) funded by the Ministry of Education (MOE, Ko-

rea) and National Research Foundation of Korea (NRF).

[1] G. Aceto, V. Persico and A. Pescap´e, ”A Survey on

Information and Communication Technologies for In-

dustry 4.0: State-of-the-Art, Taxonomies, Perspec-

tives, and Challenges,” in IEEE Communications

Surveys Tutorials, vol. 21, no. 4, pp. 3467-3501,

Fourthquarter 2019, Available: https://doi.org/

10.1109/COMST.2019.2938259

[2] L. D. Xu, W. He and S. Li, ”Internet of Things in

Industries: A Survey,” in IEEE Transactions on In-

dustrial Informatics, vol. 10, no. 4, pp. 2233-2243,

Nov. 2014, Available: https://doi.org/10.1109/

TII.2014.2300753

[3] Oztemel, E., Gursev, S. Literature review of Indus-

try 4.0 and related technologies. J Intell Manuf 31,

127–182 (2020). https://doi.org/10.1007/s10845-018-

1433-8

[4] W. Z. Khan, M. H. Rehman, H. M. Zangoti, M.

K. Afzal, N. Armi and K. Salah, ”Industrial Inter-

net of Things: Recent advances enabling technologies

and open challenges”, Comput. Electr. Eng., vol. 81,

Jan. 2020. Available: https://doi.org/10.1016/j.

compeleceng.2019.106522

[5] J. Pan and J. McElhannon, ”Future Edge Cloud

and Edge Computing for Internet of Things Appli-

cations,” in IEEE Internet of Things Journal, vol.

5, no. 1, pp. 439-449, Feb. 2018, Available: https:

//doi.org/10.1109/JIOT.2017.2767608

[6] Duc, Thang Le, et al. ”Machine learning methods for

reliable resource provisioning in edge-cloud comput-

ing: A survey.” ACM Computing Surveys (CSUR)

52.5 (2019): 1-39. https://doi.org/10.1145/3341145

[7] T. Taleb, K. Samdanis, B. Mada, H. Flinck, S. Dutta

and D. Sabella, ”On Multi-Access Edge Computing:

A Survey of the Emerging 5G Network Edge Cloud

Architecture and Orchestration,” in IEEE Commu-

nications Surveys Tutorials, vol. 19, no. 3, pp. 1657-

1681, thirdquarter 2017, Available: https://doi.

org/10.1109/COMST.2017.2705720

[8] L. Chen et al., ”IoT Microservice Deployment in

Edge-Cloud Hybrid Environment Using Reinforce-

ment Learning,” in IEEE Internet of Things Jour-

nal, vol. 8, no. 16, pp. 12610-12622, 15 Aug.15,

2021, Available: https://doi.org/10.1109/JIOT.

2020.3014970

[9] I. Satoh, ”5G-enabled Edge Computing for

MapReduce-based Data Pre-processing,” 2020

Fifth International Conference on Fog and Mobile

Edge Computing (FMEC), 2020, pp. 210-217, Avail-

able: https://doi.org/10.1109/FMEC49853.2020.

9144882

[10] P. Subedi, J. Hao, I. K. Kim and L. Ramaswamy,

”AI Multi-Tenancy on Edge: Concurrent Deep Learn-

ing Model Executions and Dynamic Model Place-

ments on Edge Devices,” 2021 IEEE 14th Interna-

tional Conference on Cloud Computing (CLOUD),

2021, pp. 31-42, Available: https://doi.org/10.

1109/CLOUD53861.2021.00016

[11] Bandyopadhyay, D., Sen, J. Internet of

Things: Applications and Challenges in Tech-

nology and Standardization. Wireless Pers

Commun 58, 49–69 (2011). Available: https:

//doi.org/10.1007/s11277-011-0288-5

[12] S. Li, L. Xu and X. Wang, ”Compressed sensing

signal and data acquisition in wireless sensor net-

works and Internet of Things”, IEEE Trans. Ind. In-

format., vol. 9, no. 4, pp. 2177-2186, Nov. 2013. Avail-

able: https://doi.org/10.1016/j.comnet.2010.

05.010

[13] F. Tao, Y. Zuo, L. D. Xu and L. Zhang, ”IoT-

Based Intelligent Perception and Access of Manufac-

turing Resource Toward Cloud Manufacturing,” in

IEEE Transactions on Industrial Informatics, vol. 10,

no. 2, pp. 1547-1557, May 2014, Available: https:

//doi.org/doi:10.1109/TII.2014.2306397

[14] L. D. Xu, W. He and S. Li, ”Internet of Things in

Industries: A Survey,” in IEEE Transactions on In-

dustrial Informatics, vol. 10, no. 4, pp. 2233-2243,

Nov. 2014, Available: https://doi.org/10.1109/

TII.2014.2300753

[15] S´anchez L´opez, T., Ranasinghe, D.C., Har-

rison, M. et al. Adding sense to the Internet

Acknowledgment

References

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.13

Sanghoon Do, Jongpil Jeong

E-ISSN: 2415-1521

103

Volume 10, 2022

of Things. Pers Ubiquit Comput 16, 291–308

(2012). Available: https://doi.org/10.1007/

s00779-011-0399-8

[16] S. Chen, Y. Zheng, W. Lu, V. Varadarajan and

K. Wang, ”Energy-Optimal Dynamic Computation

Oﬄoading for Industrial IoT in Fog Computing,”

in IEEE Transactions on Green Communications

and Networking, vol. 4, no. 2, pp. 566-576, June

2020, Available: https://doi.org/10.1109/TGCN.

2019.2960767

[17] Y. Zhang, J. -H. Liu, C. -Y. Wang and H. -Y.

Wei, ”Decomposable Intelligence on Cloud-Edge IoT

Framework for Live Video Analytics,” in IEEE Inter-

net of Things Journal, vol. 7, no. 9, pp. 8860-8873,

Sept. 2020, Available: https://doi.org/10.1109/

JIOT.2020.2997091

[18] S. Takaki and J. Yamagishi, ”A deep auto-encoder

based low-dimensional feature extraction from FFT

spectral envelopes for statistical parametric speech

synthesis,” 2016 IEEE International Conference on

Acoustics, Speech and Signal Processing (ICASSP),

2016, pp. 5535-5539, Available: https://doi.org/

10.1109/ICASSP.2016.7472736.

[19] Sarroﬀ, Andy, and Michael A. Casey. ”Musical au-

dio synthesis using autoencoding neural nets.” Pro-

ceedings of the International Society for Music Infor-

mation Retrieval Conference (ISMIR2014). Interna-

tional Society for Music Information Retrieval, 2014.

[20] Y. Zhang et al., ”Intelligent fault diagnosis

of rotating machinery using a new ensemble

deep auto-encoder method”, Measurement, pp.

151, 2020. Available: https://doi.org/10.1016/j.

measurement.2019.107232

[21] M. Sakurada and T. Yairi, ”Anomaly detection us-

ing autoencoders with non-linear dimensionality re-

duction”, Proceedings of the MLSDA 2014 2ndWork-

shop on Machine Learning for Sensory Data Anal-

ysis, 2014. Available: https://doi.org/10.1145/

2689746.2689747

[22] Z. Chen et al., ”Evolutionary multi-objective opti-

mization based ensemble autoencoders for image out-

lier detection”, Neurocomputing, vol. 309, pp. 192-

200, 2018. Available: https://doi.org/10.1016/j.

neucom.2018.05.012

[23] Goodfellow, Ian, Jean Pouget-Abadie, Mehdi

Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,

Aaron Courville, and Yoshua Bengio. ”Generative

adversarial nets.” Advances in neural information

processing systems 27 (2014). Available: https:

//proceedings.neurips.cc/paper/2014/hash/

5ca3e9b122f61f8f06494c97b1afccf3-Abstract.

html

[24] J. Zhu, T. Park, P. Isola and A. A. Efros, ”Unpaired

Image-to-Image Translation Using Cycle-Consistent

Adversarial Networks,” 2017 IEEE International

Conference on Computer Vision (ICCV), 2017, pp.

2242-2251, Available: https://doi.org/10.1109/

ICCV.2017.244

[25] Vondrick, Carl, Hamed Pirsiavash, and An-

tonio Torralba. ”Generating videos with scene

dynamics.” Advances in neural information pro-

cessing systems 29 (2016). Available: https:

//proceedings.neurips.cc/paper/2016/hash/

04025959b191f8f9de3f924f0940515f-Abstract.

html

[26] T. Schlegl, P. Seebck, S. M. Waldstein, G. Langs

and U. Schmidt-Erfurth, ”f-AnoGAN: Fast unsuper-

vised anomaly detection with generative adversarial

networks”, Medical Image Analysis, vol. 54, pp. 30-

44, 2019. Available: https://doi.org/10.1016/j.

media.2019.01.010

[27] L. Deecke, R. Vandermeulen, L. Ruﬀ, S. Mandt and

M. Kloft, ”Anomaly detection with generative ad-

versarial networks”, 2018, [online] Available: https:

//openreview.net/forum?id=S1EfylZ0Z.

[28] Li, D., Chen, D., Jin, B., Shi, L., Goh, J., Ng, SK.

(2019). MAD-GAN: Multivariate Anomaly Detection

for Time Series Data with Generative Adversarial

Networks. In: Tetko, I., K˚urkov´a, V., Karpov, P.,

Theis, F. (eds) Artiﬁcial Neural Networks and Ma-

chine Learning – ICANN 2019: Text and Time Se-

ries. ICANN 2019. Lecture Notes in Computer Sci-

ence(), vol 11730. Springer, Cham. Available: https:

//doi.org/10.1007/978-3-030-30490-4_56

[29] B. Zhou, S. Liu, B. Hooi, X. Cheng and J. Ye, ”Beat-

GAN: Anomalous Rhythm Detection using Adver-

sarially Generated Time Series”, Proc. of the 28th

Int. Joint Conf. on Artiﬁcial Intelligence (IJCAI),

pp. 4433-4439, 2019. Available: https://doi.org/

10.24963/ijcai.2019/616

[30] Yoon, Jinsung, Daniel Jarrett, and Mihaela Van

der Schaar. ”Time-series generative adversar-

ial networks.” Advances in neural information

processing systems 32 (2019). Available: https:

//proceedings.neurips.cc/paper/2019/hash/

c9efe5f26cd17ba6216bbe2a7d26d490-Abstract.

html

[31] A. Geiger, D. Liu, S. Alnegheimish, A. Cuesta-

Infante and K. Veeramachaneni, ”TadGAN:

Time Series Anomaly Detection Using Gen-

erative Adversarial Networks,” 2020 IEEE

International Conference on Big Data (Big

Data), 2020, pp. 33-43, Available: https:

//doi.org/10.1109/BigData50022.2020.9378139

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

WSEAS TRANSACTIONS on COMPUTER RESEARCH

DOI: 10.37394/232018.2022.10.13

Sanghoon Do, Jongpil Jeong

E-ISSN: 2415-1521

104

Volume 10, 2022