Mobile QR Code QR CODE

Journal of the Korea Concrete Institute

J Korea Inst. Struct. Maint. Insp.
  • Indexed by
  • Korea Citation Index (KCI)

  1. ์„œ์šธ์‹œ๋ฆฝ๋Œ€ํ•™๊ต,๊ฑด์ถ•๊ณตํ•™๊ณผ ์Šค๋งˆํŠธ์‹œํ‹ฐ์œตํ•ฉ์ „๊ณต ์„์‚ฌ
  2. ์„œ์šธ์‹œ๋ฆฝ๋Œ€ํ•™๊ต,๊ฑด์ถ•๊ณตํ•™๊ณผ ์Šค๋งˆํŠธ์‹œํ‹ฐ์œตํ•ฉ์ „๊ณต ๋ฐ•์‚ฌ๊ณผ์ •
  3. ์„œ์šธ์‹œ๋ฆฝ๋Œ€ํ•™๊ต,๊ฑด์ถ•๊ณตํ•™๊ณผ ๋ฐ•์‚ฌ๊ณผ์ •
  4. ์„œ์šธ์‹œ๋ฆฝ๋Œ€ํ•™๊ต,๊ฑด์ถ•๊ณตํ•™๊ณผ ์Šค๋งˆํŠธ์‹œํ‹ฐ์œตํ•ฉ์ „๊ณต ๊ต์ˆ˜



์ฝ˜ํฌ๋ฆฌํŠธ, ๊ณต๊ทน๋ฅ , ์˜์ƒ๋ถ„ํ• , ๋จธ์‹ ๋Ÿฌ๋‹, ๋”ฅ๋Ÿฌ๋‹
Concrete, Porosity, Image segmentation, Machine learning, Deep learning

1. ์„œ ๋ก 

์‚ฌ์šฉ์—ฐ์ˆ˜ 30๋…„์ด ์ดˆ๊ณผ๋œ ์ฒ ๊ทผ์ฝ˜ํฌ๋ฆฌํŠธ ๊ตฌ์กฐ๋ฌผ(์ดํ•˜ RC ๊ตฌ์กฐ๋ฌผ)์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๊ตฌ์กฐ๋ฌผ์˜ ๋‚ด๊ตฌ์„ฑ ํ‰๊ฐ€ ๋ฐ ์œ ์ง€๊ด€๋ฆฌ์— ๋Œ€ํ•œ ์‚ฌํšŒ์  ๊ด€์‹ฌ์ด ๋†’์•„์ง€๊ณ  ์žˆ๋‹ค. ์‹œ์„ค๋ฌผ ์œ ์ง€๊ด€๋ฆฌ ์ง€์นจ(Ministry of Land, Infrastructure and Transport, 2022)์—์„œ RC๊ตฌ์กฐ๋ฌผ์˜ ์ฃผ๋œ ๋‚ด๊ตฌ์„ฑ ํ‰๊ฐ€ํ•ญ๋ชฉ์€ ํƒ„์‚ฐํ™” ๊นŠ์ด ๋ฐ ์—ผํ™”๋ฌผ ์นจํˆฌ๋Ÿ‰์ด๋‹ค. ํƒ„์‚ฐํ™” ๋ฐ ์—ผํ™”๋ฌผ ์นจํˆฌ ์†๋„๋Š” ์ฝ˜ํฌ๋ฆฌํŠธ ๊ณต๊ทน๋ฅ ์ด ๋†’์•„์งˆ์ˆ˜๋ก ์ฆ๊ฐ€ํ•˜๊ธฐ ๋•Œ๋ฌธ์— RC ๊ตฌ์กฐ๋ฌผ์˜ ๋‚ด๊ตฌ์„ฑ์€ ์ฝ˜ํฌ๋ฆฌํŠธ ๊ณต๊ทน๋ฅ ์— ํฐ ์˜ํ–ฅ์„ ๋ฐ›๋Š”๋‹ค(Chang & Chen, 2006). ๊ณต๊ทน๋ฅ ์€ ์ˆ˜๋ถ„ ํก์ˆ˜๋ฒ•, ๊ฐ€์Šคํก์ฐฉ๋ฒ• ๋ฐ X-ray Microscope (XRM) ๋“ฑ์„ ํ†ตํ•ด ๊ณ„์ธกํ•  ์ˆ˜ ์žˆ๋‹ค. ์ˆ˜๋ถ„ ํก์ˆ˜๋ฒ• (KS F 2385, 2018)์€ ์ฝ˜ํฌ๋ฆฌํŠธ์— ํก์ˆ˜๋œ ์ˆ˜๋ถ„๋Ÿ‰์„ ํ†ตํ•ด ๊ณต๊ทน๋ฅ ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋ฉฐ, ๊ฐ€์Šคํก์ฐฉ๋ฒ•์€ ๋‚ด๋ถ€ ํ‘œ๋ฉด์ ์— ํก์ฐฉ๋œ ๊ธฐ์ฒด๋Ÿ‰์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ๊ณต๊ทน๋ฅ ์„ ์ถ”์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค. X-ray Microscope (XRM) ์€ X-ray๋ฅผ ํ†ตํ•ด ๋‚ด๋ถ€ ๋ฏธ์„ธ๊ตฌ์กฐ๋ฅผ ๊ฐ€์‹œํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•์ด๋ฉฐ, 3์ฐจ์›์œผ๋กœ ๊ณต๊ทน ๋„คํŠธ์›Œํฌ(Pore network)๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ์žฅ์ ์ด ์žˆ๋‹ค(Lu et al., 2006). ํ•œํŽธ, ์ด๋Ÿฌํ•œ ์ธก์ •๋ฒ•๋“ค์€ ํŒŒ๊ดด ์‹œํ—˜๋ฒ•์ด๊ธฐ ๋•Œ๋ฌธ์—, ๊ณ„์ธก ๋น„์šฉ์ด ๋†’๊ณ  ๋ฒˆ๊ฑฐ๋กญ๋‹ค๋Š” ํ•œ๊ณ„๋ฅผ ์ง€๋‹ˆ๊ณ  ์žˆ๋‹ค(Torres-Luque et al., 2014).

์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•์€ ์ปดํ“จํ„ฐ๊ฐ€ ์ด๋ฏธ์ง€ ๋‚ด์˜ ๊ฐ์ฒด๋ฅผ ํ”ฝ์…€ ๋‹จ์œ„๋กœ ๋‚˜๋ˆ„์–ด์ฃผ๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ, ์˜๋ฃŒ, ์ž์œจ์ฃผํ–‰ ๋ฐ ๊ฑด์„ค ๋“ฑ ๋‹ค์–‘ํ•œ ์‚ฐ์—… ๋ถ„์•ผ์—์„œ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค(Hofmarcher et al., 2019; Wang et al., 2018; Dogan et al., 2017; Jang et al., 2019; Yang et al., 2020). ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•์€ ๊ฑด์„ค ๋ถ„์•ผ์—์„œ ์ฃผ๋กœ ๊ท ์—ด ๊ฒ€์ถœ์— ํ™œ์šฉ๋˜์–ด์™”์œผ๋ฉฐ, ์ด๋Š” ๊ท ์—ด ํƒ์ง€์— ์†Œ์š”๋˜๋Š” ๋น„์šฉ ๋ฐ ์‹œ๊ฐ„์„ ํฌ๊ฒŒ ๊ฐ์ถ•ํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค(Dung, 2019). ์ด๋Ÿฌํ•œ ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•์„ ๊ณต๊ทน๋ฅ  ์˜ˆ์ธก์— ํ™œ์šฉํ•œ๋‹ค๋ฉด, ์„ ํ–‰ ๊ณต๊ทน๋ฅ  ์ธก์ •๋ฒ•์˜ ํ•œ๊ณ„๋ฅผ ๊ทน๋ณตํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์œผ๋กœ ๊ธฐ๋Œ€๋œ๋‹ค.

๋”ฐ๋ผ์„œ, ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ์ฝ˜ํฌ๋ฆฌํŠธ์˜ ๊ณต๊ทน๋ฅ ์„ ์˜ˆ์ธกํ•˜๊ณ ์ž ํ•˜์˜€๋‹ค. 2์žฅ์—์„œ๋Š” ๋ณธ ์—ฐ๊ตฌ์—์„œ ํ™œ์šฉ๋œ ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•๊ณผ ์‚ฌ์šฉ ์žฅ๋น„๋ฅผ ์†Œ๊ฐœํ•˜์˜€๊ณ , 3์žฅ์—์„œ๋Š” ์ด๋ฏธ์ง€ ์ทจ๋“, ๊ณต๊ทน๋ฅ  ๊ณ„์ธก ๋ฐ ๋ชจ๋ธํ•™์Šต์— ์‚ฌ์šฉ๋œ ๋ฐฉ๋ฒ•๋“ค์„ ์„œ์ˆ ํ•˜์˜€๋‹ค. 4์žฅ์—์„œ๋Š” ์‹คํ—˜ ๊ฒฐ๊ณผ๋ฅผ ์š”์•ฝํ•˜์˜€์œผ๋ฉฐ, 5์žฅ์—์„œ๋Š” ๋ณธ์—ฐ๊ตฌ์˜ ๊ฒฐ๋ก ์„ ์š”์•ฝํ•˜์˜€๋‹ค.

2. ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ• ๋ฐ ์žฅ๋น„

์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•์€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ๊ฐ์ฒด๋ฅผ ์ธ์‹ํ•˜์—ฌ ํ”ฝ์…€ ๋‹จ์œ„๋กœ ๊ฐ์ฒด๋ฅผ ๋ถ„๋ฅ˜ํ•˜๋Š” ๊ธฐ๋ฒ•์ด๋‹ค. ์ดˆ๊ธฐ์˜ ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•์€ ์ž„๊ณ„๊ฐ’์„ ์ •ํ•˜๊ณ  ์ž„๊ณ„๊ฐ’์„ ๊ธฐ์ค€์œผ๋กœ ์ด๋ฏธ์ง€ ๋‚ด์˜ ๊ฐ์ฒด๋ฅผ ๋ถ„๋ฅ˜ํ•˜๊ฑฐ๋‚˜(Otsu, 1979), ๋น„์Šทํ•œ ์ƒ‰๊น”์„ ๊ตฐ์ง‘ํ™”ํ•˜์—ฌ ๊ฐ์ฒด๋ฅผ ๋ถ„๋ฅ˜ (Dhanachandra et al., 2015)ํ•˜์˜€๋‹ค. ์ด๋Ÿฌํ•œ ๋ฐฉ์‹๋“ค์€ ๋ถ„๋ฅ˜์ •ํ™•๋„๊ฐ€ ๋‚ฎ์•˜์—ˆ์ง€๋งŒ, ์ตœ๊ทผ์—๋Š” ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ์ˆ ์ด ๊ธ‰์†ํ•œ ๋ฐœ์ „์„ ์ด๋ฃจ๋ฉด์„œ ๋†’์€ ์ •ํ™•๋„๋ฅผ ๋ณด์ด๋Š” ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ• ๋ชจ๋ธ๋“ค์ด ๋‹ค์ˆ˜ ๊ฐœ๋ฐœ๋˜์—ˆ๋‹ค.

2.1 ๋”ฅ๋Ÿฌ๋‹ ๊ธฐ๋ฐ˜ ์˜์ƒ ๋ถ„ํ• ๋ชจ๋ธ

2.1.1 Fully Convolutional Network (FCN)

Fig. 1์™€ ๊ฐ™์ด FCN์€ ์ด๋ฏธ์ง€์—์„œ ๋Œ€์ƒ์˜ ํŠน์ง•์„ ์ถ”์ถœํ•˜๋Š” ๋‹ค์šด์ƒ˜ํ”Œ๋ง(Down-sampling) ๊ณผ์ •๊ณผ ๋Œ€์ƒ์˜ ์œ„์น˜๋ฅผ ์ฐพ๋Š” ์—…์ƒ˜ํ”Œ๋ง(Up-sampling) ๊ณผ์ •์œผ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค(Long et al., 2015). ๋‹ค์šด์ƒ˜ํ”Œ๋ง ๊ณผ์ •์—์„œ๋Š” Fig. 2์— ๋‚˜ํƒ€๋‚ธ ๋ฐ”์™€ ๊ฐ™์ด ์ปจ๋ณผ๋ฃจ์…˜ ๋ ˆ์ด์–ด(Convolution layer)๊ฐ€ ์ด๋ฏธ์ง€์˜ ํŠน์ง• ์ •๋ณด๋ฅผ ์ถ”์ถœํ•˜๊ณ , ์ฃผ์š” ์ •๋ณด๋งŒ์„ ํ’€๋ง(Pooling) ๊ณผ์ •์œผ๋กœ ์ €์žฅํ•˜๊ฒŒ ๋œ๋‹ค. ์ด ๊ณผ์ •์„ ์—ฌ๋Ÿฌ ๋ฒˆ ๋ฐ˜๋ณตํ•˜์—ฌ ์–ป์€ ์ •๋ณด๋ฅผ ํ†ตํ•ด ๋Œ€์ƒ์˜ ํŠน์ง•์„ ํŒŒ์•…ํ•˜๊ณ , ์ด๋ฅผ ์—…์ƒ˜ํ”Œ๋ง ๊ณผ์ •์œผ๋กœ ์ด์ „์‹œํ‚จ๋‹ค. ์—…์ƒ˜ํ”Œ๋ง ๊ณผ์ •์—์„œ๋Š” ํ•ด๋‹น ํŠน์ง•์˜ ์œ„์น˜๋ฅผ ์–ธํ’€๋ง(Unpooling)ํ•˜์—ฌ ํŠน์ •ํ•จ์œผ๋กœ์จ ๋Œ€์ƒ์˜ ์œ„์น˜์ •๋ณด๋ฅผ ํŒŒ์•…ํ•œ๋‹ค. FCN์€ ์—…์ƒ˜ํ”Œ๋ง์—์„œ ๋ฐœ์ƒํ•˜๋Š” ์œ„์น˜์ •๋ณด์˜ ์†์‹ค ์ตœ์†Œํ™”ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ๋‹ค์šด์ƒ˜ํ”Œ๋ง์—์„œ ์ปจ๋ณผ๋ฃจ์…˜ ๋ ˆ์ด์–ด์˜ ์ผ๋ถ€๋ฅผ ์ฐธ์กฐํ•œ๋‹ค.

Fig. 1 Structure of FCN
../../Resources/ksm/jksmi.2023.27.1.30/fig1.png
Fig. 2 Convolution layer
../../Resources/ksm/jksmi.2023.27.1.30/fig2.png

2.1.2 U-net

Fig. 3๊ณผ ๊ฐ™์ด U-net ๋˜ํ•œ ๋‹ค์šด์ƒ˜ํ”Œ๋ง ๊ณผ์ •๊ณผ ์—…์ƒ˜ํ”Œ๋ง ๊ณผ์ •์œผ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค(Ronneberger et al., 2015). ์ด ๋ฐฉ๋ฒ•์€ ๋ชจ๋“  ์—…์ƒ˜ํ”Œ๋ง ๋ ˆ์ด์–ด์—์„œ ๋‹ค์šด์ƒ˜ํ”Œ๋ง ๋‹จ๊ณ„์˜ ์ปจ๋ณผ๋ฃจ์…˜ ๋ ˆ์ด์–ด๋ฅผ ์ฐธ์กฐํ•˜๊ธฐ ๋•Œ๋ฌธ์— ์ผ๋ฐ˜์ ์œผ๋กœ FCN๋Œ€๋น„ ๋†’์€ ๋ถ„๋ฅ˜์ •ํ™•๋„๋ฅผ ๊ฐ–๋Š”๋‹ค. ๋˜ํ•œ ์ปจ๋ณผ๋ฃจ์…˜ ๋ ˆ์ด์–ด์—์„œ์˜ ์ค‘๋ณต ๊ฒ€์ฆ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜์—ฌ ํ•™์Šต ์†๋„๊ฐ€ ๋น ๋ฅด๋‹ค๋Š” ํŠน์ง•์„ ๊ฐ–๋Š”๋‹ค.

Fig. 3 Structure of U-net
../../Resources/ksm/jksmi.2023.27.1.30/fig3.png

2.1.3 DeepLab v3+

DeepLab v3+๋Š” ๊ธฐ๋ณธ์ ์œผ๋กœ FCN๋ชจ๋ธ๊ณผ ์œ ์‚ฌํ•œ ๊ตฌ์กฐ๋กœ ๋˜์–ด ์žˆ์ง€๋งŒ, Fig. 4์— ๋‚˜ํƒ€๋‚œ ๋ฐ”์™€ ๊ฐ™์ด ๊ณต๊ฐ„ ํ”ผ๋ผ๋ฏธ๋“œ ํ’€๋ง (Spatial Pyramid pooling) ๊ธฐ๋ฒ•์ด ํ™œ์šฉ๋œ๋‹ค(Chen et al., 2017). ์ด ๊ธฐ๋ฒ•์€ ๋นˆ ๊ณต๊ฐ„์„ ๊ฐ–๋Š” ํ•„ํ„ฐ๋ฅผ ํ™œ์šฉํ•˜๋ฉฐ, ๋น„์œจ (rate)์˜ ๊ฐ’์— ๋”ฐ๋ผ ๋นˆ ๊ณต๊ฐ„์˜ ํฌ๊ธฐ๊ฐ€ ๊ฒฐ์ •๋œ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, rate=1์ผ ๊ฒฝ์šฐ ํ•„ํ„ฐ์— ๋นˆ ๊ณต๊ฐ„์ด ์—†์œผ๋ฉฐ, ๋น„์œจ์ด ์ปค์งˆ์ˆ˜๋ก ๋นˆ ๊ณต๊ฐ„์ด ๋„“์–ด์ง€๊ฒŒ ๋œ๋‹ค. ์ด๋ฅผ ํ†ตํ•ด Fig. 5์™€ ๊ฐ™์ด ํ•„ํ„ฐ๋ฅผ ์ด์šฉํ•œ ๊ณ„์‚ฐ๋Ÿ‰์„ ์ ๊ฒŒ ์œ ์ง€ํ•˜๋ฉด์„œ๋„ ํ•œ ํ”ฝ์…€๋กœ ๋ณผ ์ˆ˜ ์žˆ๋Š” ์˜์—ญ์„ ๋„“๊ฒŒ ๊ฐ€์ ธ๊ฐˆ ์ˆ˜ ์žˆ๋‹ค.

Fig. 4 Spatial pyramid pooling
../../Resources/ksm/jksmi.2023.27.1.30/fig4.png
Fig. 5 Atrous convolution
../../Resources/ksm/jksmi.2023.27.1.30/fig5.png

2.1.4 MoblieNet

Howard et al.(2017)๋Š” ๊ธฐ์กด์˜ ์ปจ๋ณผ๋ฃจ์…˜(convolution)์—์„œ ํ•œ ์ถ•์˜ ์—ฐ์‚ฐ์„ ๊ฐ์†Œ์‹œํ‚จ Depth-wise separable convolution (DWSC) ์„ ํ™œ์šฉํ•˜์—ฌ, ์ €์‚ฌ์–‘ ์ปดํ“จํŒ… ๊ธฐ๊ธฐ์— ์ ์šฉํ•  ์ˆ˜ ์žˆ๋Š” ์˜์ƒ ๋ถ„ํ•  ์•Œ๊ณ ๋ฆฌ์ฆ˜(MobileNet)์„ ๊ฐœ๋ฐœํ•˜์˜€๋‹ค(Fig. 6). ์ด ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ๋‹ค๋ฅธ ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ• ๋Œ€๋น„ ๋‚ฎ์€ ๋ถ„๋ฅ˜์„ฑ๋Šฅ์„ ๊ฐ–์ง€๋งŒ, ์šฐ์ˆ˜ํ•œ ์—ฐ์‚ฐํšจ์œจ๋กœ ์ธํ•ด ์Šค๋งˆํŠธํฐ ๋“ฑ์— ์ฃผ๋กœ ํ™œ์šฉ๋˜๊ณ  ์žˆ๋‹ค.

2.2 ์‹คํ—˜์žฅ๋น„

Fig. 6 MobileNet
../../Resources/ksm/jksmi.2023.27.1.30/fig6.png

์ฝ˜ํฌ๋ฆฌํŠธ์˜ ํ‘œ๋ฉด ์ด๋ฏธ์ง€๋ฅผ ์ทจ๋“ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ๊ด‘ํ•™ํ˜„๋ฏธ๊ฒฝ(HT004, Gasworld)์„ ํ™œ์šฉํ•˜์˜€๋‹ค. ์ฝ˜ํฌ๋ฆฌํŠธ์˜ ์‹ค์ œ ๊ณต๊ทน๋ฅ ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ X-ray Microscope (XRM, Xradia 620 Versa, Carl Zeiss)์„ ํ™œ์šฉํ•˜์˜€์œผ๋ฉฐ, Dragonfly Pro๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ XRM๋ฐ์ดํ„ฐ๋ฅผ 3์ฐจ์›์œผ๋กœ ๊ฐ€์‹œํ™”ํ•˜์˜€๋‹ค. ๋น„ํŒŒ๊ดด ์‹œํ—˜๋ฒ•์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ์ฝ˜ํฌ๋ฆฌํŠธ์˜ ์••์ถ•๊ฐ•๋„๋ฅผ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ Concrete Test & Surveyor (CTS, HJ-CTS-02v04, Heungjin)๋ฅผ ํ™œ์šฉํ•˜์˜€๋‹ค.

3. ์‹คํ—˜๋ฐฉ๋ฒ•

3.1 ์‹œํŽธ ์ œ์ž‘

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ฝ˜ํฌ๋ฆฌํŠธ์˜ ํ‘œ๋ฉด ์ด๋ฏธ์ง€๋ฅผ ์ˆ˜์ง‘ํ•˜๊ธฐ ์œ„ํ•ด ๋ฌผ์‹œ๋ฉ˜ํŠธ๋น„๋ฅผ ๋ณ€์ˆ˜๋กœ ํ•˜๋Š” 3์ข…๋ฅ˜์˜ ์ฝ˜ํฌ๋ฆฌํŠธ ์‹คํ—˜์ฒด๋“ค์„ ์ œ์ž‘ํ•˜์˜€๋‹ค. Table 1์—๋Š” ๊ฐ ์‹คํ—˜์ฒด์˜ ์ฝ˜ํฌ๋ฆฌํŠธ ๋ฐฐํ•ฉ๋น„๋ฅผ ๋‚˜ํƒ€๋‚ด์—ˆ์œผ๋ฉฐ, ํ•œ ๋ณ€์˜ ๊ธธ์ด๊ฐ€ 50mm์ธ ์ •์œก๋ฉด์ฒด ์‹คํ—˜์ฒด๋ฅผ ๋ฐฐํ•ฉ๋‹น 30๊ฐœ์”ฉ, ์ด 90๊ฐœ๋ฅผ ์ œ์ž‘ํ•˜์˜€๋‹ค(Fig. 7). ๋ชจ๋“  ์‹คํ—˜์ฒด๋Š” 28์ผ๊ฐ„ ์ˆ˜์ค‘์–‘์ƒ ๋˜์—ˆ๋‹ค.

Table 1 Summary of concrete mix-proportion

Type

water

cement

fine aggregate

coarse aggregate

Type1

175

325

910

921

Type2

162

460

785

970

Type3

177

600

681

902

๋‹จ์œ„ : kg/mยณ
Fig. 7 Specimen
../../Resources/ksm/jksmi.2023.27.1.30/fig7.png

3.2 ์ด๋ฏธ์ง€ ์ทจ๋“ ๋ฐ ์˜ˆ์ธก ๊ณต๊ทน๋ฅ  ์‚ฐ์ •

86๋ฐฐ์œจ ๋ฐ 500๋งŒ ํ™”์†Œ์˜ ๊ด‘ํ•™ํ˜„๋ฏธ๊ฒฝ์„ ํ™œ์šฉํ•˜์—ฌ ํ‘œ๋ฉด์ฒ˜๋ฆฌ๋ฅผ ๊ฑฐ์น˜์ง€ ์•Š์€ ์‹คํ—˜์ฒด๋“ค์˜ ํ‘œ๋ฉด ์ด๋ฏธ์ง€๋ฅผ ์ทจ๋“ํ•˜์˜€๋‹ค. ํ‘œ๋ฉด๊ณต๊ทน๋ฅ ์˜ ๋ณ€๋™์„ฑ์„ ๊ณ ๋ คํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ๋ชจ๋“  ์‹คํ—˜์ฒด(90๊ฐœ)๋ฅผ ์ด๋ฏธ์ง€ ์ทจ๋“์— ํ™œ์šฉํ•˜์˜€๋‹ค. ๋‹ค๋งŒ, ๊ฐ ์‹œํŽธ์˜ 6๊ฐœ ๋ฉด ์ค‘, ๊ฑฐํ‘ธ์ง‘๊ณผ ๋งž๋‹ฟ์€ 5๊ฐœ์˜ ๋ฉด์— ๋Œ€ํ•ด์„œ๋งŒ ์ดฌ์˜์ด ์ด๋ฃจ์–ด์กŒ์œผ๋ฉฐ, ๊ฐ ๋ฉด๋‹น 6๊ฐœ์˜ ๊ตฌ์—ญ์„ ๋‚˜๋ˆ„์–ด ํ‘œ๋ฉด ์ด๋ฏธ์ง€๋ฅผ ์ดฌ์˜ํ•˜์˜€๋‹ค. ๊ทธ ๊ฒฐ๊ณผ, ์ด 2,729์žฅ(90๊ฐœร—5๋ฉดร—6๊ตฌ์—ญ + 29์žฅ์˜ ์ถ”๊ฐ€ ์ด๋ฏธ์ง€)์˜ ์ด๋ฏธ์ง€๋ฅผ ์ทจ๋“ํ•˜์˜€๋‹ค. ์ด ์ค‘ ์ด 139๊ฐœ ์ด๋ฏธ์ง€์˜ ๊ณต๊ทน์„ ๋งˆ์Šคํ‚นํ•˜์—ฌ ๋ชจ๋ธํ•™์Šต, ๊ฒ€์ฆ ๋ฐ ์‹œํ—˜(model training, validation and testing)์— ํ™œ์šฉํ•˜์˜€์œผ๋ฉฐ, ๋‚˜๋จธ์ง€ 2,590์žฅ์˜ ์ด๋ฏธ์ง€๋Š” ๋ฐฐํ•ฉ ๋ณ„ ๊ณต๊ทน๋ฅ ์„ ํ‰๊ฐ€ํ•˜๋Š” ๋ฐ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. ์—ฌ๊ธฐ์„œ ์˜ˆ์ธก ๊ณต๊ทน๋ฅ ์€ ์ „์ฒด ์ด๋ฏธ์ง€ ๋ฉด์  ์ค‘ ๊ณต๊ทน์ด ์ฐจ์ง€ํ•˜๋Š” ๋ฉด์ ์˜ ๋น„์œจ๋กœ ์‚ฐ์ •๋˜์—ˆ๋‹ค. Fig. 8 (a) ๋ฐ (b)๋Š” ๊ฐ๊ฐ ๊ด‘ํ•™ํ˜„๋ฏธ๊ฒฝ์œผ๋กœ ์ดฌ์˜๋œ ์‹คํ—˜์ฒด์˜ ํ‘œ๋ฉด ์ด๋ฏธ์ง€์™€ ๋งˆ์Šคํ‚น ๋œ ์ด๋ฏธ์ง€๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.

Fig. 8 Surface images extracted by microscope
../../Resources/ksm/jksmi.2023.27.1.30/fig8.png

3.3 ์ธก์ • ๊ณต๊ทน๋ฅ  ์‚ฐ์ •

์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•์„ ํ†ตํ•ด ๋„์ถœํ•œ ์˜ˆ์ธก ๊ณต๊ทน๋ฅ ์˜ ์ •ํ™•๋„๋ฅผ ํ‰๊ฐ€ํ•˜๊ณ ์ž, XRM ๋ฐ ์ˆ˜๋ถ„ ํก์ˆ˜๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ๊ณต๊ทน๋ฅ ์„ ๊ณ„์ธกํ•˜์˜€๋‹ค. ๊ฐ ๋ฐฐํ•ฉ ๋ณ„ ์‹คํ—˜์ฒด๋ฅผ 10ร—10ร—10 mm3 ํฌ๊ธฐ๋กœ ์ ˆ๋‹จํ•˜์—ฌ XRM ์‹œํŽธ์„ ์ค€๋น„ํ•˜์˜€์œผ๋ฉฐ, Fig. 9๋Š” XRM (X-Ray Microscopy)์„ ํ†ตํ•ด ์ดฌ์˜๋œ ์‹œํŽธ์˜ 3์ฐจ์› ๊ณต๊ทน ๊ตฌ์กฐ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. ์ด๋•Œ, ๊ณต๊ทน ๊ตฌ์กฐ๋Š” Dragonfly Pro์˜ Threshold๋ฅผ ์กฐ์ ˆํ•˜๋ฉฐ ๊ณต๊ทน์„ ๊ตฌ๋ถ„ํ•˜๊ณ  ์ดฌ์˜ํ•œ ๊ฒƒ์ด๋‹ค. ์ œ์•ˆ ๋ฐฉ๋ฒ•์€ ๊ตต์€๊ณจ์žฌ๊ฐ€ ๋ณด์ด์ง€ ์•Š๋Š” ํ‘œ๋ฉด ์ด๋ฏธ์ง€๋ฅผ ๋ฐ”ํƒ•์œผ๋กœ ์˜ˆ์ธก ๊ณต๊ทน๋ฅ ์„ ์‚ฐ์ •ํ•˜๊ธฐ ๋•Œ๋ฌธ์—, ๊ตต์€๊ณจ์žฌ๊ฐ€ ์—†๋Š” 3์ฐจ์› ์˜์—ญ์˜ XRM ๊ฒฐ๊ณผ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์ธก์ • ๊ณต๊ทน๋ฅ ์„ ์‚ฐ์ •ํ•˜์˜€๋‹ค.

์ˆ˜๋ถ„ ํก์ˆ˜๋ฒ•(Ahn et al., 2013)์œผ๋กœ ๊ณต๊ทน๋ฅ ์„ ์ธก์ •ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ์‹คํ—˜์ฒด๋ฅผ 100ยฐC ํ™˜๊ฒฝ์—์„œ 24์‹œ๊ฐ„ ๊ฑด์กฐํ•œ ๋’ค, 1์ฃผ์ผ๊ฐ„ ์ˆ˜์ค‘์— ๋‘์–ด ์‹คํ—˜์ฒด์— ์ˆ˜๋ถ„์ด ์ถฉ๋ถ„ํžˆ ํก์ˆ˜๋  ์ˆ˜ ์žˆ๊ฒŒ ํ•˜์˜€๋‹ค. ์ดํ›„ ์‹คํ—˜์ฒด์˜ ๋ณ€ํ™”๋œ ์ค‘๋Ÿ‰์„ ํ†ตํ•ด ๊ณต๊ทน๋ฅ ($P$)์„

(1)
$P(\%)=1-\dfrac{V_{sol}}{V_{T}}=\dfrac{(W_{S}-W_{D})/\gamma_{W}}{V_{T}}$

์œผ๋กœ ์‚ฐ์ •ํ•˜์˜€์œผ๋ฉฐ, ์—ฌ๊ธฐ์„œ, P๋Š” ๋ฐฑ๋ถ„์œจ๋กœ ๋‚˜ํƒ€๋‚ธ ์‹œ๋ฃŒ์˜ ๊ณต๊ทน๋ฅ , $V_{sol}$๋Š” ๊ณต๊ทน์„ ์ œ์™ธํ•œ ์‹œ๋ฃŒ์˜ ๋ถ€ํ”ผ, $V_{T}$๋Š” ์‹œ๋ฃŒ์˜ ๋ถ€ํ”ผ, $W_{S}$๋Š” ์‹œ๋ฃŒ์˜ ์ˆ˜์ค‘์ค‘๋Ÿ‰, $W_{D}$๋Š” ์‹œ๋ฃŒ์˜ ๊ฑด์กฐ์ค‘๋Ÿ‰, $\gamma_{W}$๋Š” ๋ฌผ์˜ ๋‹จ์œ„์ค‘๋Ÿ‰์„ ๋‚˜ํƒ€๋‚ธ๋‹ค.

Fig. 9 3-Dimensionalized X-ray images (Type 1)
../../Resources/ksm/jksmi.2023.27.1.30/fig9.png

3.4 ์••์ถ•๊ฐ•๋„ ํ‰๊ฐ€

์••์ถ•๊ฐ•๋„๋Š” ๊ณต๊ทน๋ฅ ๊ณผ ๊ฐ•ํ•œ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ๊ฐ–๋Š” ๋Œ€์ฒด์ธ์ž ์ด๋ฏ€๋กœ(Kumar et al., 2003), CTS(๋น„ํŒŒ๊ดด ์‹œํ—˜, Nitto, 2009)์™€ Universal Testing Machine(UTM, ํŒŒ๊ดด ์‹œํ—˜, KS F 2405, 2022)์„ ํ™œ์šฉํ•˜์—ฌ ์ฝ˜ํฌ๋ฆฌํŠธ์˜ ์••์ถ•๊ฐ•๋„๋ฅผ ์ธก์ •ํ•˜์˜€๋‹ค. ์••์ถ•๊ฐ•๋„ ์ธก์ •์€ Type ๋ณ„๋กœ ๊ฐ๊ฐ ์ง€๋ฆ„์ด 100 mm์ด๊ณ , ๋†’์ด๊ฐ€ 200 mm์ธ ์›์ฃผํ˜• ๊ณต์‹œ์ฒด 3๊ฐœ์— ๋Œ€ํ•ด ์‹ค์‹œ๋˜์—ˆ๋‹ค. CTS๋ฅผ ํ™œ์šฉํ•  ๊ฒฝ์šฐ, ์›์ฃผํ˜• ๊ณต์‹œ์ฒด๋ฅผ 20ํšŒ ํƒ€๊ฒฉํ•˜์—ฌ ๋„์ถœ๋œ ๋ฐ˜๋ฐœ ๊ฒฝ๋„ ๊ฐ’์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ์••์ถ•๊ฐ•๋„๋ฅผ ์‚ฐ์ •ํ•˜์˜€๋‹ค.

3.5 ๋ชจ๋ธํ•™์Šต

3.5.1 ์ฆ๊ฐ•(Augmentation)

์ปดํ“จํ„ฐ ๋น„์ „์€ ์ด๋ฏธ์ง€๋ฅผ ์ˆ˜์น˜ํ™”ํ•˜์—ฌ ์‚ฌ์šฉํ•˜๋ฏ€๋กœ, ํšŒ์ „, ๋ฐ˜์ „ ๋ฐ ํ™•๋Œ€๋œ ๊ฐ์ฒด๋ฅผ ๋‹ค๋ฅธ ๊ฐ์ฒด๋กœ ์ธ์‹ํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด์— ๋”ฐ๋ผ ์ด๋ฏธ์ง€์˜ ํšŒ์ „๊ณผ ๋ฐ˜์ „, ํ™•๋Œ€ ๋“ฑ์˜ ๋ฐ์ดํ„ฐ ์ฆ๊ฐ•๊ณผ์ •์„ ํ†ตํ•ด ์ถ”๊ฐ€ ์ด๋ฏธ์ง€๋ฅผ ํ™•๋ณดํ•˜๊ณ , ํ•™์Šต์— ํ™œ์šฉํ•˜์˜€๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ์ด ๋‹ค์–‘ํ•œ ํฌ๊ธฐ ๋ฐ ํ•ด์ƒ๋„์˜ ์ด๋ฏธ์ง€๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ์ƒ์„ฑ์ž(generator)๋ฅผ ๊ฑฐ์ณ ์ด๋ฏธ์ง€์˜ ํฌ๊ธฐ๋ฅผ 256ร—256์œผ๋กœ ๋ณ€ํ˜•ํ•˜์˜€๋‹ค. ์ดํ›„ ์ „๋ฐ˜์ ์ธ ๋ชจ๋ธ๋„์ถœ๊ณผ์ •์€ Fig. 10์— ๋„์‹ํ™”ํ•˜์˜€๋‹ค.

Fig. 10 Flowchart for model derivation
../../Resources/ksm/jksmi.2023.27.1.30/fig10.png

3.5.2 ๋ชจ๋ธ ์„ ํƒ

ํ•™์Šต์— ์‚ฌ์šฉ๋  ์ตœ์ ์˜ ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ(Deep Learning Model)์„ ๋„์ถœํ•˜๊ธฐ ์œ„ํ•˜์—ฌ FCN, MobileNet, DeepLab v3+ ๋ฐ U-net์„ ํ™œ์šฉํ•˜์—ฌ ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ์—์„œ ๊ฐ ๋ชจ๋ธ์˜ ํ•™์Šต์—๋Š” default parameter๊ฐ€ ์‚ฌ์šฉ๋˜์—ˆ์œผ๋ฉฐ, 30์žฅ์˜ ์ด๋ฏธ์ง€๋ฅผ ํ™œ์šฉํ•˜์—ฌ ํ•™์Šต์ด ์ˆ˜ํ–‰๋˜์—ˆ๋‹ค. Table 2๋Š” ๋ฒค์น˜๋งˆํฌ ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ๋ฅผ ๋‚˜ํƒ€๋‚ธ ๊ฒƒ์ด๋‹ค.

ํ…Œ์ŠคํŠธ ๊ฒฐ๊ณผ, ํ•™์Šต ์‹œ๊ฐ„์€ MobileNet์ด ๊ฐ€์žฅ ๋‚ฎ์•˜์œผ๋ฉฐ, ๊ณต๊ทน๋ฅ  ํ‰๊ฐ€ ์ •ํ™•๋„๋Š” DeepLab v3+๊ฐ€ ๊ฐ€์žฅ ๋†’์•˜๋‹ค. U-net์€ DeepLab v3+์™€ ์œ ์‚ฌํ•œ ์ˆ˜์ค€์˜ ์ •ํ™•๋„๋ฅผ ๋‚˜ํƒ€๋‚ด์—ˆ์ง€๋งŒ, ํ•™์Šต ์‹œ๊ฐ„์€ DeepLab v3+๋ณด๋‹ค 30%๊ฐ€๋Ÿ‰ ๋‚ฎ์•˜๋‹ค. ์ด์ฒ˜๋Ÿผ ๋ฏธ๋ฏธํ•œ ํ‰๊ฐ€์ •ํ™•๋„์˜ ์ฐจ์ด๋Š” ๋งค๊ฐœ๋ณ€์ˆ˜ ์ตœ์ ํ™”(Hyper parameter tunning)๋ฅผ ํ†ตํ•˜์—ฌ ๊ทน๋ณตํ•  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ, ํ•™์Šต ๋ฐ tunning์— ์†Œ์š”๋˜๋Š” ์‹œ๊ฐ„์ด ์งง์€ U-net์„ ํ™œ์šฉํ•˜์—ฌ ์ดํ›„ ์—ฐ๊ตฌ๋‹จ๊ณ„๋ฅผ ์ง„ํ–‰ํ•˜์˜€๋‹ค.

Table 2 Result of benchmark test

Models

FCN

MobileNet

DeepLab v3+

U-net

Time

1.1h

0.3h

0.7h

0.5h

Accuracy

93.0%

89.2%

96.1%

95.8%

3.5.3 ๋งค๊ฐœ๋ณ€์ˆ˜ ์ตœ์ ํ™”

๋ชจ๋ธ์˜ ์ตœ์ ํ™”๋œ ๋งค๊ฐœ๋ณ€์ˆ˜ ์กฐํ•ฉ์„ ์ฐพ๊ธฐ ์œ„ํ•˜์—ฌ Grid search๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค(Liashchynskyi, 2019). Grid search๋Š” batch size=[2, 4, 8, 16], filter=[5, 6, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 30], layer=[2, 3, 4, 5, 6, 7, 8, 9, 10]์— ๋Œ€ํ•ด ์ˆ˜ํ–‰๋˜์—ˆ์œผ๋ฉฐ, ์—ํฌํฌ(epoch)๋Š” 1,000์œผ๋กœ ์„ค์ •๋˜์—ˆ๋‹ค. Table 3์—๋Š” ๊ฐœ๋ฐœ๋ชจ๋ธ์˜ ์ตœ์ ํ™”๋œ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด์—ˆ์œผ๋ฉฐ, ๊ฐ€์žฅ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋‚˜ํƒ€๋‚ธ ๋งค๊ฐœ๋ณ€์ˆ˜์˜ ์กฐํ•ฉ์€ ๊ฐ๊ฐ batch size=2, filter=17, layer=3์ด์—ˆ๋‹ค.

Table 3 Grid search

Parameter

Value

Parameter

Value

Input shape

(256,256,3)

*Number of layers

3

Batch norm

True

Number of classes

1

Up-sampling mode

Deconvolut-ion

Dropout type

Spatial

Attention

True

Dropout rate

0.5

*Filters

17

Activation of output

Sigmoid

Optimizer

Adam

Epochs

1000

*Batch size

2

Shuffle

True

* Parameters for grid search

3.5.4 ๋ชจ๋ธํ•™์Šต

์•ž์„œ ๋„์ถœ๋œ ๋งค๊ฐœ๋ณ€์ˆ˜๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ํ•™์Šต์‹œ์ผฐ๋‹ค. ์ „์ฒด ๋ฐ์ดํ„ฐ ์ค‘ 60%๋Š” ํ•™์Šต, 20%๋Š” ๊ฒ€์ฆ, 20%๋Š” ์‹œํ—˜์— ํ™œ์šฉ๋˜์—ˆ๋‹ค. ๋ฐ์ดํ„ฐ์˜ ์ˆœ์„œ(sequence)๊ฐ€ ๋ชจ๋ธ์˜ ํ•™์Šต์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ๊ฒƒ์„ ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•ด ์—ํฌํฌ๋งˆ๋‹ค ํ•™์Šต๋ฐ์ดํ„ฐ์˜ ์ˆœ์„œ๋ฅผ ๋ฌด์ž‘์œ„ ๋ณ€๊ฒฝํ•˜์˜€๋‹ค. ๋˜ํ•œ ๋ชจ๋ธ์˜ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์„ ํ™•๋ณดํ•˜๊ณ ์ž early-stopping ์„ ํ†ตํ•ด ๊ณผ์ ํ•ฉ์„ ๋ฐฉ์ง€ํ•˜์˜€์œผ๋ฉฐ, ์ตœ๋Œ€ ํ•™์Šต ํšŸ์ˆ˜(epoch)๋Š” 1,000 ๋ฐ ๊ฒ€์ฆ์ •ํ™•๋„์˜ ๊ฐœ์„  ์‹คํŒจ ํ•œ๊ณ„(patience)๋Š” 10์œผ๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. ํ•™์Šต๋œ ๋ชจ๋ธ์˜ ๊ฒ€์ฆ์ •ํ™•๋„๋Š” 97%์˜€์œผ๋ฉฐ, ์‹œํ—˜์ •ํ™•๋„๋Š” 92%๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค.

4. ๊ฒฐ๊ณผ ๋ฐ ๊ณ ์ฐฐ

4.1 ๊ณต๊ทน๋ฅ 

Table 4์—๋Š” XRM, ์ˆ˜๋ถ„ ํก์ˆ˜๋ฒ•, U-net์„ ํ†ตํ•ด ๋„์ถœ๋œ ๊ณต๊ทน๋ฅ ์„ ๋น„๊ตํ•˜์˜€๋‹ค. ์„ธ ๊ฐ€์ง€ ์ธก์ •๋ฒ• ๋ชจ๋‘ ๋ฌผ์‹œ๋ฉ˜ํŠธ๋น„(w/c)๊ฐ€ ๋†’์„์ˆ˜๋ก, ๋†’์€ ๊ณต๊ทน๋ฅ ์„ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค. Type 1์— ๋Œ€ํ•ด์„œ๋Š” XRM๊ณผ U-net์ด ์œ ์‚ฌํ•œ ๊ณต๊ทน๋ฅ ์„ ๋‚˜ํƒ€๋‚ด์—ˆ์œผ๋‚˜, Type 2 ๋ฐ 3์— ๋Œ€ํ•ด์„œ๋Š” U-net์ด ๊ณต๊ทน๋ฅ ์„ ์กฐ๊ธˆ ๋” ๋‚ฎ๊ฒŒ ํ‰๊ฐ€ํ•˜๋Š” ๊ฒฝํ–ฅ์„ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค. ์ด๋Š” XRM ์— ๋น„ํ•ด ์ด๋ฏธ์ง€ ๊ธฐ๋ฐ˜ Unet์„ ํ™œ์šฉํ•˜์—ฌ ๊ณ„์ธกํ•  ์ˆ˜ ์žˆ๋Š” ์ตœ์†Œ๊ณต๊ทน์˜ ํฌ๊ธฐ(XRM: 0.5ฮผm, U-net: 20~30ฮผm)๊ฐ€ ํฌ๋ฏ€๋กœ, w/c๊ฐ€ ๊ฐ์†Œํ• ์ˆ˜๋ก U-net์ด ๊ด€์ธกํ•  ์ˆ˜ ์—†๋Š” ๊ณต๊ทน๋“ค์ด ๋งŽ์•„์ง€๊ธฐ ๋•Œ๋ฌธ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค. ์ˆ˜๋ถ„ ํก์ˆ˜๋ฒ•์€ XRM ๋ฐ U-net๋ณด๋‹ค ๋น„๊ต์  ๋†’์€ ๊ณต๊ทน๋ฅ ์„ ๋‚˜ํƒ€๋‚ด์—ˆ๋Š”๋ฐ, ์ด๋Š” XRM ๋ฐ U-net์ด ๊ด€์ธกํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ทน๋ณด๋‹ค ์ž‘์€ ๊ณต๊ทน(0.5ฮผm ์ดํ•˜์˜ capilary void ๋“ฑ)์—์„œ๋„ ์ˆ˜๋ถ„ํก์ˆ˜๊ฐ€ ๋ฐœ์ƒํ•˜๊ธฐ ๋•Œ๋ฌธ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค.

Table 4 Measured porosity

Methods

Type 1

Type 2

Type 3

XRM

9.26 %

3.96 %

2.20 %

Water absorption

20.90 %

16.67 %

16.15 %

U-net

9.28 %

3.10 %

1.75 %

4.2 ํ•„์š” ํ‘œ๋ณธํฌ๊ธฐ

ํ˜„์žฅ์—์„œ ์ทจ๋“ํ•œ ์ฝ˜ํฌ๋ฆฌํŠธ์˜ ํ‘œ๋ฉด ์ด๋ฏธ์ง€๋ฅผ ํ™œ์šฉํ•˜์—ฌ ์‹ ๋ขฐ์„ฑ ์žˆ๋Š” ๊ณต๊ทน๋ฅ ์„ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•ด, ํ†ต๊ณ„์  ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ํ•„์š”ํ•œ ํ‘œ๋ณธ์˜ ์ˆ˜๋ฅผ ์กฐ์‚ฌํ•˜์˜€๋‹ค. U-net๊ณผ 2,729๊ฐœ์˜ ์ด๋ฏธ์ง€๋ฅผ ํ™œ์šฉํ•˜์—ฌ, ์ด๋ฏธ์ง€๋งˆ๋‹ค ๊ณต๊ทน๋ฅ ์„ ๋„์ถœํ•˜์˜€์œผ๋ฉฐ, ๋„์ถœ ๊ฐ’์ด ์ •๊ทœ๋ถ„ํฌ๋ฅผ ๋”ฐ๋ฅด๋Š”์ง€ ํŒ๋‹จํ•˜๊ธฐ ์œ„ํ•˜์—ฌ Shapiro-Wilk test๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ๊ท€๋ฌด๊ฐ€์„ค์€ โ€˜์ •๊ทœ๋ถ„ํฌ๋ฅผ ๋”ฐ๋ฅธ๋‹ค.โ€™๋กœ ์„ค์ •๋˜์—ˆ์œผ๋ฉฐ, ์œ ์˜์ˆ˜์ค€์€ 0.05๋กœ ์„ค์ •ํ•˜์˜€๋‹ค. Table 5์— ๋‚˜ํƒ€๋‚ธ ๋ฐ”์™€ ๊ฐ™์ด, ๋ชจ๋“  ์ข…๋ฅ˜์˜ ์‹คํ—˜์ฒด์—์„œ ์œ ์˜ํ™•๋ฅ ์ด 0.05๋ณด๋‹ค ๋‚ฎ์•˜์œผ๋ฉฐ, ์ด๋Š” ์˜ˆ์ธก ๊ณต๊ทน๋ฅ ์ด ์ •๊ทœ๋ถ„ํฌ๋ฅผ ๋”ฐ๋ฅด์ง€ ์•Š๋Š” ๊ฒƒ์„ ์˜๋ฏธํ•œ๋‹ค. ๋”ฐ๋ผ์„œ, ๋‹ค์–‘ํ•œ ํ™•๋ฅ ๋ถ„ํฌ๋ฅผ ๋ชจ์‚ฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ฐ๋งˆ๋ถ„ํฌ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ชจํ‰๊ท  ์ถ”์ •์— ํ•„์š”ํ•œ ์ตœ์†Œ ํ‘œ๋ณธํฌ๊ธฐ๋ฅผ ์‚ฐ์ •ํ•˜์˜€๋‹ค.

ํ‘œ๋ณธํฌ๊ธฐ($n$)๋Š” ์•„๋ž˜์™€ ๊ฐ™์ด ์‚ฐ์ •๋˜์—ˆ์œผ๋ฉฐ,

(2)
$n=\dfrac{(Z\times\sigma)^{2}}{d^{2}}$

$Z$๋Š” Z-score, $\sigma$๋Š” ํ‘œ์ค€ํŽธ์ฐจ, $d$๋Š” ํ—ˆ์šฉ์˜ค์ฐจ ํฌ๊ธฐ๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค. Table 6์—๋Š” ์‹ ๋ขฐ์ˆ˜์ค€ (90%, 95%, 99%) ๋ฐ ํ—ˆ์šฉ ์˜ค์ฐจ(2%, 1%, 0.5%)์— ๋”ฐ๋ฅธ ํ•„์š” ํ‘œ๋ณธ ๊ฐœ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค. ์‹ ๋ขฐ์ˆ˜์ค€์„ 90%๋กœ ํ•˜๊ณ  ํ—ˆ์šฉ์˜ค์ฐจ๋ฅผ 1% ๋‚ด์™ธ๋กœ ํ•œ๋‹ค๋ฉด Type 1, 2, 3 ์‹คํ—˜์ฒด ๋ณ„๋กœ ํ•„์š”ํ•œ ์ด๋ฏธ์ง€ ์ˆ˜๋Š” ๊ฐ๊ฐ 34, 9, 4์žฅ์ด์—ˆ๋‹ค.

Table 5 Predicted porosity using image data

 

Type 1

Type 2

Type 3

Average

9.28 %

3.10 %

1.75 %

Median

9.22 %

2.65 %

1.53 %

P-value

1.79e-12

3.99e-26

3.23e-22

Table 6 Sample size

Confidence level

Margin of error

Type 1

Type 2

Type 3

90%

2%

9

3

1

1%

34

9

4

0.5%

136

36

15

95%

2%

12

4

2

1%

49

13

5

0.5%

193

51

21

99%

2%

21

6

3

1%

84

22

9

0.5%

333

87

35

4.3 ์••์ถ•๊ฐ•๋„

Table 7์—๋Š” CTS ํ•ด๋จธ์™€ UTM์„ ์ด์šฉํ•˜์—ฌ ์ธก์ •ํ•œ ์••์ถ•๊ฐ•๋„๋ฅผ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค. CTS ํ•ด๋จธ์˜ ์‹œํ—˜ ๊ฒฐ๊ณผ๋Š” UTM์„ ํ†ตํ•œ ์••์ถ•ํŒŒ๊ดด์‹œํ—˜๊ณผ ๋น„๊ตํ•  ๋•Œ, Type 1๊ณผ 3์€ 5MPa ๋‚ด์™ธ๋กœ, Type 2๋Š” 10MPa ๋‚ด์™ธ๋กœ ์œ ์‚ฌํ•œ ์••์ถ•๊ฐ•๋„๋ฅผ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค.

Table 7 Compressive strength of specimens

 

Type 1

Type 2

Type 3

CTS

35.1 MPa

66.9 MPa

82.2 MPa

UTM

38.9 MPa

77.2 MPa

86.4 MPa

4.4 ๊ณต๊ทน๋ฅ ๊ณผ ์••์ถ•๊ฐ•๋„์˜ ์ƒ๊ด€๊ด€๊ณ„

Fig. 11์—๋Š” ์ œ์•ˆ๋ชจ๋ธ์„ ํ™œ์šฉํ•˜์—ฌ ์–ป์–ด์ง„ ๊ณต๊ทน๋ฅ ๊ณผ UTM ๋ฐ CTS ํ•ด๋จธ๋กœ ๊ณ„์ธก๋œ ์••์ถ•๊ฐ•๋„๋ฅผ ๋น„๊ตํ•œ ๊ฒƒ์ด๋‹ค. ๋ชจ๋ธ๋กœ ์ธก์ •ํ•œ ๊ณต๊ทน๋ฅ ์€ ์••์ถ•๊ฐ•๋„์™€ ๋ชจ๋‘ ๋ฐ˜๋น„๋ก€ํ•˜๋Š” ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค.

Fig. 11 Relationship between porosity and strength
../../Resources/ksm/jksmi.2023.27.1.30/fig11.png

5. ๊ฒฐ ๋ก 

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ์ฝ˜ํฌ๋ฆฌํŠธ์˜ ํ‘œ๋ฉด ์ด๋ฏธ์ง€์™€ ์˜์ƒ ๋ถ„ํ• ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ๊ณต๊ทน๋ฅ ์„ ๊ณ„์ธกํ•˜๋Š” ๋”ฅ๋Ÿฌ๋‹ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•˜์˜€์œผ๋ฉฐ, ๋ณธ ์—ฐ๊ตฌ๋กœ๋ถ€ํ„ฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ๊ฒฐ๋ก ์„ ๋„์ถœํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค.

1. ์ฝ˜ํฌ๋ฆฌํŠธ ํ‘œ๋ฉด ์ด๋ฏธ์ง€์™€ ์˜์ƒ ๋ถ„ํ•  ์•Œ๊ณ ๋ฆฌ์ฆ˜(Fully Convolution Network, MobileNet, DeepLab v3+ ๋ฐ U-net)์„ ํ™œ์šฉํ•˜์—ฌ ๋ฒค์น˜๋งˆํ‚น ํ…Œ์ŠคํŠธ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค. ํ…Œ์ŠคํŠธ๊ฒฐ๊ณผ Deep Lab v3+์™€ U-net์ด ๊ฐ€์žฅ ๋†’์€ ์ •ํ™•๋„๋ฅผ ๋‚˜ํƒ€๋‚ด์—ˆ์œผ๋ฉฐ, U-net์€ DeepLab v3+์— ๋น„ํ•ด 30% ๋‚ฎ์€ ํ•™์Šต ์‹œ๊ฐ„์„ ์†Œ์š”ํ•˜์˜€๋‹ค. ๋”ฐ๋ผ์„œ U-net์„ ํ™œ์šฉํ•˜์—ฌ ๋งค๊ฐœ๋ณ€์ˆ˜ ์ตœ์ ํ™”๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€์œผ๋ฉฐ, 97% ์ˆ˜์ค€์˜ ๊ฒ€์ฆ์ •ํ™•๋„๋ฅผ ํ™•๋ณดํ•  ์ˆ˜ ์žˆ์—ˆ๋‹ค.

2. ์ฝ˜ํฌ๋ฆฌํŠธ์˜ w/c๊ฐ€ ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ U-net์„ ํ™œ์šฉํ•˜์—ฌ ์˜ˆ์ธกํ•œ ๊ณต๊ทน๋ฅ ์ด ์ฆ๊ฐ€ํ•˜๋Š” ๊ฒฝํ–ฅ์„ ๋‚˜ํƒ€๋‚ด์—ˆ๋‹ค. ๋˜ํ•œ, ์ผ๋ฐ˜๊ฐ•๋„ ์ฝ˜ํฌ๋ฆฌํŠธ์— ๋Œ€ํ•ด์„œ X-Ray Microscope (XRM)์˜ ์ธก์ • ๊ณต๊ทน๋ฅ ๊ณผ U-net์˜ ์˜ˆ์ธก ๊ณต๊ทน๋ฅ ์ด ์œ ์‚ฌํ•œ ๊ฒƒ์œผ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค. ์ด์— ๋น„ํ•ด, ๊ณ ๊ฐ•๋„ ์ฝ˜ํฌ๋ฆฌํŠธ์—์„œ๋Š” ์ธก์ • ๊ณต๊ทน๋ฅ (XRM)๋ณด๋‹ค ์˜ˆ์ธก ๊ณต๊ทน๋ฅ (U-net)์ด ์ž‘๊ฒŒ ๋‚˜ํƒ€๋‚ฌ๋Š”๋ฐ, ์ด๋Š” w/c๊ฐ€ ๊ฐ์†Œํ•จ์— ๋”ฐ๋ผ ๊ด‘ํ•™ํ˜„๋ฏธ๊ฒฝ์œผ๋กœ ๊ด€์ธก์ด ์–ด๋ ค์šด ์ž‘์€ ๊ณต๊ทน๋“ค์ด ๋งŽ์•„์ง€๊ธฐ ๋•Œ๋ฌธ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค.

3. ์ˆ˜๋ถ„ ํก์ˆ˜๋ฒ•์œผ๋กœ ๊ด€์ธกํ•œ ๊ณต๊ทน๋ฅ ์€ XRM ๋˜๋Š” U-net์œผ๋กœ ๊ณ„์ธก๋œ ๊ณต๊ทน๋ฅ ๋ณด๋‹ค ๋†’๊ฒŒ ๋‚˜ํƒ€๋‚ฌ๋Š”๋ฐ, ์ด๋Š” XRM ๋˜๋Š” U-net์œผ๋กœ ๊ด€์ธกํ•  ์ˆ˜ ์žˆ๋Š” ์ตœ์†Œ๊ณต๊ทน ํฌ๊ธฐ๋ณด๋‹ค ์ž‘์€ ๊ณต๊ทน์—์„œ๋„ ์ˆ˜๋ถ„์ด ํก์ˆ˜๋˜๊ธฐ ๋•Œ๋ฌธ์œผ๋กœ ํŒ๋‹จ๋œ๋‹ค.

4. U-net์œผ๋กœ ๊ณ„์ธกํ•œ ๊ณต๊ทน๋ฅ ๊ณผ ์••์ถ•๊ฐ•๋„์˜ ์ƒ๊ด€๊ด€๊ณ„๋ฅผ ํ™•์ธํ•˜์˜€์œผ๋ฉฐ, ๊ณต๊ทน๋ฅ ์ด ์ฆ๊ฐ€ํ• ์ˆ˜๋ก ์••์ถ•๊ฐ•๋„๋Š” ์ง€์ˆ˜์ ์œผ๋กœ ๊ฐ์†Œํ•˜์˜€๋‹ค.

5. ์ œ์•ˆ ๋ฐฉ๋ฒ•์€ ์ ๊ฒ€ ๋ฐ ์ง„๋‹จ์ด ์‹ค์‹œ๋˜๋Š” ํ˜„์žฅ์—์„œ ์ฝ˜ํฌ๋ฆฌํŠธ์˜ ๊ณต๊ทน๋ฅ ์„ ๊ฐ„ํŽธํ•˜๊ณ  ์‹ ์†ํ•˜๊ฒŒ ๊ณ„์ธกํ•˜๋Š”๋ฐ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ๋‹ค. ํ•œํŽธ, ์ด๋ฏธ์ง€ ์ดฌ์˜์— ํ™œ์šฉ๋œ ํ˜„๋ฏธ๊ฒฝ์˜ ๋ฐฐ์œจ์— ๋”ฐ๋ผ ๊ด€์ธกํ•  ์ˆ˜ ์žˆ๋Š” ๊ณต๊ทน ํฌ๊ธฐ์˜ ๋ฒ”์œ„๊ฐ€ ํ•œ์ •๋˜๋ฏ€๋กœ, ๋ฐฐ์œจ์ด ๋‚ฎ์€ ํ˜„๋ฏธ๊ฒฝ์„ ํ™œ์šฉํ•  ๊ฒฝ์šฐ ๊ณต๊ทน๋ฅ ์„ ์‹ค์ œ๋ณด๋‹ค ๋‚ฎ๊ฒŒ ํ‰๊ฐ€ํ•  ์ˆ˜ ์žˆ๋‹ค.

5.1 ์ถ”ํ›„ ์—ฐ๊ตฌ

์ถ”ํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ๋‹ค์–‘ํ•œ ๋ฐฐ์œจ์˜ ๊ณต๊ทน ์ด๋ฏธ์ง€(๋ฉ€ํ‹ฐ์Šค์ผ€์ผ ๊ณต๊ทน ์ด๋ฏธ์ง€)๋ฅผ ํ™œ์šฉํ•˜์—ฌ U-net ๊ธฐ๋ฐ˜ ์˜์ƒ ๋ถ„ํ• ๋ชจ๋ธ์„ ํ•™์Šตํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค. ๋˜ํ•œ, ์ œ์•ˆ ๋ฐฉ๋ฒ•์€ ๊ณต๊ทน๋ฅ ์˜ ์ถ”์ •๋ฟ๋งŒ ์•„๋‹ˆ๋ผ, ๊ณจ์žฌ๋Ÿ‰ ๋ฐ ์ฝ˜ํฌ๋ฆฌํŠธ ๋ฐฐํ•ฉ๋น„๋ฅผ ์ถ”์ •ํ•˜๋Š” ๋ฐ์—๋„ ํ™œ์šฉ๋  ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ, ๋ชจ๋ธ์˜ ์‘์šฉ๋ฒ”์œ„๋ฅผ ํ™•์žฅํ•˜์—ฌ ์ฝ˜ํฌ๋ฆฌํŠธ ํ’ˆ์งˆํ‰๊ฐ€์— ์ด๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋„๋ก ํ•ด์•ผ ํ•  ๊ฒƒ์ด๋‹ค.

๊ฐ์‚ฌ์˜ ๊ธ€

์ด ๋…ผ๋ฌธ์€ 2022๋…„๋„ ์ •๋ถ€(๊ณผํ•™๊ธฐ์ˆ ์ •๋ณดํ†ต์‹ ๋ถ€)์˜ ์žฌ์›์œผ๋กœ ํ•œ๊ตญ์—ฐ๊ตฌ์žฌ๋‹จ์˜ ์ง€์›์„ ๋ฐ›์•„ ์ˆ˜ํ–‰๋œ ์—ฐ๊ตฌ์ž„(No. 2019R1A2C 2086388).

References

1 
Ministry of Land, Infrastructure and Transport. (2022), Guidelines for safety and maintenance of infrastructures, Korea Ministry of Government Legislation. https://www.law.go.kr/URL
2 
Chang, C. F., and Chen, J. W. (2006). The experimental investigation of concrete carbonation depth, Cement and Concrete Research, 36(9), 1760-1767.DOI
3 
KS F 2385. (2018). Permeable asphalt mixtures, Korea Standards Association. https://e-ks.kr/streamdocs/view/sd;streamdocsId=7 2059199200454467URL
4 
Lu, S., Landis, E. N., and Keane, D. T. (2006). X-ray microtomographic studies of pore structure and permeability in Portland cement concrete, Materials and Structures, 39(6), 611-620.DOI
5 
Torres-Luque, M., Bastidas-Arteaga, E., Schoefs, F., Sรกnchez- Silva, M., and Osma, J. F. (2014). Non-destructive methods for measuring chloride ingress into concrete: State-of-the-art and future challenges, Construction and Building Materials, 68, 68-81.DOI
6 
Hofmarcher, M., Unterthiner, T., Arjona-Medina, J., Klambauer, G., Hochreiter, S., & Nessler, B. (2019). Visual scene understanding for autonomous driving using semantic segmentation, Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, 285-296.DOI
7 
Wang, G., Li, W., Zuluaga, M. A., Pratt, R., Patel, P. A., Aertsen, M., ... & Vercauteren, T. (2018). Interactive medical image segmentation using deep learning with image-specific fine tuning, IEEE Transactions on Medical Imaging, 37(7), 1562-1573.DOI
8 
Dogan, G., Arslan, M. H., and Ceylan, M. (2017), Concrete compressive strength detection using image processing based new test method, Measurement, 109, 137-148.DOI
9 
Jang, Y., Ahn, Y., and Kim, H. Y. (2019). Estimating compressive strength of concrete using deep convolutional neural networks with digital microscope images, Journal of Computing in Civil Engineering, 33(3), 04019018.DOI
10 
Yang, H., Jiao, S. J., and Yin, F. D. (2020). Multilabel Image Classification Based Fresh Concrete Mix Proportion Monitoring Using Improved Convolutional Neural Network, Sensors, 20(16), 4638.DOI
11 
Dung, C. V. (2019). Autonomous concrete crack detection using deep fully convolutional neural network, Automation in Construction, 99, 52-58.DOI
12 
Otsu, N. (1979). A threshold selection method from gray-level histograms, IEEE Transactions on Systems, Man, and Cybernetics, 9(1), 62-66.URL
13 
Dhanachandra, N., Manglem, K., and Chanu, Y. J. (2015). Image segmentation using K-means clustering algorithm and subtractive clustering algorithm, Procedia Computer Science, 54, 764-771.DOI
14 
Long, J., Shelhamer, E., and Darrell, T. (2015). Fully convolutional networks for semantic segmentation, Proceedings of the IEEE conference on computer vision and pattern recognition, 3431-3440.URL
15 
Ronneberger, O., Fischer, P., and Brox, T. (2015). U-net: Convolutional networks for biomedical image segmentation, International Conference on Medical image computing and computer-assisted intervention, Springer, Berlin, 234-241.DOI
16 
Chen, L. C., Papandreou, G., Kokkinos, I., Murphy, K., and Yuille, A. L. (2017). Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs, IEEE transactions on pattern analysis and machine intelligence, 40(4), 834-848.DOI
17 
Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., ... and Adam, H. (2017). Mobilenets: Efficient convolutional neural networks for mobile vision applications, arXiv preprint arXiv, 1704.04861.DOI
18 
Ahn, J., Lee, Y., Vaidya, S., Kim, J. H., & Lee, S. W. (2013). Estimation the porosity of pervious concretes based on X-Ray CT and submerged weight, Journal of the Korean Society of Hazard Mitigation, 13(4), 77-82.URL
19 
Kumar, R., and Bhattacharjee, B. (2003). Porosity, pore size distribution and in situ strength of concrete, Cement and Concrete Research, 33(1), 155-164.DOI
20 
Nitto. Concrete tester and surveyor(User manual). (2009). Nitto, Osaka, Japan.URL
21 
KS F 2405. (2022). Test method for compressive strength of concrete, Korea Standards Association. https://e-ks.kr/streamdocs/ vIew/sd;streamdocsId=7205926579 6199707URL
22 
Liashchynskyi, P., and Liashchynskyi, P. (2019). Grid search, random search, genetic algorithm: A big comparison for NAS, arXiv preprint arXiv, 1912.06059.DOI