• ๋Œ€ํ•œ์ „๊ธฐํ•™ํšŒ
Mobile QR Code QR CODE : The Transactions of the Korean Institute of Electrical Engineers
  • COPE
  • kcse
  • ํ•œ๊ตญ๊ณผํ•™๊ธฐ์ˆ ๋‹จ์ฒด์ด์—ฐํ•ฉํšŒ
  • ํ•œ๊ตญํ•™์ˆ ์ง€์ธ์šฉ์ƒ‰์ธ
  • Scopus
  • crossref
  • orcid

  1. (Dept. of Electrical Engineering, Soongsil University, Korea. )



transmission line fault data, data augmentation, generative adversarial network, isolation forest, KL-divergence

1. ์„œ ๋ก 

์†ก์ „์„ ๋กœ๋Š” ๋ฐœ์ „์›๊ณผ ์†Œ๋น„์ž๋ฅผ ์ด์–ด์ฃผ๋Š” ์—ญํ• ์„ ๋‹ด๋‹นํ•˜๊ธฐ์—, ์†ก์ „์„ ๋กœ์˜ ํšจ์œจ์ ์ธ ๊ด€๋ฆฌ์™€ ์œ ์ง€๋Š” ์ „๋ ฅ ๊ณต๊ธ‰์˜ ์•ˆ์ •์„ฑ์„ ๋ณด์žฅํ•˜๋Š” ์ค‘์š”ํ•œ ์š”์†Œ์ด๋‹ค. ์†ก์ „์„ ๋กœ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๊ณ ์žฅ์€ ์ „๋ ฅ๊ณ„ํ†ต์˜ ์‹ ๋ขฐ๋„ ์œ ์ง€ ์ธก๋ฉด์—์„œ ์น˜๋ช…์ ์ธ ๋ฌธ์ œ๋ฅผ ์ผ์œผํ‚ฌ ์ˆ˜ ์žˆ๋‹ค. ์†ก์ „์„ ๋กœ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๊ณ ์žฅ์˜ ์›์ธ์œผ๋กœ ์ž์—ฐ์žฌํ•ด, ์™ธ๋ฌผ์ ‘์ด‰, ์„ค๋น„๊ฒฐํ•จ, ๋ณด์ˆ˜๊ฒฐํ•จ ๋“ฑ ๋‹ค์–‘ํ•œ ์š”์ธ์— ์˜ํ•ด ๊ณ ์žฅ์ด ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋‹ค[1]. ์†ก์ „์„ ๋กœ ๊ณ ์žฅ์€ ํ‰ํ˜• ๊ณ ์žฅ๊ณผ ๋ถˆํ‰ํ˜• ๊ณ ์žฅ์œผ๋กœ ๊ตฌ๋ถ„๋˜๋ฉฐ ์ด 10๊ฐ€์ง€ ์œ ํ˜•์œผ๋กœ ๋ถ„๋ฅ˜ํ•  ์ˆ˜ ์žˆ๋‹ค[2]. ํ‰ํ˜• ๊ณ ์žฅ์œผ๋กœ๋Š” 3์ƒ ๋‹จ๋ฝ(Three Phase Fault, LLL) 1๊ฐ€์ง€๊ฐ€ ์žˆ๊ณ , ๋ถˆํ‰ํ˜• ๊ณ ์žฅ์œผ๋กœ 1์„  ์ง€๋ฝ(Single Line to Ground Fault, LG) 3๊ฐ€์ง€, 2์„  ์ง€๋ฝ(Double Line to Ground Fault, LLG) 3๊ฐ€์ง€, ์„ ๊ฐ„ ๋‹จ๋ฝ(Line to Line Fault, LL) 3๊ฐ€์ง€์˜ ์ด 10๊ฐ€์ง€ ๊ณ ์žฅ์œผ๋กœ ๊ตฌ๋ถ„๋œ๋‹ค.

์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋Š” ์ „๋ ฅ ์‹œ์Šคํ…œ ๊ณ ์žฅ ์ง„๋‹จ ๋ฐ ์˜ˆ์ธก ๋ชจ๋ธ ๊ฐœ๋ฐœ ๋“ฑ ๋‹ค์–‘ํ•œ ์—ฐ๊ตฌ ๋ถ„์•ผ์—์„œ ํ™œ์šฉ๋˜๋Š” ์ค‘์š”ํ•œ ๋ฐ์ดํ„ฐ์ด๋‹ค. ๋”ฐ๋ผ์„œ ์ถฉ๋ถ„ํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ํ™•๋ณด๋Š” ์—ฐ๊ตฌ ๋ฐ ์šด์˜ ๋‹จ๊ณ„์—์„œ ์ค‘์š”ํ•˜๋‹ค. ํ•˜์ง€๋งŒ ์‹ค์ œ ์†ก์ „์„ ๋กœ์—์„œ ๊ณ ์žฅ ์‚ฌ๋ก€๊ฐ€ ๋“œ๋ฌผ๊ฒŒ ๋ฐœ์ƒํ•˜๊ณ  ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์ด ๋ถˆ๊ฐ€๋Šฅํ•œ ์ƒํ™ฉ๋„ ์กด์žฌํ•˜๊ธฐ์— ์ถฉ๋ถ„ํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ํ™•๋ณด๊ฐ€ ์–ด๋ ต๋‹ค. ์ด๋Ÿฌํ•œ ์–ด๋ ค์›€ ๋•Œ๋ฌธ์—, ์„ ํ–‰์—ฐ๊ตฌ์—์„œ๋Š” ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ ๋Œ€์‹  ์‹ค์ œ ๊ณ ์žฅ ์ƒํ™ฉ์„ ๋ชจ์˜ํ•  ์ˆ˜ ์žˆ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜์— ์˜์กดํ•˜์—ฌ ํ•„์š”ํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑ[3]ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ๊ณ ์žฅ ์‹œ๋‚˜๋ฆฌ์˜ค ๋ถ„์„์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•˜์˜€๋‹ค[4].

ํ•˜์ง€๋งŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์€ ๊ณ„ํ†ต์„ ์™„๋ฒฝํžˆ ๋ชจ์‚ฌํ•  ์ˆ˜ ์—†๊ธฐ์— ์‹œ๋ฎฌ๋ ˆ์ด์…˜์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋Š” ์‹ค์ œ ์†ก์ „์„ ๋กœ์—์„œ ๋ฐœ์ƒํ•˜๋Š” ๊ณ ์žฅ๋ฐ์ดํ„ฐ์™€ ์ฐจ์ด๊ฐ€ ์กด์žฌํ•œ๋‹ค[5~6]. ์ด๋Ÿฌํ•œ ๋ฐ์ดํ„ฐ์˜ ์ฐจ์ด๋กœ ์ธํ•ด ์‹ค๊ณ„ํ†ต์— ์ ์šฉ ๊ฐ€๋Šฅํ•œ ๊ณ ์žฅ ์ง„๋‹จ ๋ฐ ์˜ˆ์ธก๋ชจ๋ธ ๊ฐœ๋ฐœ๊ณผ ๊ฐ™์€ ์—ฐ๊ตฌ์— ํ•œ๊ณ„๊ฐ€ ์กด์žฌํ•œ๋‹ค[7~8].

์ด๋Ÿฌํ•œ ๋ฌธ์ œ์— ๋Œ€์‘ํ•˜๊ธฐ ์œ„ํ•ด, ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์— ์‹ค์ œ ํ™˜๊ฒฝ ํŠน์„ฑ์„ ๋ชจ์‚ฌํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ธฐ๋ฒ•์œผ๋กœ ์žก์Œ์„ ์ถ”๊ฐ€ํ•˜๋Š” ์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜์—ˆ๋‹ค[9~10]. ์ด ์—ฐ๊ตฌ์—์„œ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์— ์ธ์œ„์ ์ธ ์žก์Œ์„ ์ถ”๊ฐ€ํ–ˆ์„ ๋•Œ ์ œ์•ˆํ•˜๋Š” ๋ถ„๋ฅ˜๊ธฐ๋ฒ•์ด ์–ผ๋งˆ๋‚˜ ๊ฐ•์ธํ•œ์ง€ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. ํ•˜์ง€๋งŒ ์žก์Œ์„ ์ถ”๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์€ ์‹ค์ œ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ฐ€์ง„ ๋ณต์žกํ•œ ํŠน์„ฑ๊ณผ ํŒจํ„ด์„ ์ถฉ๋ถ„ํžˆ ๋ฐ˜์˜ํ•˜๊ธฐ ์–ด๋ ต๊ธฐ์— ์žก์Œ์„ ์ถ”๊ฐ€ํ•˜๋Š” ์„ ํ–‰์—ฐ๊ตฌ๋Š” ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ ์ธก๋ฉด์—์„œ ํ•œ๊ณ„์ ์ด ์กด์žฌํ•œ๋‹ค.

์ตœ๊ทผ ์—ฐ๊ตฌ[11~12]์—์„œ ์‹ค๊ณ„ํ†ต ํ™˜๊ฒฝ์„ ๋ฐ˜์˜ํ•˜๋ฉฐ, ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์„ ์ฆ๊ฐ€์‹œํ‚ค๊ธฐ ์œ„ํ•œ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ธฐ๋ฒ•์œผ๋กœ ์ƒ์„ฑ์ ์ ๋Œ€์‹ ๊ฒฝ๋ง(Generative Adversarial Networks, GAN)์„ ์ด์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•์ด ์—ฐ๊ตฌ๋˜์—ˆ๋‹ค. [11~12]์—์„œ๋Š” ๊ณ ์ €ํ•ญ ๊ณ ์žฅ(High Impedence Fault, HIF) ๊ฐ์ง€ ๋ฐฉ๋ฒ•์—์„œ ์ •์ƒ์ƒํƒœ ๋ฐ์ดํ„ฐ์™€ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ๊ฐ„์˜ ๋ถ„ํฌ ๋ถˆ๊ท ํ˜• ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด GAN์„ ์ ์šฉํ•˜์˜€๋‹ค. ์„ ํ–‰์—ฐ๊ตฌ์—์„œ GAN์„ ํ†ตํ•ด ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์™€ ์œ ์‚ฌํ•œ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•จ์œผ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ์˜ ํ’ˆ์งˆ๊ณผ ๋‹ค์–‘์„ฑ์„ ์ฆ์ง„์‹œํ‚ค๋Š” ๊ฒƒ์„ ๋ณด์˜€๋‹ค.

๊ธฐ์กด GAN์„ ์ด์šฉํ•œ ์—ฐ๊ตฌ๋Š” ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•œ ํ‰๊ฐ€๋ฐฉ๋ฒ•์ด ๋ชจํ˜ธํ•˜๋‹ค๋Š” ํ•œ๊ณ„์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค[13]. GAN์€ ํ•™์Šตํ•œ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ํ†ตํ•ด ์‹ค์ œ์™€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜์ง€๋งŒ, ๊ทธ ํ’ˆ์งˆ์„ ์ •๋Ÿ‰์ ์œผ๋กœ ์ธก์ •ํ•˜๊ณ  ํ‰๊ฐ€ํ•˜๋Š” ๊ฒƒ์ด ์–ด๋ ต๋‹ค. ์ผ๋ฐ˜์ ์œผ๋กœ GAN์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ํ”„๋ ˆ์ณ‡ ์ธ์…‰์…˜ ๊ฑฐ๋ฆฌ(Frรฉchet Inception Distance, FID) ์ ์ˆ˜[14], ์ธ์…‰์…˜ ์ ์ˆ˜(Inception Score, IS)[15]์™€ ๊ฐ™์€ ํ‰๊ฐ€์ง€ํ‘œ๊ฐ€ ์‚ฌ์šฉ๋œ๋‹ค. ๋‘ ์ง€ํ‘œ ๋ชจ๋‘ ์‚ฌ์ „ํ›ˆ๋ จ๋œ ์ธ์…‰์…˜ V3 ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ๋‹ค. ํ•˜์ง€๋งŒ ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ๊ฐ€ ๋ถ€์กฑํ•œ ์ƒํ™ฉ์—์„œ ๋ฐ์ดํ„ฐ์˜ ํŠน์ง•์„ ์ž˜ ๋ฐ˜์˜ํ•œ ์ธ์…‰์…˜ V3 ๋ชจ๋ธ์„ ํ›ˆ๋ จ์‹œํ‚ค๋Š” ๊ฒƒ์€ ํ˜„์‹ค์ ์œผ๋กœ ์–ด๋ ต๋‹ค. ์ถ”๊ฐ€์ ์œผ๋กœ FID์™€ IS์˜ ์ฃผ๋กœ ์ด๋ฏธ์ง€ ๋ฐ์ดํ„ฐ์…‹์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•ด ๊ฐœ๋ฐœ๋˜์—ˆ๊ธฐ์— ๊ณ ์žฅ๊ณผ ๊ฐ™์€ ์‹œ๊ณ„์—ด์ ์ธ ํŠน์ง•์ด ๋ฐ˜์˜๋œ ๋ฐ์ดํ„ฐ์…‹์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ์ง€ํ‘œ๋กœ๋Š” ์ ์ ˆํ•˜์ง€ ์•Š์„ ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ํŠน์ง•์— ๋งž๋Š” ํ‰๊ฐ€์ง€ํ‘œ๊ฐ€ ํ•„์š”ํ•˜๋‹ค.

๋ณธ ์—ฐ๊ตฌ๋Š” ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ํ•„์š”์„ฑ๊ณผ ๊ธฐ์กด ๋ฐฉ๋ฒ•์˜ ํ•œ๊ณ„์ ์„ ๊ณ ๋ คํ•˜๊ณ , ์‚ฌ์ „์—ฐ๊ตฌ[16]์„ ํ™•์žฅํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ GAN ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค. ์ œ์•ˆํ•˜๋Š” ๊ธฐ๋ฒ•์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์œผ๋กœ ์ƒ์„ฑํ•œ ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šต๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉํ•˜์˜€์œผ๋ฉฐ, ์‹œ๊ณ„์—ด ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์— ์ ํ•ฉํ•œ WGAN-GP(Wasserstein Generative Adversarial Network with Gradient Penalty) ๋ชจ๋ธ์„ ์ƒ์„ฑ๊ธฐ๋ฒ•์œผ๋กœ ์‚ฌ์šฉํ•œ๋‹ค[17]. ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ํ†ตํ•ด ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•œ ํ›„ ๋ฐ์ดํ„ฐ ๊ฒ€์ฆ์„ ์œ„ํ•ด IF(Isolation Forest) ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ์ €ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•œ๋‹ค. ๋งˆ์ง€๋ง‰์œผ๋กœ ์ œ์•ˆํ•˜๋Š” WGAN-GP ๋ชจ๋ธ์ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์…‹์„ KL-divergence(Kullback-Leibler divergence) ๊ฐ’์„ ํ†ตํ•ด ๋ฐ์ดํ„ฐ ๋ถ„ํฌ ์œ ์‚ฌ๋„๋ฅผ ํ‰๊ฐ€ํ•œ๋‹ค. ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๊ฐ€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ๋ฅผ ์ž˜ ๋”ฐ๋ฅด๋ฉด์„œ ๋‹ค์–‘์„ฑ์„ ์œ ์ง€ํ•˜๊ณ  ์žˆ์Œ์„ ์‹œ๊ฐ์ ์œผ๋กœ ํ™•์ธํ•˜์˜€๋‹ค. ๋˜ํ•œ, ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์œผ๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹๋ณด๋‹ค ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์…‹๊ณผ ๋” ์œ ์‚ฌํ•œ ๋ถ„ํฌ๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Œ์„ KL-divergence ๊ฐ’ ๋น„๊ต๋ฅผ ํ†ตํ•ด ํ™•์ธํ•˜์˜€๋‹ค.

2. ์ด๋ก ์  ๋ฐฐ๊ฒฝ

2.1 GAN

GAN[18]์€ ๋”ฅ๋Ÿฌ๋‹์˜ ํ•œ ๋ถ„์•ผ๋กœ์„œ, ๋‘ ๋„คํŠธ์›Œํฌ๊ฐ€ ์„œ๋กœ ๊ฒฝ์Ÿํ•˜๋ฉด์„œ ๋™์‹œ์— ํ•™์Šตํ•˜๋Š” ๊ตฌ์กฐ๋ฅผ ๊ฐ€์ง„ ๋ชจ๋ธ์ด๋‹ค. GAN์€ ๊ทธ๋ฆผ 1๊ณผ ๊ฐ™์ด ํฌ๊ฒŒ ์ƒ์„ฑ์ž(Generator), ํŒ๋ณ„์ž(Discriminator)์˜ ๋‘ ๋ถ€๋ถ„์œผ๋กœ ๊ตฌ๋ถ„ํ•œ๋‹ค. ์ƒ์„ฑ์ž๋Š” ๋ฌด์ž‘์œ„ ์žก์Œ์„ ์ž…๋ ฅ๋ฐ›์•„ ์‹ค์ œ์™€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ์—ญํ• ์„ ํ•˜๊ณ  ํŒ๋ณ„์ž๋Š” ์ž…๋ ฅ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ์‹ค์ œ ๋ฐ์ดํ„ฐ์ธ์ง€ ์ƒ์„ฑ์ž๊ฐ€ ๋งŒ๋“  ๊ฐ€์งœ ๋ฐ์ดํ„ฐ์ธ์ง€๋ฅผ ํŒ๋ณ„ํ•˜๋Š” ์—ญํ• ์„ ํ•œ๋‹ค. ์ƒ์„ฑ์ž์™€ ํŒ๋ณ„์ž๋Š” ์„œ๋กœ ๋ณด์™„ํ•ด๊ฐ€๋ฉด์„œ ์ ๋Œ€์ ์œผ๋กœ ํ›ˆ๋ จ๋˜์–ด ์„ฑ๋Šฅ์ด ๊ณ ๋„ํ™”๋œ๋‹ค. GAN ๋ชจ๋ธ์€ ๋‹ค์Œ ์ˆ˜์‹ (1)์„ ๋ชฉ์ ํ•จ์ˆ˜๋กœ ์‚ฌ์šฉํ•˜์—ฌ ํ•™์Šตํ•œ๋‹ค.

(1)
../../Resources/kiee/KIEE.2024.73.8.1318/eq1.png

$x$๋Š” ํ•™์Šต๋ฐ์ดํ„ฐ ๋ถ„ํฌ($P_{data}$)์—์„œ ๊ฐ€์ ธ์˜จ ์‹ค์ œ ๋ฐ์ดํ„ฐ์ด๊ณ  $z$๋Š” ์ƒ์„ฑ์ž์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ์žก์Œ ๋ฒกํ„ฐ์ด๋‹ค. $z$๋Š” ์‚ฌ์ „์— ์ •์˜๋œ ์žก์Œ ๋ถ„ํฌ $p_{z}$์—์„œ ๋ฌด์ž‘์œ„๋กœ ์ถ”์ถœ๋œ๋‹ค. $G(z)$๋Š” ์ƒ์„ฑ์ž $G$์— ์˜ํ•ด ์ƒ์„ฑ๋œ ๊ฐ€์งœ ๋ฐ์ดํ„ฐ์ด๊ณ , $D(x)$๋Š” ํŒ๋ณ„์ž $D$๊ฐ€ ๋ฐ์ดํ„ฐ $x$๋ฅผ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋กœ ํŒ๋ณ„ํ•  ํ™•๋ฅ ์ด๋‹ค. $D(G(z))$๋Š” ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ $G(z)$๊ฐ€ ์‹ค์ œ ๋ฐ์ดํ„ฐ๋กœ ํŒ๋ณ„๋  ํ™•๋ฅ ์ด๋‹ค. ํ•™์Šต์ด ์ง„ํ–‰๋จ์— ๋”ฐ๋ผ ์ƒ์„ฑ์ž $G$๋Š” ์ ์  ๋” ์ง„์งœ ๊ฐ™์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ณ , ํŒ๋ณ„์ž๋Š” ์ง„์งœ ๋ฐ์ดํ„ฐ์™€ ๊ฐ€์งœ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ๋ถ„ํ•˜๋Š” ๋Šฅ๋ ฅ์ด ๊ฐœ์„ ๋œ๋‹ค. $G$๊ฐ€ ์ด์ƒ์ ์œผ๋กœ ํ•™์Šต์ด ๋˜์—ˆ์„ ๋•Œ $D(x)= 0.5$์˜ ๊ฐ’์„ ๊ฐ–๊ฒŒ ๋˜๋ฉฐ, ํŒ๋ณ„์ž $D$๋Š” ์‹ค์ œ ๋ฐ์ดํ„ฐ์™€ ๊ฐ€์งœ ๋ฐ์ดํ„ฐ๋ฅผ ๊ตฌ๋ถ„ํ•  ์ˆ˜ ์—†์„ ์ •๋„๋กœ ์ƒ์„ฑ์ž $G$๊ฐ€ ์ง„์งœ ๊ฐ™์€ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ฒŒ ๋œ๋‹ค.

๊ทธ๋ฆผ 1. GAN ๋ชจ๋ธ ๊ตฌ์กฐ

Fig. 1. Architecture of the GAN

../../Resources/kiee/KIEE.2024.73.8.1318/fig1.png

2.2 WGAN-GP

WGAN-GP[19]๋Š” ๊ธฐ์กด์˜ GAN์˜ ๋ฌธ์ œ๋ฅผ ๊ฐœ์„ ํ•˜๊ธฐ ์œ„ํ•ด ์ œ์•ˆ๋œ ๋ชจ๋ธ์ด๋‹ค. ๋Œ€ํ‘œ์ ์ธ GAN์˜ ๋ฌธ์ œ์ ์œผ๋กœ ์ƒ์„ฑ์ž $G$๊ฐ€ ํ•™์Šต๋ฐ์ดํ„ฐ ๋ถ„ํฌ์˜ ๋‹ค์–‘์„ฑ์„ ์ถฉ๋ถ„ํžˆ ๋ฐ˜์˜ํ•˜์ง€ ๋ชปํ•˜๊ณ  ํŠน์ • ๋ฒ”์œ„์˜ ํ•™์Šต๋ฐ์ดํ„ฐ๋งŒ์„ ๋ฐ˜๋ณต์ ์œผ๋กœ ์ƒ์„ฑํ•˜๋Š” ํ˜„์ƒ์ด ์žˆ๋‹ค. ์ด๋กœ ์ธํ•ด ์ƒ์„ฑ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์ด ์†์‹ค๋˜๋Š” ์ƒํ™ฉ์ด ๋ฐœ์ƒํ•œ๋‹ค. WGAN-GP๋Š” Wassertein ๊ฑฐ๋ฆฌ๋ฅผ ๋ชฉ์ ํ•จ์ˆ˜๋กœ ์‚ฌ์šฉํ•˜๊ณ  ๊ฒฝ์‚ฌ ํŽ˜๋„ํ‹ฐ(gradient penalty) ํ•ญ์„ ์ถ”๊ฐ€ํ•จ์œผ๋กœ์จ, ํ•™์Šต ๊ณผ์ •์—์„œ ๋ฐœ์ƒํ•  ์ˆ˜ ์žˆ๋Š” ๋ฌธ์ œ์ ์„ ํ•ด๊ฒฐํ•œ๋‹ค. WGAN์˜ ๋ชฉ์ ํ•จ์ˆ˜์ธ Wassertein ๊ฑฐ๋ฆฌ๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ์…‹๊ณผ ์ƒ์„ฑ ๋ฐ์ดํ„ฐ์…‹ ์‚ฌ์ด์˜ ํ™•๋ฅ ๋ถ„ํฌ ๊ฐ„์˜ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•˜๋Š”๋ฐ ์‚ฌ์šฉ๋˜๋ฉฐ Wassertein ๊ฑฐ๋ฆฌ๋Š” ๋‹ค์Œ ์ˆ˜์‹ (2)๋กœ ํ‘œํ˜„๋œ๋‹ค.

(2)
../../Resources/kiee/KIEE.2024.73.8.1318/eq2.png

์—ฌ๊ธฐ์„œ $P_{g}$๋Š” ์ƒ์„ฑ์ž $G$๊ฐ€ ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์˜ ๋ถ„ํฌ, $\underline{\dfrac{}{}}"{sup}"_{vert vert f vert vert_{L}\le K}$๋Š” ๋ชจ๋“  Lipschitz ์ƒ์ˆ˜ $K$์— ๋Œ€ํ•ด ์ œํ•œ๋œ ํ•จ์ˆ˜ $f$์— ๋Œ€ํ•œ ์ตœ๋Œ“๊ฐ’, $f(x)$๋Š” ํŒ๋ณ„์ž๊ฐ€ ๋ฐ์ดํ„ฐ $x$๋ฅผ ๋ฐ์ดํ„ฐ๋ฅผ ์–ด๋–ป๊ฒŒ ๊ตฌ๋ถ„ํ• ์ง€ ํ•™์Šตํ•œ ํ•จ์ˆ˜์ด๋‹ค.

WGAN-GP๋Š” Wassertein ๊ฑฐ๋ฆฌ์˜ Lipschitz ์ œ์•ฝ์„ ๋งŒ์กฑ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฒฝ์‚ฌ ํŽ˜๋„ํ‹ฐ ํ•ญ์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์‚ฌ์šฉํ•˜๊ณ , ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๋ชฉ์ ํ•จ์ˆ˜๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค.

(3)
../../Resources/kiee/KIEE.2024.73.8.1318/eq3.png

์—ฌ๊ธฐ์„œ $\lambda$๋Š” ๊ฒฝ์‚ฌ ํŒจ๋„ํ‹ฐ ํ•ญ์˜ ๊ฐ€์ค‘์น˜ํ•ญ, $\hat{x}$๋Š” ํ•™์Šต๋ฐ์ดํ„ฐ์™€ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์‚ฌ์ด์˜ ๋ณด๊ฐ„๋œ ๋ฐ์ดํ„ฐ, $\nabla_{\hat{x}}D(\hat{x})$๋Š” $\hat{x}$์—์„œ ํŒ๋ณ„์ž $D$์˜ ๊ธฐ์šธ๊ธฐ๋ฅผ ์˜๋ฏธํ•œ๋‹ค. ์ด๋ ‡๊ฒŒ ์ถ”๊ฐ€๋œ ๊ฒฝ์‚ฌ ํŽ˜๋„ํ‹ฐ ํ•ญ์€ ํŒ๋ณ„์ž์˜ ๊ธฐ์šธ๊ธฐ๊ฐ€ Lipschitz ์—ฐ์† ์กฐ๊ฑด์„ ๋งŒ์กฑํ•˜๋„๋ก ๊ฐ•์ œํ•จ์œผ๋กœ์จ WGAN์˜ ํ•™์Šต์˜ ์•ˆ์ •์„ฑ๊ณผ ์„ฑ๋Šฅ์„ ๊ฐœ์„ ํ•œ๋‹ค.

2.3 Isolation Forest

Isolation Forest(IF)[20] ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์ฃผ๋กœ ์ด์ƒ์น˜(outlier) ํƒ์ง€์— ์‚ฌ์šฉ๋˜๋ฉฐ, ์ด์ƒ์น˜๊ฐ€ ์ •์ƒ๋ฐ์ดํ„ฐ์— ๋น„ํ•ด ์ƒ๋Œ€์ ์œผ๋กœ ๊ณ ๋ฆฝ์‹œํ‚ค๊ธฐ ์‰ฌ์šธ ๊ฒƒ์ด๋ผ๋Š” ์•„์ด๋””์–ด๋ฅผ ์ ์šฉํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด๋‹ค. IF๋Š” ๊ฒฐ์ • ํŠธ๋ฆฌ(Decision Tree)๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๋ฅผ ๊ณ ๋ฆฝ์‹œํ‚ค๋Š” ๊ณผ์ •์œผ๋กœ ๋™์ž‘ํ•œ๋‹ค. ํŠธ๋ฆฌ์˜ ๊ฐ ๋…ธ๋“œ์—์„œ, ๋ฌด์ž‘์œ„๋กœ ์„ ํƒ๋œ ํŠน์„ฑ์— ๋Œ€ํ•œ ๋ฌด์ž‘์œ„ ์ž„๊ณ—๊ฐ’์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ ๋‘ ๊ฐœ์˜ ๊ทธ๋ฃน์œผ๋กœ ๋ถ„ํ•  ํ•œ๋‹ค. ์ด ๊ณผ์ •์€ ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๊ฐ€ ํ•˜๋‚˜๋งŒ ๋‚จ์„ ๋•Œ๊นŒ์ง€, ๋˜๋Š” ํŠธ๋ฆฌ๊ฐ€ ์‚ฌ์ „์— ์ •์˜๋œ ์ตœ๋Œ€ ๊นŠ์ด์— ๋„๋‹ฌํ•  ๋•Œ๊นŒ์ง€ ๋ฐ˜๋ณต๋œ๋‹ค. ๋ถ„ํ•  ๊ณผ์ •์„ ๊ฑฐ์น˜๋ฉด์„œ ๊ฐ๊ฐ์˜ ๋ฐ์ดํ„ฐ์— ๋”ฐ๋ฅธ ์ด์ƒ์น˜ ์ ์ˆ˜(anomaly score)๋ฅผ ๊ณ„์‚ฐํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด์ƒ์น˜ ์ ์ˆ˜๋Š” ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ๊ฐ€ ํŠธ๋ฆฌ๋ฅผ ํ†ต๊ณผํ•˜๋Š”๋ฐ ํ•„์š”ํ•œ ๋ถ„ํ•  ํšŸ์ˆ˜๋ฅผ ํ†ตํ•ด ๊ณ„์‚ฐ๋œ๋‹ค. ์ด์ƒ์น˜๋Š” ์ ์€ ์ˆ˜์˜ ๋ถ„ํ• ๋กœ ๊ณ ๋ฆฝ๋˜๋ฏ€๋กœ ๋ถ„ํ•  ํšŸ์ˆ˜๊ฐ€ ์ ์„์ˆ˜๋ก ์ด์ƒ์น˜ ์ ์ˆ˜๊ฐ€ ๋†’๋‹ค. ๊ธฐ์กด ๊ฑฐ๋ฆฌ๊ธฐ๋ฐ˜ ๋˜๋Š” ๋ฐ€๋„๊ธฐ๋ฐ˜์˜ ์ด์ƒ์น˜ ํƒ์ง€ ๋ฐฉ๋ฒ•๋ก ๋“ค๊ณผ ๋‹ค๋ฅด๊ฒŒ ๊ณ„์‚ฐ๋Ÿ‰์ด ์ ์–ด ์‹œ๊ฐ„ ๋ณต์žก๋„๊ฐ€ ๋‚ฎ๋‹ค๋Š” ์ ๊ณผ ๊ณ ์ฐจ์› ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•ด์„œ๋„ ์„ฑ๋Šฅ์ด ๋ณด์žฅ๋˜๋Š” ๊ฒƒ์ด ์žฅ์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด๋‹ค.

2.4 KL-divergence

KL-divergence๋Š” ๋‘ ํ™•๋ฅ ๋ถ„ํฌ ๊ฐ„์˜ ์œ ์‚ฌ๋„๋ฅผ ์ธก์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค. KL-divergence๋Š” ๋ฐ์ดํ„ฐ์…‹ P์˜ ๋ถ„ํฌ๋ฅผ ๊ธฐ์ค€์œผ๋กœ ๋ฐ์ดํ„ฐ์…‹ Q์˜ ๋ถ„ํฌ๊ฐ€ P๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ๊ทผ์‚ฌํ•˜๋Š”์ง€๋ฅผ ์ธก์ •ํ•˜๋Š” ์ง€ํ‘œ๋กœ ์‚ฌ์šฉ๋œ๋‹ค. ๋‘ ์ด์‚ฐํ™•๋ฅ ๋ถ„ํฌ P์™€ Q์— ๋Œ€ํ•ด, P์—์„œ Q๋กœ์˜ KL-divergence๋Š” ์ˆ˜์‹ (4)์œผ๋กœ ๊ณ„์‚ฐ๋œ๋‹ค.

(4)
../../Resources/kiee/KIEE.2024.73.8.1318/eq4.png

$P(x)$, $Q(x)$๋Š” ๊ฐ๊ฐ ํ™•๋ฅ ๋ถ„ํฌ $P$์™€ $Q$์—์„œ ์‚ฌ๊ฑด $x$๊ฐ€ ๋ฐœ์ƒํ•  ํ™•๋ฅ ์„ ๋‚˜ํƒ€๋‚ธ๋‹ค. $\chi$๋Š” ๋ชจ๋“  ๊ฐ€๋Šฅํ•œ ์‚ฌ๊ฑด์˜ ์ง‘ํ•ฉ์ด๋‹ค. ๊ณ„์‚ฐ๋œ ๊ฐ’์€ ํ•ญ์ƒ 0 ๋˜๋Š” ์–‘์˜ ๊ฐ’์„ ๊ฐ€์ง„๋‹ค. KL-divergence ๊ฐ’์ด ์ž‘์„์ˆ˜๋ก ๋‘ ๋ถ„ํฌ๊ฐ€ ๋น„์Šทํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•˜๋ฉฐ, ๋‘ ํ™•๋ฅ ๋ถ„ํฌ๊ฐ€ ์™„์ „ํžˆ ๋™์ผํ•˜๋‹ค๋ฉด 0์˜ ๊ฐ’์„ ๊ฐ–๋Š”๋‹ค.

3. ๋‹ค์–‘ํ•œ ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ GAN ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ธฐ๋ฒ•

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ GAN ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•œ๋‹ค. ๊ทธ๋ฆผ 2๋Š” ์ œ์•ˆํ•˜๋Š” GAN ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ธฐ๋ฒ•์˜ ๊ฐœ๋…๋„์ด๋‹ค. ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์€ 1) GAN ๋ชจ๋ธ ํ›ˆ๋ จ ๋ถ€๋ถ„(3์žฅ 1์ ˆ)๊ณผ 2) ์ด์ƒ์น˜๋ฅผ ์ œ๊ฑฐํ•˜์—ฌ ์ตœ์ข… ๋ฐ์ดํ„ฐ์…‹์„ ์„ ์ •(3์žฅ 2์ ˆ)์˜ ์ด 2๋‹จ๊ณ„๋กœ ๊ตฌ๋ถ„๋œ๋‹ค.

๊ทธ๋ฆผ 2. ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ณผ์ • ์ „์ฒด๊ฐœ๋…๋„

Fig. 2. Overall framework of the data augmentation

../../Resources/kiee/KIEE.2024.73.8.1318/fig2.png

3.1 GAN ๋ชจ๋ธ ํ›ˆ๋ จ

GAN ๋ชจ๋ธ ํ›ˆ๋ จ์— ํ•„์š”ํ•œ ํ•™์Šต๋ฐ์ดํ„ฐ๋Š” ์†ก์ „์„ ๋กœ ๊ณ ์žฅ์„ ๋ชจ์˜ํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์„ ํ†ตํ•ด ์ƒ์„ฑํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•œ๋‹ค. ์‹œ๋ฎฌ๋ ˆ์ด์…˜์œผ๋กœ ์ƒ์„ฑํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋Š” GAN ๋ชจ๋ธ์˜ ํŒ๋ณ„์ž $D$์˜ ํ•™์Šต๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉ๋˜๊ธฐ ๋•Œ๋ฌธ์— ์‹ค๊ณ„ํ†ต ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์ตœ๋Œ€ํ•œ ๋ฐ˜์˜ํ•˜๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹ค. ๋˜ํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์„ ์œ„ํ•ด ๊ณ ์žฅ์œ ํ˜•, ์„ ๋กœ๊ธธ์ด, ์œ„์ƒ ๊ฐ ๋“ฑ ๋‹ค์–‘ํ•œ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์กฐ์ ˆํ•˜๋ฉฐ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•œ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์— ์ ํ•ฉํ•œ WGAN-GP ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. WGAN-GP ๋ชจ๋ธ์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์œผ๋กœ ์ƒ์„ฑํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ ๊ฐ๊ฐ ํ›ˆ๋ จ์‹œํ‚จ๋‹ค.

3.2 ์ด์ƒ์น˜ ์ œ๊ฑฐ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•œ ์ตœ์  ๋ฐ์ดํ„ฐ์…‹ ์„ ์ •

๋ณธ ๋‹จ๊ณ„๋Š” ๋‘ ๋‹จ๊ณ„์˜ ์„ธ๋ถ€ ๋‹จ๊ณ„๋กœ ๊ตฌ๋ถ„๋˜์–ด ์žˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„(3.2.1)์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์ด์ƒ์น˜ ๋น„์œจ์„ ์ ์šฉํ•˜์—ฌ ๋ณต์ˆ˜ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•œ๋‹ค. ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„(3.2.2)๋Š” ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์ด ํ›ˆ๋ จ๋ฐ์ดํ„ฐ ๋ถ„ํฌ๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ํ•™์Šตํ–ˆ๋Š”์ง€๋ฅผ ํ‰๊ฐ€ํ•˜๋Š” ๋‹จ๊ณ„์ด๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” KL-divergence ๊ฐ’์„ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํ‰๊ฐ€ํ•˜์—ฌ ์ตœ์ ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ์„ ์ •ํ•œ๋‹ค.

3.2.1 ์ด์ƒ์น˜ ๋น„์œจ์— ๋”ฐ๋ฅธ ๋ฐ์ดํ„ฐ์…‹ ์ƒ์„ฑ

ํ›ˆ๋ จ๋œ ์ƒ์„ฑ์ž $G$์—์„œ ํ™•๋ฅ ์ ์œผ๋กœ ์ƒ์„ฑ๋˜๋Š” ๋ฐ์ดํ„ฐ๋“ค ์ค‘ ์ด์ƒ ๋ฐ์ดํ„ฐ๋“ค์€ ์ „์ฒด ๋ฐ์ดํ„ฐ์…‹์˜ ํ’ˆ์งˆ์„ ์ €ํ•˜์‹œ์ผœ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์˜ ํšจ๊ณผ๋ฅผ ๊ฐ์†Œ์‹œํ‚จ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋Ÿฌํ•œ ์ €ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๋“ค์€ ์ œ๊ฑฐํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ด์ƒ์น˜ ์ œ๊ฑฐ๋ฅผ ์œ„ํ•ด IF(Isolation Forest) ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๊ตฌ์ฒด์ ์œผ๋กœ IF ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ†ตํ•ด ๊ฐ ๋ฐ์ดํ„ฐ์˜ ์ด์ƒ ์ ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•˜๊ณ , ๊ฐ€์žฅ ๋†’์€ ์ ์ˆ˜๋ฅผ ๊ฐ€์ง„ ๋ฐ์ดํ„ฐ๋“ค์„ ์ œ๊ฑฐํ•œ๋‹ค. IF ์•Œ๊ณ ๋ฆฌ์ฆ˜์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๋ณ€์ˆ˜๋Š” ์˜ค์—ผ๋„(Contamination, $C$) ์ด๋‹ค. ์˜ค์—ผ๋„ $C$์— ๋”ฐ๋ผ ์ „์ฒด ๋ฐ์ดํ„ฐ์…‹์—์„œ ์ œ๊ฑฐ๋  ๋ฐ์ดํ„ฐ๋“ค์˜ ๋น„์œจ์„ ๊ฒฐ์ •ํ•  ์ˆ˜ ์žˆ๊ณ , $C$์˜ ๊ฐ’์— ๋”ฐ๋ผ ์„œ๋กœ ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, ์ „์ฒด ๋ฐ์ดํ„ฐ์…‹์„ ๊ธฐ์ค€์œผ๋กœ 5%๊ฐ€ ์ด์ƒ์น˜๋ผ๊ณ  ๊ฐ€์ •ํ•˜๋ฉด ์˜ค์—ผ๋„ $C$๋Š” 0.05๊ฐ€ ๋œ๋‹ค.

3.2.2 KL-divergence๋ฅผ ํ†ตํ•œ ์ตœ์  ๋ฐ์ดํ„ฐ์…‹ ์„ ์ •

์˜ค์—ผ๋„ $C$์— ๋”ฐ๋ผ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์ด ํ•™์Šต ๋ฐ์ดํ„ฐ์…‹์˜ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ์–ผ๋งˆ๋‚˜ ์ž˜ ๋ฐ˜์˜ํ•˜๋Š”์ง€๋ฅผ ์ •๋Ÿ‰์ ์œผ๋กœ ํ‰๊ฐ€ํ•˜๊ณ , ์ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฐ์ดํ„ฐ์…‹์˜ ํ’ˆ์งˆ์„ ๊ฒฐ์ •ํ•˜์—ฌ ์ตœ์  ๋ฐ์ดํ„ฐ์…‹์„ ์„ ์ •ํ•œ๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๊ณ ์žฅํŒŒํ˜• ๋ฐ์ดํ„ฐ์…‹์˜ ํŠน์„ฑ์„ ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•ด ์‹คํšจ๊ฐ’(Root Mean Square, RMS)์— ๊ธฐ๋ฐ˜ํ•œ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹๊ณผ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹ ๊ฐ„์˜ KL-divergence ๊ฐ’์„ ๊ณ„์‚ฐํ•œ๋‹ค.

์œ„์˜ 3.2.1๊ณผ 3.2.2์˜ ๊ณผ์ •์— ๋Œ€ํ•œ ์ž์„ธํ•œ ๊ณ„์‚ฐ ๊ณผ์ •์€ 1)~4)๋กœ ์„ค๋ช…ํ•˜๋ฉฐ ๋ชจ๋“  ๊ณ ์žฅ ์œ ํ˜•์— ๋Œ€ํ•ด 1)~4) ๊ณผ์ •์„ ๋ฐ˜๋ณตํ•œ๋‹ค.

1) ์˜ค์—ผ๋„ $C$์— ๋”ฐ๋ผ $M$๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•œ๋‹ค.

2) ๊ฐ๊ฐ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ์ˆ˜์‹ (5)๋ฅผ ํ†ตํ•ด RMS๋กœ ๋ณ€ํ™˜ํ•˜๊ณ  ๊ฐ ์ƒ(A์ƒ, B์ƒ, C์ƒ)๋ณ„ RMS์— ๋Œ€ํ•œ ์ด์‚ฐ ํ™•๋ฅ ๋ฐ€๋„ํ•จ์ˆ˜๋ฅผ ๊ณ„์‚ฐํ•œ๋‹ค. ์ด๋•Œ $N$์€ ์ฃผ๊ธฐํ•จ์ˆ˜ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ์ค€์œผ๋กœ 1์ฃผ๊ธฐ์— ํ•ด๋‹นํ•˜๋Š” ๋ฐ์ดํ„ฐ ๊ฐœ์ˆ˜๋ฅผ ์˜๋ฏธํ•˜๋ฉฐ, $l$์€ ์ด ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ๊ฐœ์ˆ˜์ด๋‹ค.

(5)
../../Resources/kiee/KIEE.2024.73.8.1318/eq5.png

3) ์ด $M$๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹ ๋ณ„๋กœ ์ˆ˜์‹ (6)๋ฅผ ํ†ตํ•ด KL-divergence ๊ฐ’์„ ๊ณ„์‚ฐํ•œ๋‹ค. ์—ฌ๊ธฐ์„œ P์™€ Q๋Š” ๊ฐ๊ฐ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹๊ณผ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹ ๋ถ„ํฌ์˜ ํ™•๋ฅ ๋ฐ€๋„ํ•จ์ˆ˜์ด๋‹ค. KL-divergence ๊ฐ’์€ ์ƒ๋ณ„๋กœ ๊ณ„์‚ฐ๋˜๋ฉฐ ์ตœ์ข… KL-divergence ๊ฐ’์€ ์ˆ˜์‹ (7)๊ณผ ๊ฐ™์ด ๊ฐ ์ƒ๋ณ„ ๊ณ„์‚ฐ ๊ฒฐ๊ณผ์˜ ํ‰๊ท ๊ฐ’์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ์ด๋•Œ $\log$์•ˆ์˜ ๋ถ„์ž ๋ถ„๋ชจ๊ฐ€ 0์ด ๋˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด $\epsilon$๊ฐ’์„ ๋”ํ•˜์—ฌ ์‚ฌ์šฉํ•œ๋‹ค. $\epsilon$์€ $10^{-10}$์„ ์‚ฌ์šฉํ•˜์˜€๋‹ค.

(6)
../../Resources/kiee/KIEE.2024.73.8.1318/eq6.png
(7)
../../Resources/kiee/KIEE.2024.73.8.1318/eq7.png

4) ์ด $M$๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹ ์ค‘ $D_{KL}$๊ฐ’์ด ๊ฐ€์žฅ ๋‚ฎ์€ ๋ฐ์ดํ„ฐ์…‹์„ ์ตœ์ ์˜ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ์„ ์ •ํ•œ๋‹ค.

4. ์‹คํ—˜๊ฒฐ๊ณผ

4.1 ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ธฐ๋ฐ˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ

GAN ๋ชจ๋ธ์˜ ํ•™์Šต๋ฐ์ดํ„ฐ์— ์‚ฌ์šฉ๋˜๋Š” ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ PSCAD๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ƒ์„ฑํ•˜์˜€๋‹ค. ๊ณ„ํ†ต๋„๋Š” ๊ทธ๋ฆผ 3๊ณผ ๊ฐ™์ด ๊ตญ๋‚ด์˜ 345-154kV ๊ณ„ํ†ต์„ ๋‹จ์ˆœํ™”ํ•˜์—ฌ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๊ณ„ํ†ต ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ ์ƒํ™ฉ์„ ์ตœ๋Œ€ํ•œ ๋ชจ์˜ํ•˜๊ธฐ ์œ„ํ•ด ๊ตญ๋‚ด ์‹ค๊ณ„ํ†ต ๋ถ€ํ•˜ ๋ฐ ๋ณ€์••๊ธฐ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์˜€๋‹ค[21]. ๋˜ํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์„ ๊ณ ๋ คํ•˜์—ฌ ๊ณ ์žฅ ์œ ํ˜•, ์†ก์ „์„ ๋กœ ๊ธธ์ด, ์œ„์ƒ ๊ฐ์„ ํ‘œ 1๊ณผ ๊ฐ™์ด ๊ตฌ์„ฑํ•˜์—ฌ ์ด 700๊ฐœ์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ธฐ๋ฐ˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜์˜€๋‹ค. ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋Š” ๊ณ ์žฅ ํ›„ 2์ฃผ๊ธฐ๋ฅผ ์ƒ์„ฑํ•˜์˜€๋‹ค. ์ƒ˜ํ”Œ๋ง ์ฃผ๊ธฐ๋Š” 3840(60$\times$64)HZ๋กœ ํ•˜์˜€๋‹ค.

ํ‘œ 1 PSCAD ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ํŒŒ๋ผ๋ฏธํ„ฐ

Table 1 PSCAD simulation parameter

๊ณ ์žฅ ์œ ํ˜•

AG, BG, CG, ABG, ACG, BCG,

AB, AC, BC, ABC

์†ก์ „์„ ๋กœ๊ธธ์ด(km)

10 ~ 100km (10km ๊ฐ„๊ฒฉ)

์œ„์ƒ(ยฐ)

0 ~ 180ยฐ (30ยฐ ๊ฐ„๊ฒฉ)

๊ทธ๋ฆผ 3. PSCAD ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ณ„ํ†ต๋„

Fig. 3. PSCAD Simulation Topology

../../Resources/kiee/KIEE.2024.73.8.1318/fig3.png

4.2 GAN ๋ชจ๋ธ ๊ตฌ์กฐ ๋ฐ ํ•™์Šต

GAN ๊ธฐ๋ฐ˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•ด ์‚ฌ์šฉํ•œ WGAN-GP ๋ชจ๋ธ์˜ ์ƒ์„ฑ์ž, ํŒ๋ณ„์ž์˜ ๊ตฌ์กฐ๋Š” ํ‘œ 2์™€ ๊ฐ™๋‹ค. ๋‹ค์Œ ํ‘œ๋Š” ๊ฐ ๋ชจ๋ธ์˜ ์ธต์„ ์ˆœ์„œ๋Œ€๋กœ ๋‚˜์—ดํ•˜๊ณ , ๊ฐ ์ธต์—์„œ ์‚ฌ์šฉ๋œ ์ฃผ์š” ํŒŒ๋ผ๋ฏธํ„ฐ๋“ค์„ ์ •๋ฆฌํ•˜์˜€๋‹ค. ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•ด ๊ฐ ๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ WGAN-GP๋ชจ๋ธ์ด ํ›ˆ๋ จ๋œ๋‹ค. ์ƒ์„ฑ์ž์˜ ์ž…๋ ฅ์œผ๋กœ ์‚ฌ์šฉ๋˜๋Š” ์žก์Œ ๋ฒกํ„ฐ์˜ ํฌ๊ธฐ๋Š” (1, 100)์ด๊ณ  ํŒ๋ณ„์ž์— ์‚ฌ์šฉ๋˜๋Š” ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ๋Š” ๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ 70๊ฐœ์”ฉ์ด ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. ์ƒ์„ฑ์ž, ํŒ๋ณ„์ž์˜ ์ตœ์ ํ™”๊ธฐ๋กœ Adam, ํ™œ์„ฑํ™” ํ•จ์ˆ˜๋Š” Leaky ReLU, ํ•™์Šต๋ฅ ์€ 0.0001, ๋ฐฐ์น˜ ํฌ๊ธฐ๋Š” 70์œผ๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ํŒ๋ณ„์ž๋ฅผ ๋‹ค์„ฏ ๋ฒˆ ์—…๋ฐ์ดํŠธํ•  ๋•Œ๋งˆ๋‹ค ์ƒ์„ฑ์ž๋Š” ํ•œ ๋ฒˆ์”ฉ ์—…๋ฐ์ดํŠธํ•˜์—ฌ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜์˜€๋‹ค. ๊ทธ๋ฆผ 4๋Š” ABC ๊ณ ์žฅ ์œ ํ˜•์— ๋Œ€ํ•ด ์—ํฌํฌ์— ๋”ฐ๋ฅธ ์ƒ์„ฑ์ž์™€, ํŒ๋ณ„์ž์˜ ํ›ˆ๋ จ ์†์‹ค์„ ๋‚˜ํƒ€๋‚ธ ๊ทธ๋ž˜ํ”„์ด๋‹ค. ํ›ˆ๋ จ ๋ฐ˜๋ณต ํšŸ์ˆ˜๋ฅผ ์˜๋ฏธํ•˜๋Š” ์—ํฌํฌ๋Š” ์ƒ์„ฑ์ž์™€ ํŒ๋ณ„์ž๊ฐ€ ์ˆ˜๋ ดํ•˜๋Š” ๊ตฌ๊ฐ„์„ ์‹คํ—˜์ ์œผ๋กœ ๋Œ€๋žต ์—ํฌํฌ๋ฅผ 15,000์œผ๋กœ ์„ค์ •ํ•˜์˜€์„ ๋•Œ ์ž˜ ์ˆ˜๋ ดํ•˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•˜์˜€๋‹ค. WGAN-GP ๋ชจ๋ธ ํ›ˆ๋ จ๊ณผ ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์€ ๋ชจ๋‘ Python 3.8.15๋ฅผ ํ†ตํ•ด ๊ตฌํ˜„ํ•˜์˜€๋‹ค.

ํ‘œ 2 WGAN-GP ๋ชจ๋ธ ์ฃผ์š” ํŒŒ๋ผ๋ฏธํ„ฐ

Table 2 Parameters of WGAN-GP model

์ˆœ์„œ

์ธต ์ข…๋ฅ˜

์ฃผ์š” ํŒŒ๋ผ๋ฏธํ„ฐ

์ƒ์„ฑ์ž(Generator)

1

Dense

units=8192

2

Conv2DTranspose

filters=64, kernel_size=4x4

3

Conv2DTranspose

filters=64, kernel_size=4x4

4

Conv2DTranspose

filters=1, kernel_size=4x4

ํŒ๋ณ„์ž(Discrimator)

1

Conv2D

filters=32, kernel_size=4x4

2

Conv2D

filters=64, kernel_size=4x4

3

Conv2D

filters=128, kernel_size=4x4

4

Dense

units=1

๊ทธ๋ฆผ 4. Epoch์— ๋”ฐ๋ฅธ ํ›ˆ๋ จ ์†์‹ค (a) ์ƒ์„ฑ์ž์˜ ํ›ˆ๋ จ ์†์‹ค, (b) ํŒ๋ณ„์ž์˜ ํ›ˆ ๋ จ ์†์‹ค (ABC ๊ณ ์žฅ)

Fig. 4. Training Loss of Generator(a) and Discriminator(b) Over Epochs (ABC fault case)

../../Resources/kiee/KIEE.2024.73.8.1318/fig4.png

4.3 GAN ๊ธฐ๋ฐ˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ

๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ ์ด 10๊ฐœ์˜ GAN ๋ชจ๋ธ ํ›ˆ๋ จ์„ ๋งˆ์นœ ํ›„ ๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ ํ›ˆ๋ จ๋œ ์ƒ์„ฑ์ž $G$๋ฅผ ํ†ตํ•ด ์˜ค์—ผ๋„ $C$์˜ ๊ฐ’์„ 0.05๋ถ€ํ„ฐ 0.2 ๊นŒ์ง€ 0.05์”ฉ ์˜ฌ๋ ค๊ฐ€๋ฉด์„œ ์ด 4๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•œ๋‹ค. ํ•œ ๊ฐœ์˜ ๋ฐ์ดํ„ฐ์…‹์€ 1,000๊ฐœ์˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. ๊ฐ๊ฐ์˜ ๋ฐ์ดํ„ฐ ๋ณ„๋กœ RMS๋Š” ์ˆ˜์‹ (5)๋ฅผ ํ†ตํ•ด ๊ณ„์‚ฐํ•œ๋‹ค. ๊ณ ์žฅ๋ฐ์ดํ„ฐ ํ•˜๋‚˜๋Š” ์ด 3๊ฐœ์˜ ์ƒ(A, B, C ์ƒ)์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๊ณ , ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ๊ฐœ์ˆ˜์ธ $l$์€ 128๊ฐœ(2์ฃผ๊ธฐ)๋กœ ๊ตฌ์„ฑ๋˜์—ˆ์œผ๋ฉฐ $N$์€ ์ •์ƒ์ƒํƒœ ์ „๋ฅ˜์˜ 1์ฃผ๊ธฐ์— ํ•ด๋‹นํ•˜๋Š” ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ๊ฐœ์ˆ˜์ธ 64๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ํ•˜๋‚˜์˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•ด ์ƒ ๋ณ„๋กœ RMS๋Š” ์ด 65๊ฐœ ๊ฐ’์ด ๊ณ„์‚ฐ๋œ๋‹ค. ๊ฐ๊ฐ์˜ ์ƒ๋ณ„ $D_{KL}^{A}$, $D_{KL}^{B}$, $D_{KL}^{C}$๊ฐ’์˜ ํ‰๊ท ์„ ๋‚ธ $D_{KL}$๊ฐ’์ด ๊ฐ€์žฅ ๋‚ฎ์€ ๋ฐ์ดํ„ฐ์…‹์„ ์ตœ์ ์˜ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ์„ ์ •ํ•œ๋‹ค. ์˜ค์—ผ๋„ $C$์˜ ๊ฐ’์— ๋”ฐ๋ผ $D_{KL}$๊ฐ’์„ ๋น„๊ตํ•œ ๊ฒฐ๊ณผ๋ฅผ ํ‘œ 3์— ๋„์‹œํ•˜์˜€๋‹ค. ๊ฐ ๊ณ ์žฅ ์œ ํ˜•์— ๋”ฐ๋ผ ์ตœ์ ์˜ ์˜ค์—ผ๋„๊ฐ€ ๋‹ค๋ฅธ ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค.

ํ‘œ 3 ๊ณ ์žฅ ์œ ํ˜•๋ณ„ ์˜ค์—ผ๋„์— ๋”ฐ๋ฅธ KL-divergence๊ฐ’

Table 3 KL-divergence values based on fault type and contamination level

๊ณ ์žฅ ์œ ํ˜•

์˜ค์—ผ๋„ (C)

0.05(5%)

0.1(10%)

0.15(15%)

0.2(20%)

AG

0.728

0.878

0.813

0.951

BG

0.765

0.835

1.014

0.931

CG

0.657

0.666

0.699

0.693

AB

1.660

1.591

1.481

1.609

BC

1.884

1.958

1.986

1.912

AC

1.379

1.410

1.506

1.552

ABG

0.625

0.666

0.637

0.681

BCG

0.759

0.852

0.776

0.825

ACG

0.759

0.730

0.778

0.838

ABC

0.762

0.751

0.776

0.786

๊ทธ๋ฆผ 5๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ธฐ๋ฐ˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ์™€ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ ์ค‘์—์„œ ์ •์ƒ์œผ๋กœ ๋ถ„๋ฅ˜๋œ ๋ฐ์ดํ„ฐ์™€ ์ด์ƒ์น˜๋กœ ์ œ๊ฑฐ๋œ ๋ฐ์ดํ„ฐ ํŒŒํ˜•์˜ ์˜ˆ์‹œ๋ฅผ ๋„์‹œํ•˜์˜€๋‹ค. ๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ LG, LL, LLG, LLL ๊ณ ์žฅ์„ ๋Œ€ํ‘œ๋กœ AG, AB, ABG, ABC ๊ณ ์žฅ์„ ๋„์‹œํ•˜์˜€์œผ๋ฉฐ, ๋น„๊ต์˜ ํŽธ์˜๋ฅผ ์œ„ํ•ด ๋ชจ๋“  ๋ฐ์ดํ„ฐ๋Š” A, B, C์ƒ ์ „๋ฅ˜์˜ ์ตœ๋Œ“๊ฐ’์„ ๊ธฐ์ค€์œผ๋กœ ์ •๊ทœํ™”ํ•˜์˜€๋‹ค. ์ด์ƒ์น˜๋กœ ํŒ๋ณ„๋œ ๋ฐ์ดํ„ฐ๋Š” ํŒŒํ˜•์ด ๊ณผ๋„ํ•˜๊ฒŒ ๋ณ€ํ˜•๋˜์—ˆ๊ฑฐ๋‚˜ ์žก์Œ๊ณผ ๊ฐ™์ด ๋ถˆ๊ทœ์น™ํ•œ ๋ณ€๋™์ด ์‹ฌํ•œ ํŠน์ง•์„ ๋ณด์ด๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ ์ด๋Ÿฐ ์ด์ƒ์น˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ œ๊ฑฐํ•˜๋Š” ๊ฒƒ์ด ๋ณด๋‹ค ํ˜„์‹ค์ ์ธ ๊ณ ์žฅ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•˜๋Š”๋ฐ ๋„์›€์ด ๋œ๋‹ค.

๊ทธ๋ฆผ 5. ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ, ์ƒ์„ฑ ๋ฐ์ดํ„ฐ, ์ด์ƒ์น˜ ๋ฐ์ดํ„ฐ ์˜ˆ์‹œ

Fig. 5. Examples of Simulation Data, Generated Data, and Anomaly Data

../../Resources/kiee/KIEE.2024.73.8.1318/fig5.png

๋‹ค์Œ์€ T-SNE(T-distributed Stochastic Neighbor Embedding)๋ฅผ ํ†ตํ•œ ์‹œ๊ฐํ™” ๊ณผ์ •์„ ํ†ตํ•ด ํ™•๋ฅ ๋ถ„ํฌ ๊ณ„์‚ฐ ์‹œ RMS์˜ ํ•„์š”์„ฑ์„ ๋„์‹œํ•œ๋‹ค. T-SNE๋Š” ๊ณ ์ฐจ์› ๋ฐ์ดํ„ฐ ํฌ์ธํŠธ ๊ฐ„์˜ ์œ ์‚ฌ์„ฑ์„ ํ™•๋ฅ ๋ถ„ํฌ๋กœ ๋ชจ๋ธ๋งํ•˜๊ณ , ์ด ์œ ์‚ฌ์„ฑ์„ ์ €์ฐจ์›๊ณต๊ฐ„์—์„œ ์ตœ๋Œ€ํ•œ ์œ ์ง€ํ•จ์œผ๋กœ์จ ๋ฐ์ดํ„ฐ์˜ ๊ตฌ์กฐ์™€ ํŒจํ„ด์„ ์‹œ๊ฐ์ ์œผ๋กœ ๋ถ„์„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ํ•ด์ค€๋‹ค[22]. ๋‹ค์Œ ๊ทธ๋ฆผ 6์€ ๊ณ ์žฅ๋ฐ์ดํ„ฐ 700๊ฐœ๋ฅผ ๊ฐ๊ฐ ์ˆœ์‹œ์น˜, RMS๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ 2์ฐจ์›์œผ๋กœ ์‹œ๊ฐํ™”ํ•˜๊ณ  ๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ ์ƒ‰์„ ์ž…ํ˜”๋‹ค. ๊ทธ๋ฆผ 6์˜ ์ˆœ์‹œ์น˜๋ฅผ ํ†ตํ•œ ์‹œ๊ฐํ™” ๊ฒฐ๊ณผ ๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ ๊ตฐ์ง‘์ด ํ˜•์„ฑ๋˜์ง€ ์•Š๋Š” ๋ฐ˜๋ฉด, RMS๋ฅผ ํ†ตํ•œ ๊ฒฐ๊ณผ๋Š” ์ผ๋ถ€ LL-LLG๊ฐ„ ๊ณ ์žฅ์„ ์ œ์™ธํ•œ ๋‚˜๋จธ์ง€ ์œ ํ˜•๋“ค์ด ๊ตฐ์ง‘ํ™”๋˜์–ด ๋ช…ํ™•ํ•˜๊ฒŒ ๋ถ„๋ฆฌ๋˜๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Š” ๊ฐ ๊ณ ์žฅ ์œ ํ˜•์˜ ํŠน์„ฑ์ด ๋ฐ์ดํ„ฐ์— ๋ฐ˜์˜๋˜์—ˆ์Œ์„ ๋‚˜ํƒ€๋‚ด๊ณ , RMS๋ฅผ ๊ธฐ์ค€์œผ๋กœ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ๊ณ„์‚ฐํ•˜๋Š” ๊ฒƒ์ด ๊ณ ์žฅ๋ฐ์ดํ„ฐ์˜ ํŠน์„ฑ์„ ์ž˜ ๋ฐ˜์˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค€๋‹ค.

๊ทธ๋ฆผ 6. T-SNE๋ฅผ ํ†ตํ•œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์‹œ๊ฐํ™” (a) ์ˆœ์‹œ์น˜, (b) RMS

Fig. 6. Visualization of Simulation Fault Data using T-SNE (a) Instantaneous Value, (b) RMS

../../Resources/kiee/KIEE.2024.73.8.1318/fig6.png

๊ทธ๋ฆผ 7. T-SNE๋ฅผ ํ†ตํ•œ ์ƒ์„ฑ๋œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์‹œ๊ฐํ™”

Fig. 7. Visualization of Generated Fault Data using T-SNE

../../Resources/kiee/KIEE.2024.73.8.1318/fig7.png

๊ทธ๋ฆผ 7์€ ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์œผ๋กœ ์ƒ์„ฑ๋œ 10,000๊ฐœ์˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ RMS๋ฅผ ์ด์šฉํ•ด ์ „์ฒ˜๋ฆฌํ•œ ํ›„, T-SNE๋ฅผ ํ™œ์šฉํ•˜์—ฌ 2์ฐจ์›์œผ๋กœ ์‹œ๊ฐํ™”ํ•œ ๊ฒฐ๊ณผ๋ฅผ ๋ณด์—ฌ์ค€๋‹ค. ๊ทธ๋ฆผ 7์˜ ๊ฒฐ๊ณผ์—์„œ๋„ ๊ทธ๋ฆผ 6(b)์™€ ๊ฐ™์ด ๊ณ ์žฅ ์œ ํ˜•๋ณ„๋กœ ์„œ๋กœ ์˜์—ญ์„ ์นจ๋ฒ”ํ•˜์ง€ ์•Š๊ณ  ๊ตฐ์ง‘์ด ํ˜•์„ฑ๋˜๋Š” ๊ฒƒ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Š” ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ๊ธฐ์กด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹์˜ ํŠน์ง•์„ ์ž˜ ํ•™์Šตํ–ˆ์Œ์„ ๊ฐ„์ ‘์ ์œผ๋กœ ๋ณด์—ฌ์ค€๋‹ค. ๋”ฐ๋ผ์„œ, ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์ด ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ์ฆ๊ฐ•์‹œํ‚ค๋Š” ๊ฒƒ์„ ํ™•์ธํ•˜์˜€๋‹ค.

4.4 GAN ๊ธฐ๋ฐ˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ์…‹ ๋ถ„ํฌ ์‹œ๊ฐํ™”

์ถ”๊ฐ€์ ์œผ๋กœ GAN์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์˜ ๋‹ค์–‘์„ฑ์„ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ ์‹œ๊ฐํ™”๋ฅผ ํ†ตํ•ด ํ™•์ธํ•œ๋‹ค. ์ด๋•Œ ์‹ค์ œ ์†ก์ „์„ ๋กœ์—์„œ ๋ฐœ์ƒํ•œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋„ ์ถ”๊ฐ€ํ•˜์—ฌ ๋น„๊ต๋ฅผ ์ง„ํ–‰ํ•œ๋‹ค. ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋Š” ๊ตญ๋‚ด ๋‹ค์–‘ํ•œ ์œ„์น˜์—์„œ ์ทจ๋“๋˜์—ˆ์œผ๋ฉฐ, ๊ณผ๊ฑฐ ์•ฝ 9๋…„๊ฐ„์˜ ๊ธฐ๊ฐ„ ๋™์•ˆ ๋ฐœ์ƒํ•œ ์ด 51๊ฐœ์˜ ๊ตญ๋‚ด ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์ด๋‹ค. ๋ณธ ์—ฐ๊ตฌ์—์„œ ์‚ฌ์šฉํ•œ ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์˜ ๊ณ ์žฅ ์œ ํ˜•๋ณ„ ๋ฐ์ดํ„ฐ ๊ฐœ์ˆ˜๋Š” ํ‘œ 4๋ฅผ ํ†ตํ•ด ์ •๋ฆฌํ•˜์˜€๋‹ค.

ํ‘œ 4 ๊ณ ์žฅ ์œ ํ˜• ๋ณ„ ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ ๊ฐœ์ˆ˜

Table 4 Number of Actual Data by Fault Type

์œ ํ˜•

์„ธ๋ถ€ ์œ ํ˜•

๋ฐ์ดํ„ฐ ๊ฐœ์ˆ˜ (๊ฐœ)

LG

AG

5

BG

11

CG

9

LL

AB

3

BC

7

AC

0

LLG

ABG

2

BCG

0

ACG

3

LLL

ABC

11

์ด ๊ฐœ์ˆ˜

51

๋‹ค์Œ์€ ๊ณ ์žฅ๋ฐ์ดํ„ฐ์˜ RMS ์ค‘ ๊ฐ€์žฅ ํฐ ๊ฐ’์„ ์ด์šฉํ•˜์—ฌ 3์ฐจ์› ํ‰๋ฉด์— ์‹œ๊ฐํ™”ํ•˜๋Š” ๊ณผ์ •์„ ๊ทธ๋ฆผ 8์„ ํ†ตํ•ด ์„ค๋ช…ํ•œ๋‹ค. ๊ทธ๋ฆผ 8์˜ (a)๋Š” ์ˆœ์‹œ์น˜ ํŒŒํ˜•์„ ๋‚˜ํƒ€๋‚ด๊ณ  (b)๋Š” RMS๋กœ ์ „์ฒ˜๋ฆฌ๋œ ํŒŒํ˜•์„ ๋‚˜ํƒ€๋‚ด๋ฉฐ (c)๋Š” RMS ์ค‘ ์ตœ๋Œ“๊ฐ’์„ 3์ฐจ์› ํ‰๋ฉด์— ๋„์‹œํ•œ ๊ทธ๋ฆผ์ด๋‹ค.

๊ทธ๋ฆผ 9๋Š” AG, AB, ABG, ABC์˜ ๊ณ ์žฅ ์œ ํ˜•์„ ๊ทธ๋ฆผ 8์˜ ๊ณผ์ •์„ ๋”ฐ๋ผ์„œ 3์ฐจ์› ํ‰๋ฉด์— ์ฐจ๋ก€๋Œ€๋กœ (a), (b), (c), (d)์— ์‹œ๊ฐํ™”ํ•˜์˜€๋‹ค. ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ๋Š” ํŒŒ๋ž€์ƒ‰, ์ƒ์„ฑ ๋ฐ์ดํ„ฐ๋Š” ๋นจ๊ฐ„์ƒ‰, ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋Š” ์ดˆ๋ก์ƒ‰์œผ๋กœ ๋‚˜ํƒ€๋‚ด๋ฉฐ ๊ฐ ๋ฐ์ดํ„ฐ ๊ฐ„์˜ ๋ถ„ํฌ๋ฅผ ์‹œ๊ฐ์ ์œผ๋กœ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋ฆผ 9์—์„œ ๋‘๋“œ๋Ÿฌ์ง€๊ฒŒ ๋ณด์ด๋“ฏ์ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์™€ ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์‚ฌ์ด์—๋Š” ๊ฐ„๊ทน์ด ์กด์žฌํ•œ๋‹ค. ์ด๋Š”, ์‹œ๋ฎฌ๋ ˆ์ด์…˜์„ ํ†ตํ•ด ์„ ๋กœ๊ธธ์ด, ์œ„์ƒ๊ฐ์„ ๋ณ€๊ฒฝํ•˜๋ฉฐ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ๋งŒ์œผ๋กœ๋Š” ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์™€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์— ํ•œ๊ณ„๊ฐ€ ์žˆ์Œ์„ ๋ณด์—ฌ์ค€๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ ๋ถ„ํฌ์˜ ์ฃผ๋ณ€์„ ๋”ฐ๋ผ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋˜๋Š” ๋ชจ์Šต์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์ฆ‰, ์ œ์•ˆํ•˜๋Š” ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ํ™•๋ฅ ์ ์œผ๋กœ ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์™€ ์œ ์‚ฌํ•œ ๋ฐ์ดํ„ฐ๊ฐ€ ์ƒ์„ฑ๋  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค€๋‹ค.

๊ทธ๋ฆผ 8. ๊ณ ์žฅ๋ฐ์ดํ„ฐ 3์ฐจ์› ์‹œ๊ฐํ™” ๊ณผ์ • (a) ์ˆœ์‹œ์น˜, (b) RMS, (c) ๊ฐ์ƒ๋ณ„ RMS์˜ ์ตœ๋Œ“๊ฐ’์„ ํ†ตํ•œ 3์ฐจ์› ์‹œ๊ฐํ™”

Fig. 8. Process of Three-dimensional Visualization of Fault Data (a) Instantaneous value, (b) RMS values, (c) Three-dimensional visualization with the maximum of the effective values for each phase

../../Resources/kiee/KIEE.2024.73.8.1318/fig8.png

๊ทธ๋ฆผ 9. ๋ฐ์ดํ„ฐ์…‹ 3์ฐจ์› ์‹œ๊ฐํ™” (a) AG, (b) AB, (c) ABG, (d)ABC

Fig. 9. Three-dimensional Visualization of the Dataset (a) AG, (b) AB, (c) ABG, (d) ABC

../../Resources/kiee/KIEE.2024.73.8.1318/fig9.png

4.5 ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์…‹๊ณผ์˜ KL-divergence ๊ฐ’ ๋น„๊ต

์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์˜ ํ’ˆ์งˆ์„ ์ •๋Ÿ‰์ ์œผ๋กœ ๊ฒ€์ฆํ•˜๊ธฐ ์œ„ํ•ด ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ์ค€์œผ๋กœ KL-divergence ๊ฐ’์„ ๊ณ„์‚ฐํ•˜๊ณ , ๋ฐ์ดํ„ฐ์…‹ ๋ถ„ํฌ์˜ ์œ ์‚ฌ๋„๋ฅผ ๋น„๊ตํ•œ๋‹ค. ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์™€ ๋น„๊ต๋˜๋Š” ๋ฐ์ดํ„ฐ์…‹์˜ ์ข…๋ฅ˜๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹, ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์ด ์ ์šฉ๋œ ์ƒ์„ฑ ๋ฐ์ดํ„ฐ์…‹์ด๋‹ค. ๋ฐ์ดํ„ฐ์…‹์— ๋”ฐ๋ผ ๊ณ„์‚ฐ๋œ KL-divergence ๊ฐ’์„ ํ‘œ 5์— ์ •๋ฆฌํ•˜์˜€๋‹ค. ๊ณ„์‚ฐ ๊ฒฐ๊ณผ ๋ชจ๋“  ๊ณ ์žฅ ์œ ํ˜•์— ๋Œ€ํ•ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹ ๋ณด๋‹ค ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์˜ KL-divergence ๊ฐ’์ด ๋” ๋‚ฎ์Œ์„ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋Š” GAN์„ ํ†ตํ•ด ์ƒ์„ฑํ•œ ๋ฐ์ดํ„ฐ์…‹์ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹๋ณด๋‹ค ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์˜ ํ™•๋ฅ ๋ถ„ํฌ์™€ ์œ ์‚ฌํ•˜๋‹ค๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•œ๋‹ค. ๋”ฐ๋ผ์„œ ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์œผ๋กœ ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์™€ ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ ๊ฐ„์˜ ์ฐจ์ด๋ฅผ ์ค„์ผ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์™€ ๋” ์œ ์‚ฌํ•œ ํ˜•ํƒœ์˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ์Œ์„ ์˜๋ฏธํ•œ๋‹ค.

ํ‘œ 5 ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์…‹ ๊ณผ์˜ KL-divergence ๊ฐ’ ๋น„๊ต

Table 5 Comparison of KL-divergence Values with Real Fault Dataset

๊ณ ์žฅ ์œ ํ˜•

์‹œ๋ฎฌ๋ ˆ์ด์…˜

๋ฐ์ดํ„ฐ์…‹

์ƒ์„ฑ๋œ

๋ฐ์ดํ„ฐ์…‹

LG

AG

9.212

6.229

BG

8.046

5.617

CG

7.556

5.024

LL

AB

10.969

6.634

BC

7.213

4.767

AC

์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์—†์Œ

LLG

ABG

8.517

6.461

BCG

์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์—†์Œ

ACG

12.820

6.953

LL

ABC

5.992

1.394

5. ๊ฒฐ ๋ก 

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•œ GAN ๊ธฐ๋ฐ˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ธฐ๋ฒ•์„ ์ œ์•ˆํ•˜์˜€๋‹ค. ์ œ์•ˆ๋œ ๊ธฐ๋ฒ•์€ ์ƒ์„ฑ๊ณผ ์ตœ์  ๋ฐ์ดํ„ฐ์…‹ ์„ ์ •์˜ ๋‘ ๋‹จ๊ณ„๋กœ ๊ตฌ์„ฑ๋œ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ ์‹ค๊ณ„ํ†ต์˜ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋ฐ˜์˜ํ•œ ๋‹ค์–‘ํ•œ ์†ก์ „์„ ๋กœ ๊ณ ์žฅ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ๋กœ GAN์„ ํ•™์Šตํ•˜์˜€๋‹ค. ์ œ์•ˆํ•˜๋Š” ๊ธฐ๋ฒ•์—์„œ๋Š” ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•์— ์ ํ•ฉํ•œ WGAN-GP ๋ชจ๋ธ์„ ์ƒ์„ฑ ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ์‚ฌ์šฉํ•˜์˜€๋‹ค. ๋‘ ๋ฒˆ์งธ ๋‹จ๊ณ„์—์„œ๋Š” ๋จผ์ € IF ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•˜์—ฌ ๋‹ค์–‘ํ•œ ์ด์ƒ์น˜ ๋น„์œจ์„ ์ ์šฉํ•ด ๋ณต์ˆ˜์˜ ๋ฐ์ดํ„ฐ์…‹์„ ์ƒ์„ฑํ•˜๊ณ , ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹๋“ค ์ค‘์—์„œ KL-divergence๊ฐ’์„ ๊ธฐ์ค€์œผ๋กœ ํ•™์Šต๋ฐ์ดํ„ฐ ๋ถ„ํฌ์™€ ๊ฐ€์žฅ ์œ ์‚ฌํ•œ ์ตœ์ ์˜ ๋ฐ์ดํ„ฐ์…‹์„ ๊ฒฐ์ •ํ•˜์˜€๋‹ค. ์ด ๊ณผ์ •์„ ํ†ตํ•ด, ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹์˜ ํ™•๋ฅ ๋ถ„ํฌ์™€ ์œ ์‚ฌํ•˜๊ฒŒ ๋œ๋‹ค. IF ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ด์šฉํ•œ ์ด์ƒ์น˜ ์ œ๊ฑฐ๋ฅผ ํ†ตํ•ด ํŒŒํ˜•์ด ๊ณผ๋„ํ•˜๊ฒŒ ๋ณ€ํ˜•๋˜๊ฑฐ๋‚˜ ์žก์Œ์ด ๋งŽ์€ ํŒŒํ˜•์ฒ˜๋Ÿผ ๋ณด์ด๋Š” ์ €ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ๊ฐ€ ์ œ๊ฑฐ๋จ์„ ํ™•์ธํ•˜์˜€๋‹ค. ์ œ์•ˆํ•˜๋Š” ์ฆ๊ฐ•๊ธฐ๋ฒ•์œผ๋กœ ํ™•๋ณดํ•œ ๋ฐ์ดํ„ฐ์…‹์€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์™€ ์œ ์‚ฌํ•œ ๋ถ„ํฌ๋ฅผ ๋”ฐ๋ฅด๋ฉด์„œ ๋ณด๋‹ค ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๋Š” ๊ฒƒ์„ ์‹œ๊ฐ์ ์œผ๋กœ ํ™•์ธํ•˜์˜€๋‹ค. ์ตœ์ข…์ ์œผ๋กœ ์ œ์•ˆํ•˜๋Š” ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ์ƒ์„ฑ๋œ ๋ฐ์ดํ„ฐ์…‹์ด ๋ชจ๋“  ๊ณ ์žฅ ์œ ํ˜•์— ๋Œ€ํ•ด ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐ์ดํ„ฐ์…‹๋ณด๋‹ค ๋” ์‹ค๊ณ„ํ†ต ๊ณ ์žฅ๋ฐ์ดํ„ฐ์™€ ์œ ์‚ฌํ•œ ๊ฒƒ์„ KL-divergence ๊ฐ’์„ ํ†ตํ•ด ํ™•์ธํ•˜์˜€๋‹ค.

ํ›„์† ์—ฐ๊ตฌ์—์„œ๋Š” ๋” ์ •ํ™•ํ•˜๊ณ  ํ˜„์‹ค์ ์ธ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ์ƒ์„ฑ์„ ์œ„ํ•ด ์ „์•• ํŒŒํ˜•๊ณผ ์ „๋ฅ˜ ํŒŒํ˜•์˜ ์œ„์ƒ ๊ด€๊ณ„์™€ ๊ฐ™์€ ๋ฌผ๋ฆฌ์  ์ง€์‹์„ ๊ฒฐํ•ฉํ•œ ์ƒ์„ฑ๊ธฐ๋ฒ•์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ํ•  ๊ณ„ํš์ด๋‹ค. ๋˜ํ•œ ๋ณธ ์—ฐ๊ตฌ์—์„œ ์ƒ์„ฑ๋œ ์†ก์ „์„ ๋กœ ๊ณ ์žฅ๋ฐ์ดํ„ฐ๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ํ›ˆ๋ จ์‹œํ‚จ ๊ณ ์žฅ์œ ํ˜• ๋ถ„๋ฅ˜๋ชจ๋ธ๊ณผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ณ ์žฅ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜ ๊ณ ์žฅ์œ ํ˜• ๋ถ„๋ฅ˜๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋น„๊ตํ•จ์œผ๋กœ ์‹ค๊ณ„ํ†ต ์ƒํ™ฉ์—์„œ ์ œ์•ˆํ•˜๋Š” ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ธฐ๋ฒ•์˜ ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ํ™•์ธํ•  ๊ณ„ํš์ด๋‹ค.

Acknowledgements

๋ณธ ์—ฐ๊ตฌ๋Š” ํ•œ๊ตญ์ „๋ ฅ๊ณต์‚ฌ์˜ 2022๋…„ ์ฐฉ์ˆ˜ ๊ธฐ์ดˆ์—ฐ๊ตฌ๊ฐœ๋ฐœ ๊ณผ์ œ์˜ ์ง€์›(No. R22XO02-19) ๋ฐ ์‚ฐ์—…ํ†ต์ƒ์ž์›๋ถ€(MOTIE), ํ•œ๊ตญ์—๋„ˆ์ง€๊ธฐ์ˆ ํ‰๊ฐ€์›(KETEP)์˜ ์ง€์›(No. RS-2024-00398166)์„ ๋ฐ›์•„ ์ˆ˜ํ–‰ํ•œ ์—ฐ๊ตฌ ๊ณผ์ œ์ž…๋‹ˆ๋‹ค.

References

1 
H. Liang, X. Han, H. Yu, F. Li, Z. Liu, and K. Zhang, โ€œTransmission line fault-cause identification method for large-scale new energy grid connection scenarios,โ€ Global Energy Interconnection, vol. 5, no. 4, pp. 362-374, 2022.DOI
2 
A. Yadav, and Y. Dash, โ€œAn overview of transmission line protection by artificial neural network: fault detection, fault classification, fault location, and fault direction discrimination,โ€ Advances in Artificial Neural Systems, vol. 2014, no. 1, Article 230382, 2014.DOI
3 
R. Kuffel, J. Giesbrecht, T. Maguire, R.P. Wierchx, and P. Melaren, โ€œRTDS-A Fully Digital Power System Simulator Operation in Real Time,โ€ in Proc. 1995 International Conference on Energy Management and Power Delivery EMPD'95, vol. 2, pp. 19-24, 1995.DOI
4 
A. Abdullah, โ€œUltrafast transmission line fault detection using a DWT-based ANN,โ€ IEEE Transactions on Industry Applications, vol. 54, no. 2, pp. 1182-1193, 2017.DOI
5 
Y. Xing, Y. Liu, D. Lu, X. Zou, and X. He, โ€œA Physics-Informed Data-Driven Fault Location Method for Transmission Lines Using Single-Ended Measurements with Field Data Validation,โ€ arXiv preprint arXiv:2307.09740, 2023.DOI
6 
L. Yang, Z. Wang, S. Wang, F. Deng, and P. Yang, โ€œSimulation on HVDC Actual Fault and Analysis of Simulation Reliability,โ€ in Proc. 2021 3rd Asia Energy and Electrical Engineering Symposium (AEEES), IEEE, pp. 789-793, 2021.DOI
7 
T. G. Kim, S. H. Lim, K. M. Song, and S. Yoon, โ€œLSTM-based Fault Classification Model in Transmission Lines for Real Fault Data,โ€ KIEE Journal, vol. 73, no. 3, pp. 585-592, 2024.DOI
8 
M. Pazoki, โ€œA New Fault Classifier in Transmission Lines Using Intrinsic Time Decomposition,โ€ IEEE Transactions on Industrial Informatics, vol. 14, no. 2, pp. 619-628, 2018.DOI
9 
A. Mukherjee, P. K. Kundu, and A. Das, โ€œTransmission line fault classification under high noise in signal: a direct PCA-threshold-based approach,โ€ Journal of The Institution of Engineers (India): Series B, vol. 103, pp. 197-211, 2021.DOI
10 
A. Mukherjee, P. K. Kundu, and A. Das, โ€œA differential signal-based fault classification scheme using PCA for long transmission lines,โ€ Journal of The Institution of Engineers (India): Series B, vol. 102, pp. 403-414, 2021.DOI
11 
M. F. Guo, W. L. Liu, J. H. Gao, and D. Y. Chen, โ€œA data-enhanced high impedance fault detection method under imbalanced sample scenarios in distribution networks,โ€ IEEE Transactions on Industry Applications, vol. 59, no. 4, pp. 4720-4733, 2023.DOI
12 
W. L. Liu, M. F. Guo, and J. H. Gao, โ€œHigh impedance fault diagnosis method based on conditional Wasserstein generative adversarial network,โ€ in Proc. 2021 IEEE 2nd China International Youth Conference on Electrical Engineering (CIYCEE), IEEE, pp. 1-6, 2021.DOI
13 
A. Torfi, M. Beyki, and E. A. Fox, โ€œOn the evaluation of generative adversarial networks by discriminative models,โ€ in 2020 25th International Conference on Pattern Recognition (ICPR), IEEE, pp. 991-998, 2021.DOI
14 
M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, โ€œGans trained by a two time-scale update rule converge to a local nash equilibrium,โ€ Advances in Neural Information Processing Systems, vol. 30, 2017.URL
15 
S. Barratt, and R. Sharma, โ€œA note on the inception score,โ€ arXiv preprint arXiv:1801.01973, 2018.DOI
16 
K. Y. Lee, S. H. Lim, T. G. Kim, K. M. Song, S. G. Yoon, "GAN-based Generative Algorithm for Various Transmission Line Fault Data," in Proc. KIEE spring conference, 2024.URL
17 
X. Gao, F. Deng, and X. Yue, โ€œData augmentation in fault diagnosis based on the Wasserstein generative adversarial network with gradient penalty,โ€ Neurocomputing, vol. 396, pp. 487-494, 2020.DOI
18 
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, and Y. Bengio, โ€œGenerative adversarial nets,โ€ Advances in Neural Information Processing Systems, vol. 27, 2014.URL
19 
I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. C. Courville, โ€œImproved training of wasserstein gans,โ€ Advances in Neural Information Processing Systems, vol. 30, 2017.URL
20 
F. T. Liu, K. M. Ting, and Z. H. Zhou, โ€œIsolation forest,โ€ in 2008 Eighth IEEE International Conference on Data Mining, IEEE, pp. 413-422, 2008.DOI
21 
K. M. Song, T. G. Kim, C. W. Park, and S. G. Yoon, โ€œTransmission Line Fault Cause Modeling and Waveform Analysis,โ€ in Proc. KIEE spring conference, 2023.URL
22 
L. Van der Maaten, and G. Hinton, โ€œVisualizing data using t-SNE,โ€ Journal of machine learning research, vol. 9, no. 11, pp. 2579-2605, 2008.URL

์ €์ž์†Œ๊ฐœ

์ด๊ฒฝ์˜ (Kyeong-Yeong Lee)
../../Resources/kiee/KIEE.2024.73.8.1318/au1.png

He received her B.S. degree in Electrical Engineering from Soongsil University, Seoul, South Korea, in 2021. Currently, he is pursuing M.E. at Soongsil University, Seoul, Korea.

E-mail: dlru7755@naver.com

์ž„์„ธํ—Œ (Se-Heon Lim)
../../Resources/kiee/KIEE.2024.73.8.1318/au2.png

She received her B.S. degree in Electrical Engineering from Soongsil University, Seoul, South Korea, in 2018. Currently, she is pursuing Ph.D. degree at Soongsil University, Seoul, Korea.

E-mail: seheon0223@naver.com

๊น€ํƒœ๊ทผ (Tae-Geun Kim)
../../Resources/kiee/KIEE.2024.73.8.1318/au3.png

He received his B.S. degree in Electrical and Electronics Engineering from Kangwon University, Chuncheon, South Korea, in 2020. Currently he is pursuing Ph.D. degree at Soongsil University, Seoul, Korea.

E-mail: taegeun1520@gmail.com

์†ก๊ฒฝ๋ฏผ (Kyung-Min Song)
../../Resources/kiee/KIEE.2024.73.8.1318/au4.png

He received her B.S. degree in Electrical Engineering from Soongsil University, Seoul, South Korea, in 2022. Currently, he is pursuing M.E. at Soongsil University, Seoul, Korea.

E-mail: songlk111@naver.com

์œค์„ฑ๊ตญ (Sung-Guk Yoon)
../../Resources/kiee/KIEE.2024.73.8.1318/au5.png

He received the B.S. and Ph.D. degrees in Electrical Engineering and Computer Science from Seoul National University, Seoul, South Korea, in 2006 and 2012, respectively. He is currently with Soongsil University as an associate professor. His research interests include energy big data, game theory for power system, and power system optimization.

E-mail: sgyoon@ssu.ac.kr