Mobile QR Code QR CODE : The Transactions P of the Korean Institute of Electrical Engineers

  1. (Korea Electric Power Research Institute, Korea.)
  2. (Electronics and Telecommunications Research Institute, Korea.)



Autoencoder, Stacked Autoencoder, Self-attention, Anomaly-Detection, Underground Cable Tunnel, Adversarial-Autoencoder

1. ์„œ ๋ก 

์ „๋ ฅ์†Œ๋น„ ๋ฐ ๋„์‹œ ๋ฏธ๊ด€์— ๋Œ€ํ•œ ๊ด€์‹ฌ ์ฆ๊ฐ€๋กœ ์ง€์ค‘ํ™” ์ „๋ ฅ์„ค๋น„๋กœ์˜ ์ „ํ™˜ ์š”๊ตฌ๊ฐ€ ๋†’์•„์ง€๊ณ  ์žˆ์œผ๋ฉฐ, 2019๋…„ ์ „๊ตญ ๋ฐฐ์ „์„ค๋น„ ์ง€์ค‘ํ™”์œจ์€ 10.8%๋กœ ์„œ์šธ์€ 60%์ˆ˜์ค€์— ๋‹ฌํ•˜๊ณ , ์ง€์ค‘๋ฐฐ์ „์„ ๋กœ ๊ธ์žฅ์€ 54,618 circuit-km๋กœ ๋งค๋…„ ์ฆ๊ฐ€ํ•˜๊ณ  ์žˆ๋‹ค. ์ง€ํ•˜์ „๋ ฅ๊ตฌ๋Š” ๋‹ค๋Ÿ‰์˜ ์ผ€์ด๋ธ” ๋ฐ ๋ถ€์†์ž์žฌ, ์ผ€์ด๋ธ”์˜ ์ ‘์†๊ณต๊ฐ„์„ ํฌํ•จํ•˜๋Š” ๊ธฐ๋Šฅ์„ ๊ฐ€์ง„ ๊ตฌ์กฐ๋ฌผ์ด๋‹ค. ์ง€ํ•˜์ „๋ ฅ๊ตฌ์˜ ์‚ฌ๊ณ  ์œ ํ˜•์€ ํ™”์žฌ๊ฐ€ ๋งŽ์ด ๋ฐœ์ƒํ•˜๋ฉฐ, ์ด๋Š” ์‹œ์„ค๋ฌผ์— ๊ณต๊ธ‰๋˜๋Š” ์ „๊ธฐ์˜ ์ ‘์†๋ถ€ ๋ฐ ์ ˆ์—ฐ๋ถ€์˜ ๊ณผ์—ด๋กœ ์ธํ•œ ์‚ฌ๊ณ ์ด๋‹ค. ๊ณ ์žฅ์œจ ์—ญ์‹œ ๋งค๋…„ ์ฆ๊ฐ€ํ•˜๋Š” ์ถ”์„ธ๋กœ ์ง€ํ•˜์ „๋ ฅ๊ตฌ์— ๋Œ€ํ•œ ์›๊ฒฉ๊ฐ์‹œ๊ฐ€ ํ•„์ˆ˜์ ์ด๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ์ง€ํ•˜์ „๋ ฅ๊ตฌ ๋‚ด๋ถ€์— ์˜จ๋„, ํŽŒํ”„๋™์ž‘ ์„ผ์„œ๋ฅผ ์ฃผ์š” ์ง€์ ์— ์„ค์น˜ํ•œ ํ›„ ์ฃผ๋กœ ์ˆœ์‹œ์ ๊ฒ€์„ ํ†ตํ•ด ์ „๋ ฅ๊ตฌ ์ ๊ฒ€์„ ์‹œํ–‰ํ•˜๊ณ  ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์„ผ์‹ฑ๋œ ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜์˜ ๋ถ„์„ ๊ธฐ๋Šฅ์ด ์ ์šฉ๋˜์ง€ ์•Š๊ณ  ์ผ์ •ํ•œ ๊ฐ’ ๊ธฐ๋ฐ˜์˜ ๊ฒฝ๊ณ ๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์žˆ์–ด ์ข…ํ•ฉํŒ๋‹จ์„ ๋‚ด๋ฆฌ๊ธฐ์—” ๋ถ€์กฑํ•˜๋‹ค. ๋”ฐ๋ผ์„œ, ์•ˆ์ •์ ์ธ ์ „๋ ฅ์„œ๋น„์Šค๋ฅผ ์ œ๊ณตํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ์—ด์•…ํ•œ ํ™˜๊ฒฝ์˜ ์„ค๋น„ ์ƒํƒœ๋ฅผ ๋ชจ๋‹ˆํ„ฐ๋งํ•˜์—ฌ ์‹ค์‹œ๊ฐ„์œผ๋กœ ์ง„๋‹จํ•˜๊ณ , ์‚ฌ์ „์— ์„ค๋น„ ์žฅ์• ๋ฅผ ์˜ˆ์ง€ํ•˜์—ฌ ์‚ฌ์ „์— ์กฐ์น˜ํ•˜๋Š” ๊ฒƒ์ด ๋ฌด์—‡๋ณด๋‹ค ์ค‘์š”ํ•˜๋‹ค.

์ตœ๊ทผ ์‚ฐ์—…ํ˜„์žฅ์—์„œ๋Š” ์„ค๋น„์— ๋ถ€์ฐฉํ•œ ์„ผ์„œ ์ •๋ณด ๊ธฐ๋ฐ˜์œผ๋กœ ๊ธฐ๊ณ„ํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ ์šฉํ•˜์—ฌ ์„ค๋น„ ์ƒํƒœ๋ฅผ ์ง„๋‹จํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ์˜ ํŒจ๋Ÿฌ๋‹ค์ž„ ์ „ํ™˜์ด ์ด๋ฃจ์–ด์ง€๊ณ  ์žˆ๋‹ค. ์ด์ƒ์ง„๋‹จ์€ ๊ณตํ•™์‹œ์Šคํ…œ์˜ ์‹ค์‚ฌ์šฉ ํ™˜๊ฒฝ์—์„œ ๊ฐ์ข… ์„ผ์„œ๋ฅผ ํ†ตํ•ด ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ ์„ค๋น„์˜ ์ด์ƒ์ƒํƒœ๋ฅผ ํŒ๋‹จํ•˜๊ณ , ์„ค๋น„์˜ ์ˆ˜๋ช…์„ ์˜ˆ์ง€ํ•˜์—ฌ ์‚ฌ์ „ ์ •๋น„๋ฅผ ์‹ค์‹œํ•˜๋Š” ๊ธฐ์ˆ ๋กœ(1) ์•ˆ์ „๋ฌธ์ œ, ๋ถˆ๋Ÿ‰ ๊ฒ€์‚ฌ ๋“ฑ ๋ชจ๋“  ์‚ฐ์—…์— ๊ฑธ์ณ ์ ์šฉ๋˜๊ณ  ์žˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•๋“ค์ด ์‚ฌ์šฉ๋˜์–ด ์™”์ง€๋งŒ, ์ตœ๊ทผ์—๋Š” ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ™œ์šฉํ•˜์—ฌ ์—ฌ๋Ÿฌ ๋น„์„ ํ˜• ๋ณ€ํ™˜์„ ํ†ตํ•ด ์›์‹œ๋ฐ์ดํ„ฐ์—์„œ ๋Œ€ํ‘œ ํŠน์ง•์„ ์ฐพ์•„๋‚ด๋Š” ๋ฐฉ๋ฒ•์œผ๋กœ ์ด์ƒ์ง„๋‹จ์— ์‚ฌ์šฉํ•˜๊ณ  ์žˆ๋‹ค(2). Vincent ๋“ฑ์€ ํ•ญ๊ณต๊ธฐ ์—”์ง„ ๊ณ ์žฅ์ง„๋‹จ์„ ์œ„ํ•ด ํŒจํ„ด์ถ”์ถœ๊ธฐ๋กœ ์ ์ธต์žก์Œ์ œ๊ฑฐ ์˜คํ† ์ธ์ฝ”๋”๋ฅผ ์‚ฌ์šฉํ•˜๊ณ , ์†Œํ”„ํŠธ๋งฅ์Šค ๋ถ„๋ฅ˜๊ธฐ๋ฅผ ์ด์šฉํ•˜์—ฌ ๊ฒฐํ•จ์„ ๋ถ„๋ฅ˜ํ•˜๋Š” ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค(3). ๋˜ํ•œ ์ „๋™๊ธฐ์˜ ๊ณ ์žฅ์ง„๋‹จ์„ ์œ„ํ•ด ์ž…๋ ฅ๋ฐ์ดํ„ฐ๋ฅผ FFT ๋ณ€ํ™˜์„ ํ†ตํ•ด CNN(Convolution Neural Network) ๊ตฌ์กฐ๋กœ ๋ชจ๋ธ๋งํ•˜์—ฌ ์‹ค์‹œ๊ฐ„ ๋ถ„์„ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์—ฐ๊ตฌ ์‚ฌ๋ก€๋„ ์žˆ๋‹ค(4).

Autoencoder(AE) ๋ชจ๋ธ์€ ๋Œ€ํ‘œ์ ์ธ ์ด์ƒ ์ง„๋‹จ ๋„คํŠธ์›Œํฌ ๋ชจ๋ธ๋กœ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์ €์ฐจ์›์˜ ๋ถ€ํ˜ธํ™” ๋ฐ์ดํ„ฐ๋กœ ๋ฐ”๊พธ๋Š” ์ธ์ฝ”๋” ๊ณ„์ธต๊ณผ ์ด๋ฅผ ๋‹ค์‹œ ์ž…๋ ฅ๋ฐ์ดํ„ฐ ์ฐจ์›์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ๋””์ฝ”๋” ๊ณ„์ธต์œผ๋กœ ์ด๋ฃจ์–ด์ ธ์žˆ๋‹ค. ์ •์ƒ์ƒํƒœ์˜ ํ•™์Šต๋ฐ์ดํ„ฐ์™€๋Š” ๋‹ค๋ฅธ ์ด์ƒ ๋ฐ์ดํ„ฐ ๋ฐœ์ƒ์‹œ autoencoder ๋ชจ๋ธ์ถœ๋ ฅ์—์„œ ์†์‹ค๊ฐ’์ด ๋‹ค๋ฅด๊ฒŒ ์ถœ๋ ฅ๋œ๋‹ค. ์ด๋ฅผ ์ด์ƒ์ƒํ™ฉ ๊ฒ€์ถœ์—์„œ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค.

ํ•œ๊ตญ์ „๋ ฅ์˜ ์ „๋ ฅ์„ค๋น„ ์ค‘ ์ง€ํ•˜์ „๋ ฅ๊ตฌ์—๋Š” ๋ฐฐ์ „์„ ๋กœ ๋˜๋Š” ์†ก์ „์„ ๋กœ์™€ ๋”๋ถˆ์–ด ์ „๋ ฅ์„ค๋น„ ์•ˆ์ „์„ฑ ์ง„๋‹จ์„ ์œ„ํ•œ ๊ฐ์ข… ์ ๊ฒ€์„ค๋น„๊ฐ€ ๊ตฌ๋น„๋˜์–ด์žˆ๋‹ค. ์ „๋ ฅ์„ค๋น„ ์™ธ ์ž‘์—…์ž๋ฅผ ์œ„ํ•œ ๋งจํ™€, ํ™˜๊ธฐ๊ตฌ ๋“ฑ์ด ์žˆ์œผ๋ฉฐ, ํ™”์žฌ ๊ฐ์‹œ, ์นจ์ˆ˜๊ฐ์ง€, ์ถœ์ž…๋ฌธ๊ฐ์‹œ, ํ™˜๊ธฐํŒฌ๊ฐ์‹œ, ๊ตฌ๋‚ด์˜จ๋„ ์„ผ์„œ๋“ฑ์ด ์ง€ํ•˜์ „๋ ฅ๊ตฌ๋‚ด ์ฃผ์š” ์ง€์ ๋งˆ๋‹ค ์„ค์น˜๋˜์–ด ์žˆ๋‹ค. ์ „๋ ฅ์„ค๋น„ ์™ธ ๊ฐ์ข… ์ ๊ฒ€ ์„ค๋น„ ์—ญ์‹œ ํ•œ์ „์˜ ์ ๊ฒ€ ๋Œ€์ƒ ํ•ญ๋ชฉ์œผ๋กœ ์ง€ํ•˜์ „๋ ฅ๊ตฌ ํ™˜๊ฒฝ์— ๋Œ€ํ•œ ์ด์ƒ๊ฐ์ง€๋ฅผ ํ†ตํ•ด ๋” ํฐ ๋ฌธ์ œ๋กœ์˜ ํ™•์‚ฐ์„ ๋ฐฉ์ง€ํ•œ๋‹ค.

ํ•œ๊ตญ์ „๋ ฅ์€ ์ง€ํ•˜์ „๋ ฅ๊ตฌ์˜ ํ™˜๊ฒฝ์„ ๋ฌด์ธ ๊ฐ์‹œํ•˜๊ธฐ ์œ„ํ•ด Light Detection and Ranging(Lidar), RGB, ์Œํ–ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ Autoencoder ๋ชจ๋ธ์„ ์ด์šฉํ•œ ์ง€ํ•˜์ „๋ ฅ๊ตฌ ํ™˜๊ฒฝ ๊ฐ์‹œ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•˜์˜€๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์ง€ํ•˜์ „๋ ฅ๊ตฌ์—์„œ ์ทจ๋“ ๋ฐ ์ƒ์„ฑํ•œ ์Œํ–ฅ๋ฐ์ดํ„ฐ๋ฅผ ๋Œ€์ƒ์œผ๋กœ ๋จธ์‹ ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ์‹œ์„ค๋ฌผ์˜ ์ƒํƒœ๋ฅผ ๋ถ„์„ํ•˜๊ณ  ์ด์ƒ๋ฐœ์ƒ์‹œ ๊ฒฝ๋ณด๋ฅผ ์ œ๊ณตํ•œ๋‹ค.

๋ณธ ๋…ผ๋ฌธ์˜ ๊ตฌ์„ฑ์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค. 2์žฅ์—์„œ๋Š” ์ด์ƒํƒ์ง€๋ฅผ ๊ฒ€์ถœํ•˜๊ธฐ ์œ„ํ•œ ์˜คํ† ์ธ์ฝ”๋” ๊ด€๋ จ ์—ฐ๊ตฌ์— ๋Œ€ํ•ด ์†Œ๊ฐœํ•˜๊ณ , 3์žฅ์€ ์ง€ํ•˜์ „๋ ฅ๊ตฌ ๋‚ด์˜ ์Œํ–ฅ๋ฐ์ดํ„ฐ ์ทจ๋“ ๋ฐ ์ „์ฒ˜๋ฆฌ ๋ฐฉ๋ฒ•, 4์žฅ์—์„œ๋Š” ์ด์ƒํƒ์ง€ ๋ชจ๋ธ๋ง ๋ฐ ์‹œํ—˜๊ฒฐ๊ณผ์— ๋Œ€ํ•ด์„œ ์„ค๋ช…ํ•˜์˜€๊ณ , ๋งˆ์ง€๋ง‰ 5์žฅ์€ ๊ฒฐ๋ก ์— ๊ด€ํ•ด ๊ธฐ์ˆ ํ•˜์˜€๋‹ค.

2. Autoencoders

2.1 Autoencoder

๊ทธ๋ฆผ 1์€ Autoencoder์˜ ๊ตฌ์กฐ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. Autoencoder๋Š” encoder์™€ decoder๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๊ณ , ์ž…๋ ฅ๊ณผ ์ถœ๋ ฅ์˜ ๋…ธ๋“œ ์ˆ˜๊ฐ€ ๊ฐ™๋‹ค. Hidden Layer์˜ ์ˆ˜๊ฐ€ input layer์˜ ์ˆ˜๋ณด๋‹ค ์ž‘์•„์„œ ๋ฐ์ดํ„ฐ๋ฅผ ์ฐจ์› ์ถ•์†Œํ•˜๊ฑฐ๋‚˜ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ์— ๋…ธ์ด์ฆˆ๋ฅผ ์ถ”๊ฐ€ํ•œ ํ›„ ์›๋ณธ ์ž…๋ ฅ์„ ๋ณต์›ํ•  ์ˆ˜ ์žˆ๋„๋ก ๋„คํŠธ์›Œํฌ๋ฅผ ํ•™์Šต์‹œํ‚ค๋Š” ๋ชจ๋ธ์ด๋‹ค(5).

๊ทธ๋ฆผ. 1. Autoencoder์˜ ๊ตฌ์กฐ

Fig. 1. Architecture of Autoencoder

../../Resources/kiee/KIEEP.2020.69.2.69/fig1.png

Autoencoder์˜ architecture๋ฅผ ์„ค๋ช…ํ•˜๊ธฐ ์œ„ํ•ด์„œ ์•„๋ž˜์™€ ๊ฐ™์€ ์ •์˜๊ฐ€ ํ•„์š”ํ•˜๋‹ค.

1) $F, G$๋Š” ์ผ๋ฐ˜ ์ง‘ํ•ฉ์ด๋‹ค.

2) $n, p$๋Š” ์ž์—ฐ์ˆ˜์ด๋ฉฐ, $0<p<n$์„ ๋งŒ์กฑํ•œ๋‹ค.

3) $A$๋Š” $G^{P}$์—์„œ $F^{n}$๋กœ์˜ ํ•จ์ˆ˜๋“ค์˜ ํด๋ž˜์Šค๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.

4) $B$๋Š” $F^{n}$์—์„œ $G^{P}$๋กœ์˜ ํ•จ์ˆ˜๋“ค์˜ ํด๋ž˜์Šค๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.

5) $X=\left\{x_{1},\: x_{2},\:\cdots ,\: x_{m}\right\}$์€ $F^{n}$์˜ ํ•™์Šต ๋ฒกํ„ฐ๋“ค์˜ ์ง‘ํ•ฉ์„ ์˜๋ฏธํ•œ๋‹ค.

$Y=\left\{y_{1},\: y_{2},\:\cdots ,\: y_{m}\right\}$์€ $F^{n}$์˜ ํ•™์Šต ๋ฒกํ„ฐ๋“ค์— ๋Œ€์‘๋˜๋Š” ๋ชฉํ‘œ ๋ฒกํ„ฐ๋“ค์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.

6) ฮ”๋Š” $F^{n}$์—์„œ ์ •์˜๋œ ๋น„์œ ์‚ฌ ํ•จ์ˆ˜ (dissimilarity function)์ด๋‹ค.

Autoencoder์˜ ์ตœ์ข… ์ถœ๋ ฅ ์‹์€ (1)๊ณผ ๊ฐ™๋‹ค.

(1)
$$Y=A \circ B(X)$$

์—ฌ๊ธฐ์„œ, ์ž„์˜์˜ ํ•จ์ˆ˜ $A, B$๋Š” $AโˆˆA$, $BโˆˆB$์ด๊ณ , autoencoder ๋ณ€ํ™˜์˜ ์ž…๋ ฅ ๋ฒกํ„ฐ๋Š” $x_{t}$์ด๋ฉฐ autoencoder ๋ณ€ํ™˜์˜ ์ถœ๋ ฅ ๋ฒกํ„ฐ๋Š” (1)๊ณผ ๊ฐ™์ด $Y$์ด๋‹ค.

์ตœ์ ์˜ autoencoder ๋ณ€ํ™˜์„ ์œ„ํ•œ ์ตœ์ ํ•ด $A, B$๋Š” (2)์™€ ๊ฐ™์ด ์ •์˜๋œ ๋น„์œ ์‚ฌ ํ•จ์ˆ˜๋ฅผ ์ตœ์†Œํ™” ์‹œ์ผœ์•ผ ํ•œ๋‹ค.

(2)
$$\min_{A,\:B}E(A,\:B)=\min_{A,\:B}\sum_{t=1}^{m}E(x_{t})$$ $$=\min_{A,\:B}\sum_{t=1}^{m}\triangle(A\circ B(x_{t}),\:x_{t})$$

์—ฌ๊ธฐ์„œ $\Delta\left(a_{i}, b_{i}\right)=\left\|a_{i}-b_{i}\right\|^{2}$์ด๋‹ค.

2.2 Stacked Autoencoder

Stacked Autoencoder(SAE)๋Š” ์—ฌ๋Ÿฌ ๊ฐœ์˜ ํžˆ๋“  ๋ ˆ์ด์–ด๋ฅผ ๊ฐ€์ง€๋Š” ์˜คํ† ์ธ์ฝ”๋”์ด๋ฉฐ, ๋ ˆ์ด์–ด๋ฅผ ์ถ”๊ฐ€ํ• ์ˆ˜๋ก ์˜คํ† ์ธ์ฝ”๋”๊ฐ€ ๋” ๋ณต์žกํ•œ ํ•™์Šต์ด ๊ฐ€๋Šฅํ•˜๋‹ค(6). ์ž…๋ ฅํ•  ๋•Œ ์ž…๋ ฅ ๊ฐ’์ด ์••์ถ•๋จ์œผ๋กœ์จ ์†์‹ค ์—†๋Š” ์™„๋ฒฝํ•œ ์ถœ๋ ฅ ๋ณต์‚ฌ๋Š” ๋ถˆ๊ฐ€๋Šฅํ•˜๋ฉฐ, ํžˆ๋“  ๋ ˆ์ด์–ด์—์„œ ์ž…๋ ฅ์˜ ํŠน์„ฑ์„ ํ•™์Šตํ•œ๋‹ค. Stacked Autoencoder ๊ธฐ๋ฒ•์€ ์‚ฌ์šฉ์˜ ํŽธ๋ฆฌ์„ฑ๊ณผ ์œ ์—ฐ์„ฑ์œผ๋กœ ์ธํ•ด, ํ˜„์žฌ๋„ ๋งŽ์€ ์ด์ƒ ์ง„๋‹จ ๋ชจ๋ธ์—์„œ ๋‹ค์–‘ํ•œ ๋ฐฉ๋ฒ•์œผ๋กœ ์‚ฌ์šฉ๋˜๊ณ  ์žˆ๋Š” ๋ชจ๋ธ์ด๊ณ , ๋ ˆ์ด์–ด ์ถ”๊ฐ€ ์‚ญ์ œ๊ฐ€ ์šฉ์ดํ•˜์—ฌ ์ด์ƒ์ง„๋‹จ์— ํญ๋„“๊ฒŒ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค.

๊ทธ๋ฆผ. 2. Self-Attention module ์ƒ์„ฑ๋ฐฉ๋ฒ•

Fig. 2. Generation of Self-Attention module

../../Resources/kiee/KIEEP.2020.69.2.69/fig2.png

2.3 Attention Mechanism

Self attention ๋ชจ๋“ˆ์€ ์ด๋ฏธ์ง€์—์„œ ๋„“์€ ๋ฒ”์œ„์™€ ์—ฌ๋Ÿฌ ๋‹จ๊ณ„์˜ dependency์— ๋Œ€ํ•ด ๋ชจ๋ธ๋ง ํ•  ์ˆ˜ ์žˆ์–ด ๊ธฐ์กด convolution์˜ ๋‹จ์ ์„ ๋ณด์™„ํ•  ์ˆ˜ ์žˆ๋Š” ํŠน์ง•์ด ์žˆ๋‹ค. ๊ทธ๋ฆผ 2๋Š” self-attention์˜ ๋ฉ”์นด๋‹ˆ์ฆ˜์„ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์œผ๋กœ, ํ•œ ํ”ฝ์…€๊ณผ ๋‹ค๋ฅธ ํฌ์ง€์…˜์˜ ํ”ฝ์…€๋“ค ๊ฐ„์— ๊ด€๊ณ„๋ฅผ ๋ช…์‹œ์ ์œผ๋กœ ํ•™์Šตํ•˜๋„๋ก ์„ค๊ณ„๋œ ๊ฒƒ์ด๋‹ค.

Self attention์˜ ๋ฉ”์นด๋‹ˆ์ฆ˜์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค(7). Convolution feature maps(x)์€ 3๊ฐœ๋กœ ๋ณต์ œ๋˜์–ด ๋‚˜๋ˆ ์ง€๋ฉฐ, ์ด์ „ hidden layer $x\in R^{C\times N}$์˜ image feature๋Š” attention์„ ๊ณ„์‚ฐํ•˜๊ธฐ ์œ„ํ•ด ๋‘ ๊ฐœ์˜ ํ”ผ์ณ ๊ณต๊ฐ„ $f,\:{g} $๋กœ ๋ณ€ํ™˜๋˜๋ฉฐ, ์—ฌ๊ธฐ์„œ $f(x)= W_{f}x$, ${g}(x)= W_{{g}}x$,

(3)
$\beta_{j,\:i}=\dfrac{\exp(s_{i,\:j})}{\sum_{i=1}^{N}\exp(s_{i,\:j})},\: where \quad s_{i,\:j}=f(x_{i})^{T}{g}(x_{i})$

$\beta_{j,\:i}$๋Š” $j $๋ฒˆ์งธ ์ง€์—ญ์„ ํ•ฉ์„ฑํ•  ๋•Œ $i $๋ฒˆ ์œ„์น˜์— ์ง‘์ค‘ํ•˜๋Š” ์ •๋„๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์—ฌ๊ธฐ์„œ $C $๋Š” ์ฑ„๋„์˜ ์ˆ˜์ด๊ณ , $N $์€ ์ด์ „ hidden layer์—์„œ feature๋“ค์˜ feature location์˜ ๊ฐœ์ˆ˜์ด๋‹ค. Attention layer์˜ ์ถœ๋ ฅ์€ $O=(o_{1},\: o_{2},\:\cdots ,\: o_{j},\:\cdots ,\: o_{N})\in R^{C\times N}$

(4)
$o_{j}=v\left(\sum_{i=1}^{N}\beta_{j,\:i}h(x_{i})\right),\: h(x_{i})=W_{h}x_{i},\: v(x_{i})=W_{v}x_{i}$

$W_{{g}}\in R^{\bar{C}\times C},\: W_{f}\in R^{\bar{C}\times C},\: W_{h}\in R^{\bar{C}\times C},\:$$W_{v}\in R^{C\times\bar{C}}$๋Š” ํ•™์Šต๊ฐ€์ค‘ ๋งคํŠธ๋ฆญ์Šค์ด๋‹ค.

์ถœ๋ ฅ $y $๋Š” ์‹(5)์™€ ๊ฐ™์ด self-attention feature map(o)์— scale parameter $\gamma$๋ฅผ ๊ณฑํ•˜๊ณ , original input feature map์„ ๋”ํ•œ๋‹ค.

(5)
$y_{i}=x_{i}+\gamma o_{i}$

scale parameter $\gamma$๋Š” ์ดˆ๊ธฐ๊ฐ’์ด 0์ด์ง€๋งŒ ํ•™์Šต๊ณผ์ •์—์„œ ์ ์ฐจ ์ฆ๊ฐ€ํ•˜๊ณ , ๋„คํŠธ์›Œํฌ๋Š” ์ฒ˜์Œ์—๋Š” ๋กœ์ปฌ ์˜์—ญ์—๋งŒ ์˜์กดํ•˜๋‹ค๊ฐ€ ์ ์ฐจ ๋ฉ€๋ฆฌ ์žˆ๋Š” ์˜์—ญ์— ๋” ๋งŽ์€ ๊ฐ€์ค‘์น˜๋ฅผ ์ฃผ๋Š” ๋ฐฉ๋ฒ•์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋„๋ก ๊ตฌ์„ฑํ•œ๋‹ค. ๋”ฐ๋ผ์„œ Self-attention์ด ์ ์šฉ๋œ SAE๋Š” ์˜คํ† ์ธ์ฝ”๋”์˜ ์ธ์ฝ”๋”ฉ ๋‹จ๊ณ„์—์„œ ์‚ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๋Š” ๋””ํ…Œ์ผํ•œ feature ํŠน์„ฑ์„ ์ž˜ ์ฒ˜๋ฆฌํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 3. Convolutional filter์™€ self attention๊ฐ„์˜ ์ฐจ์ด

Fig. 3. Difference between convolutional filter and self attention

../../Resources/kiee/KIEEP.2020.69.2.69/fig3.png

์ผ๋ฐ˜์ ์ธ Convolution ๋„คํŠธ์›Œํฌ์—์„œ ํ•˜๋‚˜์˜ ํ”ฝ์…€์€ ์ž‘์€ ๋กœ์ปฌ ์˜์—ญ์œผ๋กœ ์ œํ•œํ•˜๋Š” ๋ฐ˜๋ฉด, Self-Attention์€ ๊ทธ๋ฆผ 3์—์„œ์™€ ๊ฐ™์ด, ์ผ๋ฐ˜์ ์ธ CNN ๊ตฌ์กฐ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ํ”ฝ์…€๊ณผ ์ด์›ƒ ๊ฐ„์˜ ํŠน์ง•์  ์œ„์ฃผ๋กœ ์ถ”์ถœํ•˜๋Š” ๋‹จ์ ์„ ๋ณด์™„ํ•˜์—ฌ ๋ฉ€๋ฆฌ ๋–จ์–ด์ง„ ์˜์—ญ์ด๋”๋ผ๋„ ์‰ฝ๊ฒŒ ์ „ ์˜์—ญ์—์„œ์˜ ์˜์กด์„ฑ์„ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค.

2.4 Adversarial-Auto Encoder

Adversarial-Auto Encoder(AAE)๋Š” Variational Auto-Encoder (VAE)์™€ Generative Adversarial Network(GAN)์„ ๊ฒฐํ•ฉํ•œ ์ค€์ง€๋„ ํ•™์Šต Neural Network์ด๋‹ค(8). AAE๋Š” VAE์˜ ๋‹จ์ ์ธ ํ‘œ๋ณธ์˜ ์ •๊ทœ๋ถ„ํฌ ๊ฐ€์ •๊ณผ ์ƒํ˜ธ ๊ฑฐ๋ฆฌ๊ณ„์‚ฐ ๋ฌธ์ œ๋ฅผ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด GAN ๊ตฌ์กฐ๋ฅผ ์ถ”๊ฐ€ํ•จ์œผ๋กœ์จ, ๋ณด๋‹ค ํšจ๊ณผ์ ์ธ ๋„คํŠธ์›Œํฌ ํ•™์Šต์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค. ์ƒ์„ฑ์ž์˜ encoder๋Š” ๋ฐ์ดํ„ฐ x๋ฅผ ์ž…๋ ฅ๋ฐ›์•„์„œ ์ž ์žฌ๋ณ€์ˆ˜ z๋ฅผ ์ƒ˜ํ”Œ๋งํ•˜๊ณ  ์ž ์žฌ ๊ณต๊ฐ„์—์„œ ๋ฐ์ดํ„ฐ๋ฅผ ๋งคํ•‘ํ•œ๋‹ค. ์ƒ์„ฑ์ž์˜ Generator๋Š” ์ด๋กœ๋ถ€ํ„ฐ ๋‹ค์‹œ x๋ฅผ ๋ณต์›ํ•œ๋‹ค. AAE๋Š” ๊ธฐ์กด VAE์—์„œ GAN์˜ ๊ตฌ๋ถ„์ž(discriminator)์—ญํ• ์„ ํ•˜๋Š” ๋„คํŠธ์›Œํฌ๊ฐ€ ์ถ”๊ฐ€๋˜์—ˆ๋‹ค. ์ด ๊ตฌ๋ถ„์ž๋Š” encoder๊ฐ€ ์œ„์กฐํ•œ z์™€ p(z)๋กœ๋ถ€ํ„ฐ ์ง์ ‘ ์ƒ˜ํ”Œ๋งํ•œ real z๋ฅผ ๊ตฌ๋ถ„ํ•œ๋‹ค. ํ•™์Šต์‹œ์—๋Š” ์ •์ƒ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜๊ณ , ์‹คํ–‰์‹œ Query Image์— ๋Œ€ํ•˜์—ฌ ๋ณต๊ตฌ ์˜์ƒ์„ ๊ตฌํ•œ ๋’ค MSE๋ฅผ ๋„์ถœํ•œ๋‹ค. ์ด MSE์˜ ์ตœ๋Œ€ ์ตœ์†Œ์น˜๋ฅผ ๊ณ ๋ คํ•œ Regularity score๋กœ ์ •์ƒ๊ณผ ๋น„์ •์ƒ์„ ํŒ๋ณ„ํ•  ์ˆ˜ ์žˆ๋‹ค.

3. AE๋ฅผ ์ด์šฉํ•œ ์ง€ํ•˜์ „๋ ฅ๊ตฌ ์ด์ƒํƒ์ง€์‹œ์Šคํ…œ

3.1 ์ด์ƒ ํƒ์ง€์‹œ์Šคํ…œ ๊ตฌ์กฐ

๊ทธ๋ฆผ 4๋Š” ์ด์ƒํƒ์ง€ ์‹œ์Šคํ…œ ๊ตฌ์กฐ๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ๊ฒƒ์œผ๋กœ ๋ฐ์ดํ„ฐ ์ž…๋ ฅ ํ›„ ์ „์ฒ˜๋ฆฌ๋ฅผ ํ†ตํ•ด ์ด๋ฏธ์ง€ํ™” ์‹œํ‚จ ํ›„ ๋”ฅ๋Ÿฌ๋‹ ํ•™์Šต์„ ์‹œํ‚จ ํ›„ ์†์‹ค๊ฐ’์„ ์ถ”์ถœํ•˜์—ฌ ์ •์ƒ, ๋น„์ •์ƒ์œ ๋ฌด๋ฅผ ํŒ๋‹จํ•œ๋‹ค. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ์Œํ–ฅ ๋ฐ์ดํ„ฐ๋งŒ์„ ๋Œ€์ƒ์œผ๋กœ 4๊ฐ€์ง€ ์ด์ƒํƒ์ง€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ ์šฉํ•˜๊ณ  ๊ทธ ์„ฑ๋Šฅ์— ๋Œ€ํ•ด ๋…ผํ•˜๊ณ ์ž ํ•œ๋‹ค.

๊ทธ๋ฆผ. 4. ์ด์ƒํƒ์ง€ ์‹œ์Šคํ…œ ๊ตฌ์กฐ๋„

Fig. 4. Architecture of Anomaly-Detection System for underground cable tunnel

../../Resources/kiee/KIEEP.2020.69.2.69/fig4.png

3.2 ์ด์ƒํƒ์ง€์šฉ ์Œํ–ฅ ๋ฐ์ดํ„ฐ ์ทจ๋“ ๋ฐ ์ „์ฒ˜๋ฆฌ

์Œํ–ฅ ๋ฐ์ดํ„ฐ๋Š” 9์ข…(rain, water drop, human walking, convert- sation, water pump, engine, drilling, jack hammer, construction)์˜ ์ •์ƒ ๋ฐ์ดํ„ฐ ์…‹๊ณผ 1์ข…(siren)์˜ ์ด์ƒ ๋ฐ์ดํ„ฐ ์…‹์œผ๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. Water drop, water pump, human walking์˜ ์Œ์›์€ ํ˜„์žฅ ์ง€ํ•˜ ์ „๋ ฅ๊ตฌ์—์„œ ์ƒํ™ฉ์„ ๋ฐœ์ƒ์‹œ์ผœ ๋…น์Œํ•˜์˜€๊ณ , rain, conversation, engine, drilling, jack hammer, construction, siren 7์ข…์˜ ๋ฐ์ดํ„ฐ๋Š” ๋‰ด์š•๋Œ€์˜ MARL(Music and Audio Research Lab)(9)์—์„œ ๋‹ค์šด๋ฐ›์•„ ํ™œ์šฉํ•˜์˜€๋‹ค. ์œ„ 7์ข…์˜ ์Œํ–ฅ์„ ์žฌ์ƒํ•˜์—ฌ ์ง€ํ•˜์ „๋ ฅ๊ตฌ ํ˜„์žฅ ํŠน์œ ์˜ ์žก์Œ๊ณผ ๊ฐ™์ด ๋…น์Œํ•˜์—ฌ ์Œํ–ฅ ๋ฐ์ดํ„ฐ๋ฅผ ํ™•๋ณดํ•˜์˜€๋‹ค. Mel Frequency Cepstral Coefficient(MFCC) ๊ธฐ๋ฐ˜์˜ ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ด 10์ข… ์Œํ–ฅ ํด๋ž˜์Šค ์ค‘ 9์ข…์˜ ์œ ์‚ฌ ์Œํ–ฅ ์‹ ํ˜ธ(water drop, jackhammer ๋“ฑ)์„ ์ •์ƒ ์‹ ํ˜ธ๋กœ ๋ถ„๋ฅ˜ํ•˜๊ณ , ํŠน์„ฑ์ด ๋‹ค๋ฅธ 1์ข…์˜ ์Œํ–ฅ์‹ ํ˜ธ(siren)๋ฅผ ๋น„์ •์ƒ ์‹ ํ˜ธ๋กœ ์ƒ‰์ธํ•˜์—ฌ, ๋น„์ •์ƒ ์‹ ํ˜ธ ํŒ๋ณ„ ์—ฌ๋ถ€๋ฅผ ์‹œํ—˜ํ•˜์˜€๋‹ค.

๊ทธ๋ฆผ. 5. ์ธก์ •ํ•œ 10์ข… ์Œํ–ฅ ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ

Fig. 5. Measured Sound Data

../../Resources/kiee/KIEEP.2020.69.2.69/fig5.png

๊ทธ๋ฆผ. 6. ์Œํ–ฅ ์‹ ํ˜ธ ์ „์ฒ˜๋ฆฌ(MFCC) ๊ฒฐ๊ณผ

Fig. 6. MFCC Preprocessed Sound Signal

../../Resources/kiee/KIEEP.2020.69.2.69/fig6.png

๊ทธ๋ฆผ. 7. ์ทจ๋“ ๋ฐ์ดํ„ฐ ์ˆ˜๋Ÿ‰

Fig. 7. Number of Sample data

../../Resources/kiee/KIEEP.2020.69.2.69/fig7.png

์˜ค๋””์˜ค ๋ฐ์ดํ„ฐ๋Š” MFCC ์ „์ฒ˜๋ฆฌ ๊ธฐ๋ฒ•์„ ์ ์šฉํ•˜์—ฌ 1์ฐจ์›์˜ ์‹œ๊ณ„์—ด ๋ฐ์ดํ„ฐ๋ฅผ 2์ฐจ์›์˜ ์ด๋ฏธ์ง€ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ๋กœ ๊ฐ€๊ณตํ•˜์˜€๋‹ค. MFCC๋Š” ์Œํ–ฅ์‹ ํ˜ธ ์ฒ˜๋ฆฌ์— ์žˆ์–ด, ๋‹จ๊ธฐ ์‹ ํ˜ธ์˜ ํŒŒ์›Œ ์ŠคํŽ™ํŠธ๋Ÿผ์„ ํ‘œํ˜„ํ•˜๋Š” ๋ฐฉ๋ฒ• ์ค‘ ํ•˜๋‚˜๋กœ์จ, ์ „๋ ฅ ์ŠคํŽ™ํŠธ๋Ÿผ์— ๋น„์„ ํ˜•์˜ Mel ์Šค์ผ€์ผ ์ฃผํŒŒ์ˆ˜ ํ•„ํ„ฐ๋ฅผ ์ ์šฉํ•˜์—ฌ ์–ป๋Š” ๋ฐฉ๋ฒ•์„ ๋งํ•œ๋‹ค(10). MFCC ํŠน์ง• ๋ฒกํ„ฐ๋Š” ๊ทธ๋ฆผ 6๊ณผ ๊ฐ™์ด 48๊ฐœ์˜ MFCC ๊ณ„์ˆ˜๋ฅผ ์ด์šฉํ•˜๊ณ , 12kHz ์ƒ˜ํ”Œ๋ง ์ฃผํŒŒ์ˆ˜์— ๋”ฐ๋ผ 4์ดˆ ์Œํ–ฅ ์ƒ˜ํ”Œ์˜ ์œˆ๋„์šฐ ์ค‘์ฒฉ์„ ํ†ตํ•ด 188๊ฐœ์˜ ํŠน์ง• ๋ฒกํ„ฐ๋ฅผ ๊ฐ–๋„๋ก ๊ตฌ์„ฑํ•˜์—ฌ, ์ƒ˜ํ”Œ ๋‹น 48ร—188์˜ ํ•™์Šต ์ž…๋ ฅ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ–๋„๋ก ํ•˜์˜€๋‹ค. ์‚ฌ์šด๋“œ๋ฅผ ์ž˜ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ์ŠคํŽ™ํŠธ๋Ÿผ ํ˜•ํƒœ์˜ ์ด๋ฏธ์ง€๋กœ ๋ณ€ํ™˜ํ•˜์˜€๊ณ , ํ•œ ์ŠคํŽ™ํŠธ๋Ÿผ ๋‹น 4์ดˆ์”ฉ ๋ฐ์ดํ„ฐ๋ฅผ ๋งŒ๋“ค๊ณ , MFCC๋ฅผ ํฌํ•จํ•œ ์ „์ฒ˜๋ฆฌ ๊ณผ์ •์„ ๊ฑฐ์ณ ๊ทธ๋ฆผ 7๊ณผ ๊ฐ™์ด ๋ฐ์ดํ„ฐ๋ณ„๋กœ 1,000๊ฐœ ์ด์ƒ์”ฉ 10์ข…์œผ๋กœ 10,000๊ฐœ ์ด์ƒ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜์˜€๋‹ค.

4. ๋ชจ๋ธ๋ง ๋ฐ ์„ฑ๋Šฅ ๊ฒ€์ฆ

4.1 Stacked Autoencoder ๋ชจ๋ธ๋ง

์Œํ–ฅ ์ด์ƒ์ง„๋‹จ์šฉ SAE ๋ชจ๋ธ์€ convolution layer, max-pooling layer, upsampling layer, batch_normalization layer๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. ์Œํ–ฅ์‹ ํ˜ธ์— ๋Œ€ํ•œ SAE ๋ชจ๋ธ์˜ ์ธ์ฝ”๋”๋Š” Convolution layer- Leaky Relu layer-Maxpooling layer์˜ ๋ฐ˜๋ณต์œผ๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. ๋””์ฝ”๋”๋Š” ๋Œ€์นญ๊ตฌ์กฐ๋กœ ๊ฐœ๋ฐœํ•˜๊ณ  ๋งˆ์ง€๋ง‰ Activation layer๋Š” sigmoid ํ•จ์ˆ˜๋ฅผ ํ™œ์šฉํ•˜์˜€๋‹ค. z-dimension์€ 8๋กœ ์••์ถ•ํ•˜์˜€๊ณ , epoch์€ 1000์œผ๋กœ ์ง„ํ–‰ํ•˜์˜€๋‹ค. Optimizer๋Š” Adam ๊ธฐ๋ฒ•์„, Lossํ•จ์ˆ˜๋กœ๋Š” MSE๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์Œํ–ฅ ๋ฐ์ดํ„ฐ๋Š” MFCC ์ „์ฒ˜๋ฆฌ๊ธฐ๋ฒ•์„ ํ†ตํ•ด 2D ์ด๋ฏธ์ง€ํ™”ํ•˜์—ฌ 48ร—188ร—1์˜ ์ž…๋ ฅ์‚ฌ์ด์ฆˆ๋กœ ์ •์ƒ์ƒํƒœ์˜ ๋ฐ์ดํ„ฐ ์…‹ 9์ข…์˜ ์Œํ–ฅ ํด๋ž˜์Šค๋ฅผ SAE ๋ชจ๋ธ์— ํ•™์Šต ์‹œํ‚จ๋‹ค. Criteria ๊ฐ’์€ ํ•™์Šต์„ ํ†ตํ•ด ๋„์ถœ๋œ ์ตœ๋Œ€๊ฐ’์„ ๊ธฐ์ค€์œผ๋กœ ์ •ํ•˜์˜€๊ณ , criteria๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์ •์ƒ๊ณผ ๋น„์ •์ƒ์„ ๊ตฌ๋ถ„ํ•˜์˜€๋‹ค. ๊ทธ๋ฆผ 8์€ ๋ฐ˜๋ณต์ ์ธ ์‹คํ—˜์„ ํ†ตํ•ด ๊ตฌ์„ฑํ•œ ์ตœ์ ์˜ SAE ๋ชจ๋ธ์„ ๋ณด์—ฌ์ค€๋‹ค. ์ด ๋ชจ๋ธ์˜ ์ด์ƒ์ง„๋‹จ ์„ฑ๋Šฅ์€ 87%๋ฅผ ๋ณด์˜€๋‹ค.

๊ทธ๋ฆผ. 8. Stacked Autoencoder ๊ตฌ์กฐ

Fig. 8. Abnormal Sound Diagnostic Model using Stacked Autoencoder

../../Resources/kiee/KIEEP.2020.69.2.69/fig8.png

4.2 Adversarial Autoencoder ๋ชจ๋ธ๋ง

Adversarial Autoencoder(AAE) ๋ชจ๋ธ์€ ํฌ๊ฒŒ Encoder, Generator, Discriminator ๊ตฌ์กฐ๋กœ ๊ตฌ๋ถ„๋œ๋‹ค. ์ธ์ฝ”๋”, Generator ๊ตฌ์กฐ๋Š” AE ๊ตฌ์กฐ์™€ ์œ ์‚ฌํ•˜๊ณ , Discriminator ๊ตฌ๋ถ„์ž์˜ ๊ฒฝ์šฐ Dense Layer๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. ๊ทธ๋ฆผ 9์˜ AAE ๋ชจ๋ธ์—์„œ๋Š” ํ‘œ์ค€ํŽธ์ฐจ(ฯƒ)์™€ ํ‰๊ท ๊ฐ’(ฮผ)๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ latent space z๋ฅผ ์„ค์ •ํ•œ๋‹ค. ์ดํ›„ Discriminator์—์„œ ์ •์ƒ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šต์„ ์ง„ํ–‰ํ•œ๋‹ค.

AE ๋ชจ๋ธ์—์„œ๋Š” ์•„๋ž˜ ๊ทธ๋ฆผ๊ณผ ๊ฐ™์ด MFCC๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ 48ร—188ํฌ๊ธฐ์˜ ์Œํ–ฅ ํŠน์ง•์ž๋ฅผ ๊ตฌํ˜„ํ•˜๊ณ , ์ด๋ฅผ AAE ๋ชจ๋ธ์˜ ์ž…๋ ฅ์œผ๋กœํ•œ๋‹ค. ์ธ์ฝ”๋” ๋‹จ์€ 4ร—188ร—48(1stLayer)+4ร—94ร—24(2ndLayer) +8ร— 94ร—24(3rd Layer)+8ร—47ร—12(4th Layer)+4512(5th Layer)๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. ์ƒ์„ฑ์ž ๋‹จ์€ 4512(1st Layer)+8ร—47ร—12(2nd Layer)+8ร— 94ร—24(3rd Layer)+4ร—94ร—24(4th Layer)+4ร—188ร—48(5th Layer)๋กœ ๊ตฌ์„ฑ๋˜์–ด์žˆ๋‹ค. ์ธ์ฝ”๋” ๋‹จ์˜ 5๋ฒˆ์งธ ๋ ˆ์ด์–ด์—์„œ ์ž„์˜์˜ ํ‘œ์ค€ํŽธ์ฐจ(ฯƒ)์™€ ํ‰๊ท ๊ฐ’(u)์„ ๋ฐ”ํƒ•์œผ๋กœ Kullback Leibler-Divergence๋ฅผ ์ด์šฉํ•˜์—ฌ q(z)์™€ p(z) ์ƒ์„ฑํ•˜์˜€๋‹ค. ์ดํ›„ ๊ตฌ๋ถ„์ž๋‹จ์—์„œ๋Š” Fully connected Layer๋กœ 16(1st Layer)+8(2nd Layer)+1(3rd Layer)๋กœ ๊ตฌ์„ฑํ•˜์—ฌ ๋ชจ๋ธ๋งํ•˜์˜€๋‹ค.

๊ทธ๋ฆผ. 9. AAE๋ฅผ ์‚ฌ์šฉํ•œ ์ด์ƒ์Œํ–ฅ ์ง„๋‹จ ๋ชจ๋ธ

Fig. 9. Abnormal Sound Diagnostic Model using AAE

../../Resources/kiee/KIEEP.2020.69.2.69/fig9.png

4.3 Attention ์ถ”๊ฐ€ SAE ๋ชจ๋ธ๋ง

๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์ด ํŠน์ • ๋ฒกํ„ฐ์— ์ฃผ๋ชฉํ•˜์—ฌ ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋†’์ด๋Š” ๊ธฐ๋ฒ•์ธ self-attention์„ SAE์— ์ ์šฉํ•˜์˜€๋‹ค. Attention ๊ธฐ๋ฐ˜ SAE ๋ชจ๋ธ์€ SAE ๋ชจ๋ธ ์ฝ”๋“œ์— โ€˜Attentionโ€™ ๋ ˆ์ด์–ด๋ฅผ ์ธ์ฝ”๋” ํŒŒํŠธ์— ์ถ”๊ฐ€ํ•œ๋‹ค. Attention ๋ชจ๋“ˆ์˜ ํฌ๊ธฐ ์กฐ์ • ๋ฐ ์œ„์น˜๋ณ€๊ฒฝ ์‹คํ—˜์„ ํ†ตํ•ด ์ธ์ฝ”๋” ๋งˆ์ง€๋ง‰ hidden layer์˜ ์•ž๋‹จ์— ์œ„์น˜์‹œํ‚ค๊ณ , ์‚ฌ์ด์ฆˆ๋ฅผ 24ร—94ร—16๋กœ ํ•˜์˜€์„ ๋•Œ ์ด์ƒ ํƒ์ง€ ์„ฑ๋Šฅ์ด ๊ฐœ์„ ๋จ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ๋‹ค. Attention ๊ธฐ๋ฐ˜ SAE ๋ชจ๋ธ์€ 91%๋กœ, attention ๋ชจ๋“ˆ์„ ์ถ”๊ฐ€ํ•จ์œผ๋กœ ์ธํ•ด ์„ฑ๋Šฅ์ด 4% ์ฆ๊ฐ€ํ•˜์˜€์Œ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 10. Attention ๊ธฐ๋ฐ˜ SAE๋ชจ๋ธ ๋ฐ ํ”Œ๋กœ์šฐ ์ฐจํŠธ

Fig. 10. Abnormal Sound Diagnostic Model using Attention based SAE

../../Resources/kiee/KIEEP.2020.69.2.69/fig10.png

๊ทธ๋ฆผ. 11. Attention ๊ธฐ๋ฐ˜ AAE๋ชจ๋ธ

Fig. 11. Attention based AAE Model

../../Resources/kiee/KIEEP.2020.69.2.69/fig11.png

4.4 Attention ์ถ”๊ฐ€ AAE ๋ชจ๋ธ๋ง

Attention layer๋ฅผ ์ธ์ฝ”๋“œ์™€ generator ๋‹จ์— ์˜ฎ๊ฒจ๊ฐ€๋ฉด์„œ ๋ฐ˜๋ณต์‹คํ—˜์„ ํ•œ ๊ฒฐ๊ณผ, AAE ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋–จ์–ด ์กŒ์ง€๋งŒ, generator ๋‹จ์— ์ตœ์†Œ ์‚ฌ์ด์ฆˆ์™€ ๊ฐ™์€ ํฌ๊ธฐ์˜ attention layer(12ร—47ร—8)๋ฅผ ์‚ฝ์ž…ํ•˜์˜€์„ ๊ฒฝ์šฐ, ์ด์ƒํƒ์ง€ ์ •ํ™•๋„๊ฐ€ ๊ทธ๋‚˜๋งˆ ๋œ ์ €ํ•˜๋จ์„ ํ™•์ธํ•˜์˜€๋‹ค. ๋ณธ ์ด์ƒํƒ์ง€ ์‹œ์Šคํ…œ์—์„œ๋Š” AAE ๋ชจ๋ธ์— attention layer ์ถ”๊ฐ€ํ•  ๊ฒฝ์šฐ ์„ฑ๋Šฅ์ด 70%๋กœ, AAE ๋ชจ๋ธ์— ๋น„ํ•ด ์„ฑ๋Šฅ์ด ์˜คํžˆ๋ ค 14% ๋‚ฎ์•„์ง์„ ๋ณด์˜€๋‹ค.

4.5 ์Œํ–ฅ ํ•™์Šต๋ชจ๋ธ ์‹œํ—˜ ๊ฒฐ๊ณผ

์‹ค์ œ ์ •์ƒ์œผ๋กœ ์šด์˜๋˜๊ณ  ์žˆ๋Š” ์ง€ํ•˜์ „๋ ฅ๊ตฌ๋ฅผ ๋Œ€์ƒ์œผ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์ทจ๋“ํ•˜์˜€์œผ๋ฏ€๋กœ ์ˆ˜์ง‘๋œ ๋ฐ์ดํ„ฐ๋Š” ๋Œ€๋ถ€๋ถ„ ์ •์ƒ์ด๋ฉฐ, ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ์ธ์œ„์ ์œผ๋กœ ์ด์ƒ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜์—ฌ ์„ฑ๋Šฅ ๋ถ„์„์„ ์‹œํ–‰ํ•˜์˜€๋‹ค. ๋ชจ๋ธ์€ ์ด 10,000๊ฐœ์˜ ๋ฐ์ดํ„ฐ ์ค‘ ์ •์ƒ๋ฐ์ดํ„ฐ 9000๊ฐœ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šต์‹œ์ผฐ๊ณ , ํ…Œ์ŠคํŠธ์‹œ์—๋Š” ํ•™์Šต์— ์‚ฌ์šฉ๋˜์ง€ ์•Š์€ 500๊ฐœ์˜ ์ •์ƒ ๋ฐ์ดํ„ฐ์™€ 500๊ฐœ์˜ ์ด์ƒ ๋ฐ์ดํ„ฐ๋ฅผ ์ž…๋ ฅํ•˜์—ฌ MSE ํ•จ์ˆ˜๋ฅผ ์ ์šฉํ•˜์—ฌ ์†์‹ค๊ฐ’์„ ๋น„๊ตํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ ์ด์ƒ ์Œํ–ฅ ์—ฌ๋ถ€๋ฅผ ํŒ๋ณ„ํ•˜๊ณ  ์„ฑ๋Šฅ์„ ํ‰๊ฐ€ํ•˜์˜€๋‹ค.

๊ทธ๋ฆผ 12๋Š” attention ๊ธฐ๋ฐ˜ SAE ๋ชจ๋ธ์„ ๋ฐ˜๋ณต ํ•™์Šต์‹œํ‚ฌ ๋•Œ์˜ ์†์‹ค๊ฐ’์„ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์†์‹ค๊ฐ’์ด ์ง€์ˆ˜ํ•จ์ˆ˜์ ์œผ๋กœ ๊ฐ์†Œํ•˜์—ฌ ์ˆ˜๋ ดํ•˜๋ฏ€๋กœ ๋ชจ๋ธ์ด ์ •์ƒ์ ์œผ๋กœ ํ•™์Šต๋˜์—ˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 12. ์–ดํ…์…˜ ๊ธฐ๋ฐ˜ ์˜คํ† ์ธ์ฝ”๋” ๋ชจ๋ธ ํ•™์Šต ์ปค๋ธŒ

Fig. 12. Loss value curve for Attention based SAE Model

../../Resources/kiee/KIEEP.2020.69.2.69/fig12.png

๊ทธ๋ฆผ. 13. ์–ดํ…์…˜๊ธฐ๋ฐ˜ AAE ๋ชจ๋ธ์—์„œ์˜ ์ž ์žฌ๋ณ€์ˆ˜ ICA ๊ตฐ์ง‘ํ™” ๋น„๊ต

Fig. 13. ICA Clustering of Latent Variables for Attention Based AAE

../../Resources/kiee/KIEEP.2020.69.2.69/fig13.png

๊ทธ๋ฆผ 13์€ Independent Component Analysis(ICA) ๋ฐฉ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ AAE ๋ชจ๋ธ์˜ ์ž ์žฌ๋ณ€์ˆ˜ ๊ตฐ์ง‘ํ™”๋ฅผ 3์ฐจ์›์œผ๋กœ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์ด๋‹ค. latent ์‚ฌ์ด์ฆˆ๊ฐ€ 6์ผ ๋•Œ ๋นจ๊ฐ„์ƒ‰์€ ์ •์ƒ, ๋…ธ๋ž€์ƒ‰์€ ๋น„์ •์ƒ ๋ฐ์ดํ„ฐ์˜ ๊ตฐ์ง‘ํ™”๋ฅผ ๋‚˜ํƒ€๋‚ด๋ฉฐ ์ •์ƒ์ ์œผ๋กœ ๋ชจ๋ธ๋ง์ด ๋˜์—ˆ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ 14๋Š” attention ๊ธฐ๋ฐ˜ SAE ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์ด์ƒ ํŒ์ • ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ฃผ๋Š” ๊ฒƒ์œผ๋กœ, ๋นจ๊ฐ„์ƒ‰ โ–ณ๋Š” ์ด์ƒ๋ฐ์ดํ„ฐ, ํŒŒ๋ž€์ƒ‰ xํ‘œ๋Š” ์ •์ƒ๋ฐ์ดํ„ฐ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. Loss๊ฐ’์ด โ€“250~-300 ๊ทผ์ฒ˜์ผ ๊ฒฝ์šฐ ์ด์ƒ ์ง„๋‹จ์— ์—๋Ÿฌ๊ฐ€ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค.

๊ทธ๋ฆผ. 14. ์ •์ƒ/์ด์ƒ TEST Data์— ๋Œ€ํ•œ ์–ดํ…์…˜ ๊ธฐ๋ฐ˜ ์˜คํ† ์ธ์ฝ”๋” ๋ชจ๋ธ์˜ ์ถœ๋ ฅ ๊ตฐ์ง‘

Fig. 14. Normal/Abnormal test data cluster of attention based AE Model output

../../Resources/kiee/KIEEP.2020.69.2.69/fig14.png

๋ชจ๋ธ ์ถœ๋ ฅ์˜ ์ด์ƒํŒ๋‹จ ๊ธฐ์ค€์€ ํ…Œ์ŠคํŠธ์šฉ ๋ฐ์ดํ„ฐ์„ธํŠธ๋ฅผ ํ•™์Šต์ด ์™„๋ฃŒ๋œ ๋ชจ๋ธ์— ์ž…๋ ฅํ•œ ๋’ค, ์ž…๋ ฅ๋œ ์ด๋ฏธ์ง€์™€ ์ถœ๋ ฅ๋œ ์ด๋ฏธ์ง€ ๊ฐ„์˜ MSE๋ฅผ ๊ธฐ์ค€์œผ๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. Criteria ๊ฐ’์€ ํ•™์Šต์‹œ ์ •์ƒ ์ƒ˜ํ”Œ๋ง ๋ฐ์ดํ„ฐ 100๊ฐœ๋ฅผ ๊ธฐ์ค€์œผ๋กœ 80ํผ์„ผํƒ€์ผ์„ ์ •์ƒ/๋น„์ •์ƒ์„ ํŒ๋ณ„ํ•˜๋„๋ก ํ•˜์˜€๋‹ค. ์„ฑ๋Šฅํ‰๊ฐ€ Accuracy๋Š” confusion matrix์˜ Accuracy = (TP + TN) / (TP + FP + FN + TN)๋กœ ์‚ฌ์šฉํ•˜์˜€๋‹ค.

์œ„์˜ 4๊ฐ€์ง€ ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ด์ƒํŒ๋ณ„ ์„ฑ๋Šฅ์€ ํ‘œ 1๊ณผ ๊ฐ™๋‹ค. SAE ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ 87%, attention layer๋ฅผ ์ถ”๊ฐ€ํ•œ SAE ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ 91%๋กœ ์„ฑ๋Šฅ์ด 4% ์ฆ๊ฐ€ํ•˜์˜€์Œ์„ ์•Œ ์ˆ˜ ์žˆ๋‹ค. AAE ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์€ 84%, ์—ฌ๊ธฐ์— attention layer๋ฅผ ์ถ”๊ฐ€ํ•˜์˜€์„ ๊ฒฝ์šฐ 70%๋กœ ์„ฑ๋Šฅ์ด ์˜คํžˆ๋ ค ๋‚ฎ์•„์ง์„ ๋ณผ ์ˆ˜ ์žˆ๋‹ค. AAE ๋ชจ๋ธ ํ›ˆ๋ จ ์‹œ์—๋„ ์ด๋ฏธ ๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ ์ˆ˜ ๋Œ€๋น„ํ•˜์—ฌ ํ›ˆ๋ จ๋ฐ์ดํ„ฐ๊ฐ€ ๋งŽ์ง€ ์•Š์•˜๋Š”๋ฐ, ์ด์— attention ๋ชจ๋ธ์„ ์ถ”๊ฐ€ํ•จ์œผ๋กœ์จ, ๋ชจ๋ธ ๋‚ด์—์„œ ์–ธ๋”ํ”ผํŒ… ํ˜„์ƒ์ด ๋ฐœ์ƒํ–ˆ์„ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค. ํ–ฅํ›„ Attention-based AAE ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ๋†’์ด๊ธฐ ์œ„ํ•ด ํ›ˆ๋ จ๋ฐ์ดํ„ฐ๋ฅผ ์ถ”๊ฐ€ํ•˜์—ฌ ์‹คํ—˜ ์˜ˆ์ •์ด๋‹ค.

ํ‘œ 1. ์„ฑ๋Šฅ ๋น„๊ต

Table 1. Performance Comparison of AE(Jetson TX2)

์„œ๋น„์Šค

๋Œ€์ƒ

SAE

AAE

Attention-

based SAE

Attention-

based AAE

์ด์ƒ์Œํ–ฅ ํƒ์ง€ ์„ฑ๋Šฅ

87%

84%

91%

70%

5. ๊ฒฐ ๋ก 

๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” stacked autoencoder, adversarial autoencoder, attention based SAE ๋ฐ attention based AAE ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์Œํ–ฅ ๋ฐ์ดํ„ฐ๋ฅผ ์ด์šฉํ•˜์—ฌ ์ง€ํ•˜์ „๋ ฅ๊ตฌ์˜ ์ด์ƒ์„ ์ง„๋‹จํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ๋ฐœํ•˜์˜€๋‹ค. ์ง€ํ•˜์ „๋ ฅ๊ตฌ ํ˜„์žฅ์—์„œ ์‹ค์‹œ๊ฐ„ ๋ถ„์„์„ ์œ„ํ•˜์—ฌ Edge Device(Jetson Tx2)ํ™˜๊ฒฝ ํ•˜์— ๋™์ž‘ํ•˜๋„๋ก ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํƒ‘์žฌํ•˜์˜€๋‹ค.

์˜คํ† ์ธ์ฝ”๋”์˜ ํŠน์ง•์ธ ์žฌ๊ตฌ์„ฑ์˜ค์ฐจ(์†์‹ค๊ฐ’)๊ณผ ์ž ์žฌ๋ณ€์ˆ˜ ๊ตฐ์ง‘ํ™”๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ „๋ ฅ์„ค๋น„ ํ™˜๊ฒฝ์˜ ์ƒํƒœ๋ณ€ํ™”๋ฅผ ํƒ์ง€ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์˜€๋‹ค. ์ ์ธต ์˜คํ† ์ธ์ฝ”๋”์˜ ๊ตฌ์กฐ๋ฅผ ๋‹ค์–‘ํ•˜๊ฒŒ ๋ณ€๊ฒฝํ•˜์—ฌ ์ตœ์ ์˜ ๋ชจ๋ธ์„ ๊ตฌ์„ฑํ•˜์˜€๊ณ , ์„ฑ๋Šฅํ–ฅ์ƒ์„ ์œ„ํ•ด attention ๋ฉ”์นด๋‹ˆ์ฆ˜์„ SAE ๋ชจ๋ธ์— ์ถ”๊ฐ€ ์ ์šฉํ•˜์˜€๋‹ค. ์‹คํ—˜๊ฒฐ๊ณผ SAE ๋ชจ๋ธ์—์„œ 87%์˜ ๊ฒ€์ถœ์œจ์„ ๋ณด์˜€๊ณ , attention ๋ฉ”์นด๋‹ˆ์ฆ˜์„ SAE ๋ชจ๋ธ์— ์ถ”๊ฐ€ํ•˜์˜€์„ ๊ฒฝ์šฐ 91% ์ด์ƒ์˜ ๋น„๊ต์  ๋†’์€ ์ •ํ™•๋„๋ฅผ ๋ณด์˜€๋‹ค. Adversarial autoencoder์˜ ๊ฒฝ์šฐ SAE ๋ณด๋‹ค ๋‚ฎ์€ 84%์˜ ๊ฒ€์ถœ์œจ์„ ๋ณด์˜€๊ณ , attention ๋ฉ”์นด๋‹ˆ์ฆ˜์„ AAE ๋ชจ๋ธ์— ์ถ”๊ฐ€ํ•˜์˜€์„ ๊ฒฝ์šฐ 70%์˜ ๊ฒ€์ถœ์œจ์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ ์˜คํžˆ๋ ค ์„ฑ๋Šฅ์ด ์ €ํ•˜๋จ์„ ์•Œ ์ˆ˜ ์žˆ์—ˆ๋‹ค.

์ง€ํ•˜์ „๋ ฅ๊ตฌ๋ผ๋Š” ์—ด์•…ํ•œ ํ™˜๊ฒฝ์—์„œ ๋ณธ ์—ฐ๊ตฌ๊ฒฐ๊ณผ์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ ์šฉํ•œ ์ด์ƒํƒ์ง€์‹œ์Šคํ…œ์„ ํ™œ์šฉํ•˜์—ฌ ์‹ค์‹œ๊ฐ„ ์›๊ฒฉ ๊ฐ์‹œ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•จ์œผ๋กœ ์ด์ƒ๋ฐœ์ƒ์‹œ ์ฆ‰๊ฐ์ ์ธ ๋Œ€์ฒ˜๋ฅผ ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜์˜€๋‹ค. ํ–ฅํ›„ ์ง€ํ•˜์ „๋ ฅ๊ตฌ ์ด์ƒํƒ์ง€์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์œ„ํ•ด์„œ๋Š” ์Œํ–ฅ ๋ฐ์ดํ„ฐ ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ Lidar, ์˜์ƒ๋ฐ์ดํ„ฐ, ์˜จ์Šต๋„ ๋ฐ์ดํ„ฐ ๋“ฑ์„ ๋ณตํ•ฉ์ ์œผ๋กœ ํ™œ์šฉํ•œ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค.

References

1 
S. H. Lee, B. D. Yun, 2015, Industry 4.0 and Direction of Failure Prediction and Health Management Tech- nology(PHM), Journal of the Korean Society for Noise and Vibration Engineering, Vol. 25, No. 1, pp. 22-28Google Search
2 
Chalapathy Raghavendra, Chawla Sanjay, , Deep Learning for Anomaly Detection : A Survey, arxive.org/pdf/1901. 03407.pdfGoogle Search
3 
P. Vincent, H. Larochelle, I. Lajoie, Y. Bengio, P. A. Manzagol, 2010, Stacked Denoising Autoencoder Learning Useful Representations in a Deep Network with a Local Denoising Criterion, The Journal of Machine Learning Research, Vol. 11, No. 3, pp. 3371-3408DOI
4 
Dong-Jin Choi, , Real-Time Self Complement System of Fault Diagnosis for Induction Motor Using Machine Learning and IoT Technique, The Transaction of the Korean Institute of Electrical Engineers, Vol. 68, No. 5, pp. 662-669Google Search
5 
Baldi Pierre, 2012, Autoencoders, Unsupervised Learning, and Deep Architectures, JMLR: Workshop and Conference Proceedings, Vol. 27, pp. 37-50Google Search
6 
Zabalza Jamie, 2015, Novel Segmented Stacked AutoEncoder for Effective Dimensionality Reduction and Feature Extrac- tion in Hyperspectral Imaging, Neurocomputing JournalGoogle Search
7 
Zhang Han, Goodfellow Ian, Metaxas Dimitris, Odena Augustus, 2019, Self-Attention Generative Adversarial Networks, Proceedings of the 36th International Conference on Machine LearningGoogle Search
8 
Mkhzani Alireza, J. Shlens, N. Jaitlyl, I. Goodfeloow, B. Frey, 2015, Adversarial Autoencoder, arXive:1511.05644Google Search
9 
https://steinhardt.nyu.edu/marlGoogle Search
10 
B. Logan, 2000, Mel Frequency Cepstral Coefficients for Music Modeling, The International Sysposium on Music Information Retrieval, Vol. 270, pp. 1-11Google Search

์ €์ž์†Œ๊ฐœ

๊ฐ•์ˆ˜๊ฒฝ(Su-Kyung Kang)
../../Resources/kiee/KIEEP.2020.69.2.69/au1.png

2008๋…„ 8์›” ํ•œ์–‘๋Œ€ ์ „์ž์ „๊ธฐ์ปดํ“จํ„ฐ๊ณตํ•™๋ถ€(ํ•™์‚ฌ)

2008๋…„ 7์›”~ํ˜„์žฌ ํ•œ์ „ ์ „๋ ฅ์—ฐ๊ตฌ์›

<๊ด€์‹ฌ๋ถ„์•ผ> ์ธ๊ณต์ง€๋Šฅ, ์Šค๋งˆํŠธ๊ทธ๋ฆฌ๋“œ, ์˜์ƒ๋ถ„์„

๋ฐ•๋ช…ํ˜œ(Myung-Hye Park)
../../Resources/kiee/KIEEP.2020.69.2.69/au2.png

1993๋…„ 2์›” ๊ฒฝ๋ถ๋Œ€ํ•™๊ต ์ „์ž๊ณตํ•™๊ณผ(ํ•™์‚ฌ)

1995๋…„ 2์›” ๊ฒฝ๋ถ๋Œ€ํ•™๊ต ์ „์ž๊ณตํ•™ (์„์‚ฌ)

1995๋…„~ํ˜„์žฌ ํ•œ์ „ ์ „๋ ฅ์—ฐ๊ตฌ์› ์ฑ…์ž„์—ฐ๊ตฌ์›

<๊ด€์‹ฌ๋ถ„์•ผ> ์ „๋ ฅ์ •๋ณดํ†ต์‹ , ์‚ฌ๋ฌผ์ธํ„ฐ๋„ท, ๋„คํŠธ์›Œํฌ ์‹œ๋ฎฌ๋ ˆ์ด์…˜

๊น€์˜ํ˜„(Young-Hyun Kim)
../../Resources/kiee/KIEEP.2020.69.2.69/au3.png

2002๋…„ 2์›” ํ•œ๊ตญํ•ญ๊ณต๋Œ€ํ•™๊ต ํ†ต์‹ ์ •๋ณด๊ณตํ•™ (ํ•™์‚ฌ)

2004๋…„ 2์›” ๊ด‘์ฃผ๊ณผํ•™๊ธฐ์ˆ ์› ์ •๋ณดํ†ต์‹ ๊ณตํ•™ (์„์‚ฌ)

2004๋…„~ํ˜„์žฌ ํ•œ์ „ ์ „๋ ฅ์—ฐ๊ตฌ์› ์ฑ…์ž„์—ฐ๊ตฌ์›

<๊ด€์‹ฌ๋ถ„์•ผ> ์œ ๏ฝฅ๋ฌด์„ ํ†ต์‹ ์‹œ์Šคํ…œ, ์‚ฌ๋ฌผ์ธํ„ฐ๋„ท

๊น€๋‚™์šฐ(Nac-Woo Kim)
../../Resources/kiee/KIEEP.2020.69.2.69/au4.png

1997๋…„ 2์›” ์ค‘์•™๋Œ€ํ•™๊ต ์ œ์–ด๊ณ„์ธก๊ณตํ•™๊ณผ(ํ•™์‚ฌ)

2002๋…„ 2์›” ์ค‘์•™๋Œ€ํ•™๊ต ์ฒจ๋‹จ์˜์ƒ๋Œ€ํ•™์› ์˜์ƒ๊ณตํ•™๊ณผ (์„์‚ฌ)

2006๋…„ 2์›” ์ค‘์•™๋Œ€ํ•™๊ต ์ฒจ๋‹จ์˜์ƒ๋Œ€ํ•™์› ์˜์ƒ๊ณตํ•™๊ณผ(๋ฐ•์‚ฌ)

2006๋…„ 3์›”~ํ˜„์žฌ ํ•œ๊ตญ์ „์žํ†ต์‹ ์—ฐ๊ตฌ์› ๊ด‘๋Œ€์—ญํ†ตํ•ฉ๋ง ์—ฐ๊ตฌ๋‹จ

<๊ด€์‹ฌ๋ถ„์•ผ> ์˜์ƒ์••์ถ•, ์˜์ƒ์ •๋ณด๊ธฐ์ˆ 

์„œ์ธ์šฉ(In-Yong Seo)
../../Resources/kiee/KIEEP.2020.69.2.69/au5.png

1984๋…„ 2์›” ์„ฑ๊ท ๊ด€๋Œ€ ์ „๊ธฐ๊ณตํ•™์‚ฌ(ํ•™์‚ฌ)

1989๋…„ 8์›” ๋ถ€์‚ฐ๋Œ€ ์ „๊ธฐ์ „์ž๊ณตํ•™(์„์‚ฌ)

2003๋…„ 5์›” ๋ฏธ๊ตญ BROWN๋Œ€ ์ „๊ธฐ๊ณตํ•™๊ณผ(๋ฐ•์‚ฌ)

1990๋…„ 1์›”~2012๋…„ 12์›” ํ•œ์ „ ์ „๋ ฅ์—ฐ๊ตฌ์› ๊ทผ๋ฌด(์ฑ…์ž„์—ฐ๊ตฌ์›)

2013๋…„ 1์›”~2018๋…„ 6์›” ํ•œ์ „ ๊ฒฝ์ œ๊ฒฝ์˜์—ฐ๊ตฌ์›(์ˆ˜์„์—ฐ๊ตฌ์›)

2018๋…„ 7์›”~2020๋…„ 2์›” ํ•œ์ „ ์ „๋ ฅ์—ฐ๊ตฌ์›(์ˆ˜์„์—ฐ๊ตฌ์›)

2020๋…„ 3์›”~ํ˜„์žฌ ์„ฑ๊ท ๊ด€๋Œ€ํ•™๊ต ์ •๋ณดํ†ต์‹ ๋Œ€ํ•™ ์—ฐ๊ตฌ๊ต์ˆ˜

<๊ด€์‹ฌ๋ถ„์•ผ> ์‹ ๋ฐฐ์ „์‹œ์Šคํ…œ, ์Šค๋งˆํŠธ๊ทธ๋ฆฌ๋“œ, ์ธ๊ณต์ง€๋Šฅ