• ๋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. (DAEYOUN ENGINEERING CO., LTD.)



Condition Based Maintenance, Machine Learning, Signal Processing, Dense KNN

1. ์„œ ๋ก 

2014๋…„ ์ดํ›„๋ถ€ํ„ฐ ๊ตญ๋‚ด ์ฒ ๋„์ฐจ๋Ÿ‰ ๋ฐ ์ฒ ๋„๊ฑด์„ค ์‚ฌ์—…์— ์‹œ์Šคํ…œ ์—”์ง€๋‹ˆ์–ด๋ง ์š”๊ฑด์ด ์ถ”๊ฐ€๋˜๊ณ  ์‹ ๋ขฐ์„ฑ๊ณผ ์•ˆ์ „์„ฑ ์š”๊ฑด์„ ์ถ”๊ฐ€ํ•˜์—ฌ ์ฒ ๋„์•ˆ์ „๋ฒ•์ด ๊ฐ•ํ™”๋˜์–ด ์‹œํ–‰๋˜๊ณ  ์žˆ๋‹ค.

ํŠนํžˆ ์•ˆ์ „์žฅ์น˜๋“ค์— ๋Œ€ํ•œ ์š”๊ตฌ์‚ฌํ•ญ์€ IEC61508, IEC62278, IEC62279, IEC62425์˜ ํ‘œ์ค€์„ ์ค€์šฉํ•˜๋„๋ก ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ RAMS (Reliability, Availability, Maintainability, Safety)์™€ SIL(Safety Integrity Level)์ธ์ฆ์„ ์š”๊ตฌํ•˜๊ณ  ์žˆ๋‹ค.ํŠนํžˆ ์•ˆ์ „์žฅ์น˜๋“ค์— ๋Œ€ํ•œ ์š”๊ตฌ์‚ฌํ•ญ์€ IEC61508, IEC62278, IEC62279, IEC62425์˜ ํ‘œ์ค€์„ ์ค€์šฉํ•˜๋„๋ก ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ RAMS (Reliability, Availability, Maintainability, Safety)์™€ SIL(Safety Integrity Level)์ธ์ฆ์„ ์š”๊ตฌํ•˜๊ณ  ์žˆ๋‹ค.ํŠนํžˆ ์•ˆ์ „์žฅ์น˜๋“ค์— ๋Œ€ํ•œ ์š”๊ตฌ์‚ฌํ•ญ์€ IEC61508, IEC62278, IEC62279, IEC62425์˜ ํ‘œ์ค€์„ ์ค€์šฉํ•˜๋„๋ก ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ RAMS (Reliability, Availability, Maintainability, Safety)์™€ SIL(Safety Integrity Level)์ธ์ฆ์„ ์š”๊ตฌํ•˜๊ณ  ์žˆ๋‹ค.ํŠนํžˆ ์•ˆ์ „์žฅ์น˜๋“ค์— ๋Œ€ํ•œ ์š”๊ตฌ์‚ฌํ•ญ์€ IEC61508, IEC62278, IEC62279, IEC62425์˜ ํ‘œ์ค€์„ ์ค€์šฉํ•˜๋„๋ก ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ RAMS (Reliability, Availability, Maintainability, Safety)์™€ SIL(Safety Integrity Level)์ธ์ฆ์„ ์š”๊ตฌํ•˜๊ณ  ์žˆ๋‹ค.

์ฒ ๋„์ฐจ๋Ÿ‰ ์ถœ์ž…๋ฌธ์€ ์ „๋™์ฐจ๋ฅผ ์ด์šฉํ•˜๋Š” ์Šน๊ฐ์ด ์ตœ์ดˆ๋กœ ์ ‘ํ•˜๋Š” ์žฅ์น˜์ด๊ณ  ์•ˆ์ „์—๋„ ๋ฐ€์ ‘ํ•œ ๊ด€๋ จ์„ฑ์ด ์žˆ๋‹ค.

์ถœ์ž…๋ฌธ์ด ๊ณ ์žฅ๋‚  ๊ฒฝ์šฐ ์Šน๊ฐ์˜ ์Šน. ํ•˜์ฐจ๊ฐ€ ๋ถˆ๊ฐ€๋Šฅํ•˜๊ณ  ๋˜ํ•œ ์ฐจ๋Ÿ‰ ๋‚ด๋ถ€์˜ ํ™”์žฌ๋‚˜ ํ„ฐ๋ ˆ ๋“ฑ ๊ธด๊ธ‰์ƒํ™ฉ ๋ฐœ์ƒ์‹œ ํƒˆ์ถœ์„ ๋ชปํ•˜๊ฒŒ ๋˜์–ด ์ƒ๋ช…์— ์œ„ํ˜‘์„ ๋Š๋‚„ ์ˆ˜ ์žˆ๋‹ค.

์ถœ์ž…๋ฌธ์€ ์ˆ˜๋Ÿ‰์ด ๋งŽ๊ณ (Car๋‹น 6~8๊ฐœ) ๋™์ž‘ ํšŸ์ˆ˜๊ฐ€ ๋นˆ๋ฒˆํ•˜๋ฉฐ, ์˜จ๋„๋ณ€ํ™”์— ๋’คํ‹€๋ฆผ์ด ๋ฐœ์ƒํ•˜๊ณ , ์Šน๊ฐ•์žฅ ์•ˆ์ „๋ฌธ๊ณผ ์—ฐ๋™ํ•˜๋Š” ๋“ฑ ๋‹ค์–‘ํ•œ Data ๊ธฐ๋ฐ˜์˜ ๊ด€๋ฆฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค.

๊ทธ๋ž˜์„œ ์ถœ์ž…๋ฌธ ์ž์ฒด์˜ ๊ณ ์žฅ์„ ๋ฏธ์—ฐ์— ๋ฐฉ์ง€ํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ๊ฐœ๋ณ„ ์ถœ์ž…๋ฌธ DCU(Door Control Unit)์™€ ์—ฐ๋™ํ•˜์—ฌ Train Door CBM (Condition Based Monitoring) Unit์„ ์„ค์น˜ํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜์—ฌ ๋ถ„์„ํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค.

๋„์‹œ์ฒ ๋„ ์šด์˜๊ธฐ๊ด€ ํ†ต๊ณ„ ๋ฐ์ดํ„ฐ์— ์˜ํ•˜๋ฉด 5๋ถ„ ์ด์ƒ ์ง€์—ฐ๋˜๋Š” ์ „๋™์ฐจ ๊ณ ์žฅ์˜ ์›์ธ์ค‘ ์ถœ์ž…๋ฌธ ๊ณ ์žฅ์ด 23%๋กœ ์กฐ์‚ฌ๋˜๊ณ  ์žˆ๋‹ค(10).

๋„์‹œ์ฒ ๋„์ฐจ๋Ÿ‰์— ์ ์šฉ๋˜๋Š” ์ถœ์ž…๋ฌธ์€ ํฌ์ผ“ ์Šฌ๋ผ์ด๋”ฉ ํ˜•์‹(Pocket Sliding Type), ์™ธ๋ถ€์Šฌ๋ผ์ด๋”ฉ ํ˜•์‹(Outside Sliding Type), ํ”Œ๋Ÿฌ๊ทธ ์Šฌ๋ผ์ด๋”ฉ ํ˜•์‹(Plug Sliding Type)๋“ฑ์ด ์žˆ๋‹ค(11).

์ค€ ๊ณ ์† ์—ด์ฐจ์ธ GTX ์ฐจ๋Ÿ‰์—๋Š” ์•ˆ์ „์„ฑ ๋ฐ ๊ธฐ๋ฐ€์„ฑ์ด ์š”๊ตฌ๋˜์–ด Plug Sliding Type ์ถœ์ž…๋ฌธ์„ ์ฑ„ํƒํ•˜๊ณ  ์žˆ์œผ๋ฉฐ ํ–ฅํ›„์—๋Š” ์„ ํ˜•๋ชจํ„ฐ(Liner Motor)๊ตฌ๋™๋ฐฉ์‹์˜ ์ถœ์ž…๋ฌธ์ด ๊ฐœ๋ฐœ๋˜์–ด ๊ธฐ๋ฐ€์„ ์š”ํ•˜๋Š” ์ฐจ๋Ÿ‰์— ํƒ‘์žฌ๋  ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋˜๊ณ  ์žˆ๋‹ค.

๊ตฌ๋™๋ฐฉ์‹์€ ๊ณต์••์‹๊ณผ ์ „๊ธฐ์‹์ด ์žˆ๋‹ค. ๊ณผ๊ฑฐ ์ˆ˜๋…„์ „๊นŒ์ง€๋งŒ ํ•˜์—ฌ๋„ ๊ณต์••์‹์ด ์ฃผ๋กœ ์‚ฌ์šฉ๋˜์–ด ์™”๊ณ  ๊ทผ๋ž˜ ๋“ค์–ด์™€์„œ ์ฐจ๋Ÿ‰ ์ œ์ž‘๋ฐœ์ฃผ ์ถœ์ž…๋ฌธ ์‚ฌ์–‘์€ ์†Œํ˜• ๋ชจํ„ฐ ๋ฐ ์ œ์–ด๊ธฐ์ˆ ์˜ ๋ฐœ๋‹ฌ๋กœ ์Šคํฌ๋ฅ˜ ํƒ€์ž…์˜ ์ „๊ธฐ์‹ ๊ตฌ๋™๋ฐฉ์‹์„ ์ฑ„ํƒํ•˜๊ณ  ์žˆ๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋„์‹œ์ฒ ๋„ ์ฐจ๋Ÿ‰์˜ Pocket Sliding Type ์ถœ์ž…๋ฌธ ์—”์ง„์„ ๊ตฌ๋™ํ•˜๋Š” DC๋ชจํ„ฐ์˜ ์—ด๋ฆผ, ๋‹ซํž˜์‹œ ์ •์ƒ ๋ชจํ„ฐ์™€ ๋น„์ •์ƒ ๋ชจํ„ฐ์˜ ์ „๋ฅ˜์ฃผํŒŒ์ˆ˜ ๊ฐ’์„ ๊ฐ๊ฐ 2300ํšŒ ์ด์ƒ ์ธก์ •ํ•˜์—ฌ ์‹œ๊ฐ„์˜์—ญ(Time Domain) ํ†ต๊ณ„ ์ธ์ž 13๊ฐ€์ง€๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์ถœ์ž…๋ฌธ ๋ชจํ„ฐ์˜ ํŠน์„ฑ์— ์ ํ•ฉํ•œ ํ†ต๊ณ„์ธ์ž๋ฅผ ์ถ”์ถœํ•˜๊ณ  ์ถ”์ถœํ•œ ํ†ต๊ณ„์ธ์ž์™€ ๋ชจํ„ฐ ์—ด๋ฆผ, ๋‹ซํž˜์‹œ ์ „๋ฅ˜์น˜์˜ ๋ถ„ํฌ๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์ •์ƒ๋ชจํ„ฐ์™€ ๋น„์ •์ƒ ๋ชจํ„ฐ์˜ ์ƒํƒœ๋ฅผ ๋ถ„๋ฅ˜ํ•˜์—ฌ Machine Learning ๊ธฐ๋ฒ•์œผ๋กœ ๊ณ ์žฅ๋ฐœ์ƒ์„ ์˜ˆ์ธกํ•จ์œผ๋กœ์จ ํ–ฅํ›„ ์ƒํƒœ๊ธฐ๋ฐ˜ ์œ ์ง€๋ณด์ˆ˜์— ํ™œ์šฉํ•˜๊ณ ์ž ํ•œ๋‹ค.

2. ๋ณธ ๋ก 

2.1 ์ถœ์ž…๋ฌธ ๋™์ž‘ ๊ฐœ์š” ๋ฐ ์‹œํ—˜ ๋ฐฉ๋ฒ•

2.1.1 ์ถœ์ž…๋ฌธ ๋™์ž‘ ๊ฐœ์š”

๋„์‹œ์ฒ ๋„ ์ „๋™์ฐจ์˜ ์ถœ์ž…๋ฌธ์€ 1๋Ÿ‰ ๊ธฐ์ค€ ํŽธ์ธก์— 4๊ฐœ๊ฐ€ ์žˆ๊ณ  ๊ฐ๊ฐ 1,300ใŽœ ์–‘๋ฌธํ˜• ํญ์œผ๋กœ ๋˜์–ด์žˆ๋‹ค.

์ถœ์ž…๋ฌธ์€ ์–‘์ชฝ์œผ๋กœ ์—ด๊ณ  ๋‹ซํžˆ๋Š” ๊ตฌ์กฐ๋กœ ๋˜์–ด ์žˆ์œผ๋ฉฐ ์—ด๋ฆด ๋•Œ๋Š” ์ถœ์ž…๋ฌธ ์ „์ฒด๊ฐ€ ์™„์ „ํžˆ ์—ด๋ฆฌ๊ฒŒ ํ•˜์—ฌ ์Šน๊ฐ์˜ ์ถœ์ž…์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•œ๋‹ค.

์ถœ์ž…๋ฌธ์€ ๊ธฐ๊ด€์‚ฌ์˜ ์—ด๋ฆผ๋ฒ„ํŠผ ์กฐ์ž‘์œผ๋กœ ๊ฐœ๋ณ„์ถœ์ž…๋ฌธ์˜ ์ œ์–ด์žฅ์น˜(์ดํ•˜ DCU๋กœ ํ‘œ๊ธฐ)๋ฅผ ํ†ตํ•ด ๋ชจํ„ฐ์— ์ „๋ฅ˜๊ฐ’์ด ์ธ๊ฐ€๋˜์–ด ์Šคํฌ๋ฅ˜๊ฐ€ ํšŒ์ „ํ•˜์—ฌ ์—ด๋ฆฌ๊ณ  ๋‹ซํžˆ๋ฉฐ, ํŠธ๋กค๋ฆฌ(ํ–‰๊ฑฐ)์— ์˜ํ•ด ๋งค๋‹ฌ๋ ค ์žˆ์œผ๋ฉฐ, ํŠธ๋กค๋ฆฌ์˜ ๋ณผํŠธ๋ฅผ ์ œ๊ฑฐํ•˜๋ฉด ์ถœ์ž…๋ฌธ์„ ํƒˆ๊ฑฐํ•  ์ˆ˜ ์žˆ๋‹ค.

Pocket Sliding ์ „๊ธฐ์‹ ์ถœ์ž…๋ฌธ์‹œ์Šคํ…œ์€ ๊ธฐ์กด์˜ ๊ณต์••์‹ ์ถœ์ž…๋ฌธ์‹œ์Šคํ…œ์— ๋น„ํ•˜์—ฌ ๋‚ด๊ตฌ์„ฑ ๋ฐ ์œ ์ง€๋ณด์ˆ˜์„ฑ์ด ์šฐ์ˆ˜ํ•œ ์‹œ์Šคํ…œ์œผ๋กœ์จ ์Šน๊ฐ์˜ ์•ˆ์ „์ด ๊ฐ€์žฅ ์šฐ์„ ์‹œ ๋˜๋„๋ก ์„ค๊ณ„ ๋˜์–ด์žˆ๋‹ค(12).

์šด์ „์‹ค์—์„œ๋Š” ๊ฐ DCU๋ฅผ ํ†ตํ•ด TCMS(Train Computer Moni- toring System)๋กœ ์ „์†ก๋œ ์ถœ์ž…๋ฌธ์˜ ์ƒํƒœ์ •๋ณด๋ฅผ ์šด์ „์‹ค์˜ ๋‹จ๋ง๊ธฐ๋ฅผ ํ†ตํ•˜์—ฌ ํŒŒ์•…ํ•  ์ˆ˜ ์žˆ๋‹ค.

๊ฐ ์ฐจ๋Ÿ‰์—๋Š” 8์„ธํŠธ์˜ ์Šน๊ฐ์šฉ ์ธก ์ถœ์ž…๋ฌธ์ด ์žˆ์œผ๋ฉฐ, 1Set์˜ ์ „๊ธฐ์‹ ์ถœ์ž…๋ฌธ ์‹œ์Šคํ…œ์€ ํฌ๊ฒŒ 4๋ถ€๋ถ„(์ œ์–ด ์žฅ์น˜๋ถ€, ์˜คํผ๋ ˆ์ดํ„ฐ๋ถ€, ํŒ๋„ฌ๋ถ€, ๋น„์ƒ์žฅ์น˜๋ถ€)์œผ๋กœ ๊ตฌ์„ฑ๋˜์–ด ์žˆ๋‹ค.

๊ทธ๋ฆฌ๊ณ  ๊ฐ ๋ถ€๋ถ„์€ ๋ชจ๋“ˆํ™” ๋˜์–ด ์žˆ์–ด ์ œํ’ˆ์˜ ์œ ์ง€ ๋ฐ ๋ณด์ˆ˜๊ฐ€ ์šฉ์ดํ•œ ์ƒํƒœ์ด๋‹ค. ์ถœ์ž…๋ฌธ์˜ ๊ตฌ์กฐ๋Š” Fig. 1์„ ์ฐธ์กฐํ•˜๋ฉด ๋œ๋‹ค.

Fig. 1. Door Structure

../../Resources/kiee/KIEE.2021.70.1.096/fig1.png

2.1.2 ์ถœ์ž…๋ฌธ ์‹œํ—˜ ๋ฐฉ๋ฒ•

์ถœ์ž…๋ฌธ์˜ ์ •์ƒ ๋™์ž‘์ƒํƒœ๋ฅผ ํ™•์ธํ•˜๊ธฐ ์œ„ํ•˜์—ฌ ๋™์ž‘ ์ƒํƒœ๋ฅผ ์•„๋ž˜์™€ ๊ฐ™์ด 6๋‹จ๊ณ„๋กœ ๋‚˜๋ˆ„์–ด ์‹ค์‹œํ•œ๋‹ค.

1) ์ถœ์ž…๋ฌธ ๊ฐœํ ํ™•์ธ

2) ์ถœ์ž…๋ฌธ ๋™์ž‘ ์‹œ๊ฐ„ ํ™•์ธ

3) ์ถœ์ž…๋ฌธ ์ €ํ•ญ๋ ฅ ํ™•์ธ

4) ์ถœ์ž…๋ฌธ ์ž‘๋™ ์••๋ ฅ ํ™•์ธ

5) ์ถœ์ž…๋ฌธ ์ธํ„ฐ๋ก ๊ฐ„๊ฒฉ ํ™•์ธ

6) ์ถœ์ž…๋ฌธ์˜ ํ†ต์‹  ์ƒํƒœ ํ™•์ธ ๋ฐ ์ฃผํŒŒ์ˆ˜ ๋ถ„์„

2.2 ๊ณ ์žฅ์˜ˆ์ง€์•Œ๊ณ ๋ฆฌ์ฆ˜ ์†Œ๊ฐœ

2.2.1 ์ผ๋ฐ˜์ ์ธ ๊ฑด์ „์„ฑ ํ‰๊ฐ€ ๋ฐ ๊ณ ์žฅ ์˜ˆ์ง€ ๋ฐฉ๋ฒ•๋ก 

ํ˜„์žฌ ์šด์˜๊ธฐ๊ด€์—์„œ๋Š” ๋Œ€๋ถ€๋ถ„ ์ฒ ๋„์ฐจ๋Ÿ‰์˜ ์œ ์ง€๋ณด์ˆ˜ ์ฃผ๊ธฐ๋Š” ๊ฑฐ๋ฆฌ ๊ธฐ๋ฐ˜, ์‹œ๊ฐ„ ๊ธฐ๋ฐ˜(TBM)์œ ์ง€๋ณด์ˆ˜๊ฐ€ ํ˜„์žฌ๋Š” ๊ฐ€์žฅ ๋ณดํŽธํ™”๋œ ์œ ์ง€๋ณด์ˆ˜ ๋ฐฉ์•ˆ์œผ๋กœ ์ ์šฉ๋˜๊ณ  ์žˆ๋‹ค(1).

์‹ ๋ขฐ์„ฑ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ๊ณ ์žฅ์„ ์˜ˆ๋ฐฉํ•˜๋Š” RCM(Reliability Centered Maintenance)๊ณผ RBM (Risk-Based Maintenance)์—ฐ๊ตฌ๊ฐ€ ์ง„ํ–‰๋˜์—ˆ์ง€๋งŒ ์šด์˜ ์ค‘์— ๋ฐœ์ƒํ•˜๋Š” ์ „๋ฐ˜์ ์ธ ์ƒํ™ฉ์„ ๋Œ€์ฒ˜ํ•˜๊ธฐ์—๋Š” ์–ด๋ ค์›€์ด ์žˆ์–ด ํ˜„์žฅ ์ ์šฉ์—๋Š” ํ•œ๊ณ„๊ฐ€ ์žˆ๋‹ค(2).

์ตœ๊ทผ ICT(Information and Communication Technology) ๋ฐœ์ „๊ณผ Industry 4.0์˜ ๊ฐœ๋…์„ ๋ฐ”ํƒ•์œผ๋กœ ์‹œ์„ค๊ณผ ์žฅ๋น„์— ๋‹ค์–‘ํ•œ ์„ผ์„œ๋“ค์„ ๋ถ€์ฐฉํ•˜์—ฌ ์‹ค์‹œ๊ฐ„์œผ๋กœ ๋ฐ์ดํ„ฐ๋ฅผ ์ˆ˜์ง‘ํ•˜๊ณ  ๋ถ„์„๊ธฐ๋ฒ•์„ ํ†ตํ•˜์—ฌ ๊ณ ์žฅ์„ ์ง„๋‹จํ•˜๊ณ  ๊ณ ์žฅ ์‹œ๊ธฐ์™€ ์ž”์กด์ˆ˜๋ช…์„ ์˜ˆ์ธกํ•˜๋Š” PHM(Prognostics and Health Management) ๊ธฐ๋ฒ•์ด ์—ฐ๊ตฌ๋˜๊ณ  ์žˆ๋‹ค(6).

PHM์€ ์šด์˜ ์ค‘์ธ ์ฒ ๋„์ฐจ๋Ÿ‰, ์‹œ์„ค๊ณผ ์žฅ๋น„์˜ ๊ฒฐํ•จ์ด๋‚˜ ์„ฑ๋Šฅ ์ €ํ•˜๋ฅผ ์ง€์†์ ์œผ๋กœ ๋ชจ๋‹ˆํ„ฐ๋ง ํ•˜์—ฌ ์ด์ƒ ์ง•ํ›„๋ฅผ ์ง„๋‹จํ•˜๊ณ  ๊ณ ์žฅ ์ˆ˜์ค€์ด๋‚˜ ์‹œ๊ธฐ๋ฅผ ์˜ˆ์ง€ํ•˜์—ฌ ์˜ˆ์ธก ์ •๋น„๋ฅผ ํ†ตํ•ด ๊ณ ์žฅ์œผ๋กœ ์ธํ•ด ๋ฐœ์ƒํ•˜๋Š” ์†์‹ค์„ ๋ฐฉ์ง€ํ•˜๊ณ  ์œ ์ง€ ๋ณด์ˆ˜์˜ ์•ˆ์ „์„ฑ๊ณผ ๊ฐ€์šฉ์„ฑ์„ ํ™•๋ณดํ•˜์—ฌ ์šด์šฉ ํšจ์œจ์„ฑ์„ ํ–ฅ์ƒํ•˜๋Š” ๊ธฐ์ˆ ์ด๋‹ค(9).

๊ทธ๋ฆฌ๊ณ  ์˜ˆ์ธก ์ •๋น„์˜ ์‹œ๊ธฐ๋ฅผ ์œ„ํ•ด PHM์€ ๋ฐ์ดํ„ฐ๋ฅผ ํ†ต๊ณ„์ ์œผ๋กœ ๋ถ„์„ํ•˜๊ณ  ๋‹ค์–‘ํ•œ ์ธ๊ณต์ง€๋Šฅ ๊ธฐ๋ฒ•์„ ๊ธฐ๋ฐ˜์œผ๋กœ ์ ‘๊ทผํ•˜์—ฌ ๊ณ ์žฅ์ด ๋ฐœ์ƒํ•˜๊ธฐ ์ „์˜ ์œ ํšจ ์‹œ๊ฐ„์ธ RUL(Remaining Useful Life)์„ ์ธก์ • ํ•œ๋‹ค.

PHM์€ ์˜ˆ๋ฐฉ์ •๋น„ ๋ณด๋‹ค ๋ณด์ „์„ฑ๊ณผ ๊ฒฝ์ œ์„ฑ, ์‹ ๋ขฐ์„ฑ, ๊ฐ€์šฉ์„ฑ์„ ํ–ฅ์ƒํ•  ์ˆ˜ ์žˆ์–ด ๊ตญ๋‚ด์—์„œ๋„ ์—ฐ๊ตฌ๋ฅผ ์ง„ํ–‰ ํ•˜๊ณ  ์žˆ์ง€๋งŒ ์•„์ง์€ ๋ฏธํกํ•œ ์‹ค์ •์ด๋‹ค.

๊ณ ์žฅ์ง„๋‹จ์„ ์œ„ํ•ด Machine Learning๊ณผ Deep Learning์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ฐ์ดํ„ฐ๋ฅผ ๋ถ„์„ํ•˜๊ณ  ์ƒํƒœ๋‚˜ ๊ณ ์žฅ์„ ์ง„๋‹จํ•˜๋Š” ๋ชจ๋ธ๊ณผ ๊ณ ์žฅ์œจ์„ ์ถ”์ •ํ•˜์—ฌ ๊ณ ์žฅ์˜ ์‹œ๊ธฐ๋ฅผ ์˜ˆ์ธกํ•˜๋Š” ์˜ˆ์ง€ ๋ชจ๋ธ์„ ๊ฐœ๋ฐœํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค(6).

์ตœ๊ทผ ์ฒ ๋„์ฐจ๋Ÿ‰ ์ œ์ž‘ ์š”๊ตฌ์‚ฌํ•ญ์—๋Š” ์ฃผ์š” ์•ˆ์ „์žฅ์น˜(์ถœ์ž…๋ฌธ, ์ œ๋™, ์‹ ํ˜ธ ๋“ฑ)๋“ค์˜ ์ƒํƒœ๋ฅผ ์ง„๋‹จํ•˜๊ณ  ๊ด€๋ฆฌํ•˜์—ฌ ํšจ์œจ์ ์ธ ์œ ์ง€๋ณด์ˆ˜๋ฅผ ํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ์•ˆ์„ ์ œ์‹œํ•˜๋„๋ก ํ•˜๊ณ  ์žˆ๋‹ค.

์ด๋Ÿฌํ•œ ์‹œ๋Œ€์ ์ธ ๋ฐฐ๊ฒฝ ๋ฐ ๊ธฐ์ˆ  Trend์— ๊ธฐ๋ฐ˜ํ•˜์—ฌ ์ตœ๊ทผ ์ƒํƒœ๊ธฐ๋ฐ˜ ์œ ์ง€๋ณด์ˆ˜(CBM) ๋ฐ ๊ฑด์ „์„ฑ ๊ด€๋ฆฌ (PHM)๋ฐฉ์•ˆ ๋“ฑ์˜ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š”ํ•˜๋‹ค(3).

Machine Learning์€ ์ง€๋„ํ•™์Šต(supervised learning), ์ž์œจํ•™์Šต(unsupervised learning), ๊ฐ•ํ™”ํ•™์Šต (reinforcement learning) ์„ธ ๊ฐ€์ง€๋กœ ๋ถ„๋ฅ˜๋˜๋Š”๋ฐ ๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” ๋จธ์‹  ๋Ÿฌ๋‹ ๊ธฐ๋ฒ• ์ค‘ โ€˜์ง€๋„ํ•™์Šตโ€™๊ณผ ์ž์œจํ•™์Šตโ€™๋ฐฉ๋ฒ•์œผ๋กœ ์—ฐ๊ตฌ๋ฅผ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค(7,8).

์—ด์ฐจ์˜ ์‹œํ—˜ ์šด์ „ ๋‹จ๊ณ„ ๋“ฑ์˜ ๊ณผ๊ฑฐ ๋ฐ์ดํ„ฐ๊ฐ€ ์กด์žฌํ•˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ์—์„œ๋Š” ํ‘œ์ค€ ๋ชจ๋ธ์˜ ๊ฐ๊ฐ์˜ ์ด๋ฒคํŠธ๊ฐ€ ์–ด๋–ค ๊ฐ’์— ์ˆ˜๋ ดํ• ์ง€ ์˜ˆ์ƒํ•  ์ˆ˜ ์—†๊ธฐ ๋•Œ๋ฌธ์— ์ž์œจ ํ•™์Šต ๊ธฐ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ ์ดˆ๊ธฐ ํ‘œ์ค€ ๋ชจ๋ธ์„ ์ƒ์„ฑํ•˜์—ฌ์•ผ ํšจ์œจ์ ์ด๋‹ค.

์ถฉ๋ถ„ํ•œ ๋ฐ์ดํ„ฐ๊ฐ€ ์ˆ˜์ง‘๋˜์–ด ์‹ ๋ขฐํ•  ์ˆ˜์ค€์˜ ํ‘œ์ค€ ๋ชจ๋ธ์ด ์ด๋ฏธ ์ƒ์„ฑ๋˜์–ด ์žˆ๋Š” ๊ฒฝ์šฐ์—๋Š” ์‹ค์‹œ๊ฐ„ ์ˆ˜์ง‘๋œ ์—ด์ฐจ ์‹œ์Šคํ…œ์˜ ์ด์ƒ ์ง•ํ›„๋ฅผ ํŒ๋‹จ ํ•  ์ˆ˜ ์žˆ๋‹ค. ๋”ฐ๋ผ์„œ, ํ‘œ์ค€ ๋ชจ๋ธ์— ๋น„ํ•ด ์˜ค์ฐจ๋ฒ”์œ„ ์ด์ƒ์„ ๋ฒ—์–ด๋‚œ ์ด๋ฒคํŠธ ๊ธฐ๋ก์€ ์ด์ƒ ์ง•ํ›„๋กœ ํŒ๋‹จํ•˜๊ณ  ํ•ด๋‹น ๋ฐ์ดํ„ฐ๋ฅผ ์„ ๋ณ„์ ์œผ๋กœ ํ†ต๊ณ„์ž๋ฃŒ์—์„œ ์ œ์™ธํ•˜๋Š” ์ง€๋„ ํ•™์Šต ๋ฐฉ๋ฒ•์„ ์ฐจ์šฉํ•ด์•ผ๋งŒ ๋”์šฑ ์‹ ๋ขฐ์„ฑ ์žˆ๋Š” ํ‘œ์ค€ ๋ชจ๋ธ์„ ์ƒ์„ฑํ•ด ๋‚ผ ์ˆ˜ ์žˆ๋‹ค.

2.2.2 ์ „๋™์ฐจ ์ถœ์ž…๋ฌธ ๋ชจํ„ฐ ์ „๋ฅ˜ ๊ฐ’ ์ธก์ •

์ผ๋ฐ˜์ ์ธ ๊ณ ์žฅ ์ง„๋‹จ ์ ˆ์ฐจ๋Š” ์ˆ˜์ง‘๋œ ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ(Prepro- cessing)-์‹ ํ˜ธ์ฒ˜๋ฆฌ(Signal Processing)-ํŠน์ง• ์ถ”์ถœ(Feature Extrac- tion)-ํŠน์ง• ์„ ํƒ(Feature Selection)-๋ถ„๋ฅ˜ (Classification) ๋“ฑ์˜ ์ ˆ์ฐจ๋ฅผ ๋”ฐ๋ฅธ๋‹ค(4,6).

์‹ ํ˜ธ์ฒ˜๋ฆฌ์— ์•ž์„œ ์ธก์ •๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฒ€์ฆํ•˜๋Š” ์ „์ฒ˜๋ฆฌ(Prepro- cessing) ๊ณผ์ •์ด ์ค‘์š”ํ•˜๋‹ค(5).

์˜ˆ์ง€๋ชจ๋ธ ์—ฐ๊ตฌ์˜ ์ฒซ ๋ฒˆ์งธ ๋‹จ๊ณ„์ธ ๋ฐ์ดํ„ฐ ์ˆ˜์ง‘์„ ์œ„ํ•ด ์•„๋ž˜ Fig. 2์™€ ๊ฐ™์ด ์ถœ์ž…๋ฌธ Zig๋ฅผ ๊ตฌ๋น„ํ•œ ๊ตญ๋‚ด ์ถœ์ž…๋ฌธ ์ „๋ฌธ ์ œ์ž‘ ํšŒ์‚ฌ์˜ ๋„์›€์„ ๋ฐ›์•„ ์ง„ํ–‰ํ•˜์˜€๋‹ค.

Fig. 2. Door zig & DCU Monitoring Program (DC Motor current data acquisition- HEUNG IL)

../../Resources/kiee/KIEE.2021.70.1.096/fig2.png

Fig. 3์€ DC 100V, 150W, 2.1A์˜ ์‹ค๋ฌผ๊ณผ ์ „๋ฅ˜ ์ฃผํŒŒ์ˆ˜ ํŒŒํ˜• ๊ทธ๋ž˜ํ”„๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.

Fig. 3. DC Motor & Current Frequency

../../Resources/kiee/KIEE.2021.70.1.096/fig3.png

์ถœ์ž…๋ฌธ ์—ด๋ฆผ. ๋‹ซํž˜ ์‹œ ์‹œ๊ฐ„ ํ•ด์ƒ๋„(Time Resolution)์ •๋ณด๋Š” ์•„๋ž˜์™€ ๊ฐ™๋‹ค.

โˆ™์‹œ๊ฐ„ ์ฆ๊ฐ€๋Š” 1msec

โˆ™Sampling ์ฃผ๊ธฐ๋Š” 10msec

โˆ™์ถœ์ž…๋ฌธ ๊ด€๋ จ ์ œ์–ด ํ•จ์ˆ˜ ๋™์ž‘ ์ฃผ๊ธฐ๋Š” 8msec

โˆ™์ถœ์ž…๋ฌธ Zig์˜ ์—ด๋ฆผ, ๋‹ซํž˜ ๋™์ž‘ ์‹œ๊ฐ„์€ 13.8sec

์—ด๋ฆผ. ๋‹ซํž˜ ์‹œ ์ „๋ฅ˜ ํ”„๋กœํŒŒ์ผ(Current Profile) ๊ทธ๋ž˜ํ”„์˜ ํŒŒํ˜•์˜ ํ•ด์„์€ ๋‹ค์Œ๊ณผ ๊ฐ™๋‹ค

โˆ™์—ด๋ฆผ/๋‹ซํž˜ ์‹œ๋งˆ๋‹ค DCU์—์„œ ์‹œ๊ฐ„ ๊ฒฝ๊ณผ์— ๋”ฐ๋ฅธ ๊ฑฐ๋ฆฌ๋ฅผ ์ถ”์‚ฐํ•˜์—ฌ ๊ฐ€์† ๊ตฌ๊ฐ„๊ณผ ๊ฐ์† ๊ตฌ๊ฐ„์„ ๊ตฌ๋ณ„ ํ›„, ํ•ด๋‹น ๊ฑฐ๋ฆฌ ๊ตฌ๊ฐ„์— ๋ชจํ„ฐ ์ธก์œผ๋กœ ์ ์ ˆํ•œ ์ถœ๋ ฅ ์ „๋ฅ˜๋ฅผ ๋ณ€์กฐ (PWM ์ œ์–ด)ํ•˜์—ฌ ์†๋„๋ฅผ ์ œ์–ดํ•˜๊ณ , DCU์—์„œ ์—ด๋ฆผ/๋‹ซํž˜ ์‹œ ๋งˆ๋‹ค ํ‰๊ท  ์ „๋ฅ˜ ๊ฐ’์„ ๊ณ„์‚ฐํ•˜๊ณ  ์ €์žฅํ•  ์ˆ˜ ์žˆ๋‹ค.

โˆ™์—ด๋ฆผ์‹œ๊ฐ„ ๊ธฐ์ค€ 3,000msec, ๋‹ซํž˜์‹œ๊ฐ„ ๊ธฐ์ค€ 2,500 msec ๊ธฐ์ค€์ด๊ณ  ํŒŒํ˜• ๊ทธ๋ฆผ์€ ๋ชจํ„ฐ์— ํ๋ฅด๋Š” ์ „๋ฅ˜๊ฐ’์ด๋‹ค.

์ถœ์ž…๋ฌธ ๊ฑฐ๋ฆฌ ๊ณ„์‚ฐ์˜ ๊ทผ๋ณธ์€ ๋ชจํ„ฐ์— ๊ฐ€ํ•ด์ง€๋Š” ํž˜๊ณผ ์‹œ๊ฐ„์ด๋‹ค.

โˆ™์—ด๋ฆผ ํ–‰์ •์„ ๊ตฌ๊ฐ„๋ณ„๋กœ ๋‚˜๋ˆ„๋ฉด Unlock, ๊ฐ€์†, ๊ฐ์†, ์ •์ง€.

โˆ™๋‹ซํž˜ ํ–‰์ •์„ ๊ตฌ๊ฐ„๋ณ„๋กœ ๋‚˜๋ˆ„๋ฉด ๊ฐ€์†, ๊ฐ์†, ์ •์ง€.

โˆ™๊ฐ ํ–‰์ •๋ณ„ ์ „์••์ด ๋‹ค๋ฅด๋ฉฐ ์ถœ์ž…๋ฌธ์˜ ๊ตฌ์†๋ ฅ์ด๋‚˜ ๊ฐ€์†๋ ฅ์— ๋”ฐ๋ผ ์ „๋ฅ˜์˜ ํŒŒํ˜•์ด ๋‹ค๋ฅด๊ฒŒ ๋‚˜ํƒ€๋‚œ๋‹ค.

์•„๋ž˜ Fig. 4์™€ 5๋Š” ์ถœ์ž…๋ฌธ์˜ ์—ด๋ฆผ. ๋‹ซํž˜์‹œ ์ฃผํŒŒ์ˆ˜ ํŒŒํ˜•์ด๋‹ค

Fig. 4. Current frequency when door motor is open

../../Resources/kiee/KIEE.2021.70.1.096/fig4.png

Fig. 5. Current frequency when door motor is close

../../Resources/kiee/KIEE.2021.70.1.096/fig5.png

2.3 ์ถœ์ž…๋ฌธ ์ „๋ฅ˜๊ฐ’ ์‹œ๊ฐ„์˜์—ญ ํ†ต๊ณ„์ธ์ž ์œ ์˜์„ฑ ๊ฒ€์ฆ

์ „๋™์ฐจ ์ „๊ธฐ์‹ ์ถœ์ž…๋ฌธ์˜ ์ƒํƒœ์ •๋ณด ๋ฐ ๋ชจํ„ฐ์˜ ๊ตฌ๋™ ์ „๋ฅ˜๊ฐ’์„ ์•„๋ž˜ Table 1 ๊ธฐ๋ฐ˜์œผ๋กœ ์‹œ๊ฐ„์˜์—ญ(Time Domain) ํ†ต๊ณ„์ธ์ž 13๊ฐ€์ง€๋ฅผ ๋ถ„์„ํ•˜์—ฌ ํ†ต๊ณ„์  ์œ ์˜์„ฑ์„ ํ™•์ธํ•˜๊ณ  ๊ฒ€์ฆํ•˜๋Š” ์ ˆ์ฐจ ๋ฐ ๋ฐฉ์•ˆ์„ ์ œ์‹œํ•œ๋‹ค(6,12).

Table 1. Extracted time domain Features

../../Resources/kiee/KIEE.2021.70.1.096/table1.png

์ „๋™์ฐจ ์ถœ์ž…๋ฌธ์˜ ํŠน์ง•(Feature)์€ ์ •ํ•ด์ง„ ์‹œ๊ฐ„๋™์•ˆ ์ถœ์ž…๋ฌธ ๋ชจํ„ฐ์˜ ๊ตฌ๋™์œผ๋กœ ์Šคํฌ๋ฅ˜๋ฅผ ํšŒ์ „ํ•˜์—ฌ ์ถœ์ž…๋ฌธ์„ ์—ด๊ณ  ๋‹ซ๋Š” ๋ฐ˜๋ณต ๋™์ž‘(๊ฐ€์†-์ •์†-๊ฐ์†)์ด๋ฏ€๋กœ ์ „๋ฅ˜ ์ฃผํŒŒ์ˆ˜ ํŒŒํ˜•์„ ํ™•์ธํ•˜๋ฉด ์•Œ ์ˆ˜ ์žˆ๋‹ค.

์ธก์ •ํ•œ 2300ํšŒ์˜ ๋ฐ˜๋ณต๋™์ž‘์—์„œ ๋ชจํ„ฐ ์ „๋ฅ˜์˜ ์ฃผํŒŒ์ˆ˜๊ฐ€ ์–ด๋Š์ •๋„ ์ง‘์ค‘์ ์œผ๋กœ ์ค‘์‹ฌ์— ๋ถ„ํฌ๋˜์–ด ์žˆ๋Š”์ง€๋ฅผ ํ™•์ธํ•˜์—ฌ ์ •์ƒ, ๋น„์ •์ƒ ๋ชจํ„ฐ์˜ ์ „๋ฅ˜ ์ฃผํŒŒ์ˆ˜๊ฐ€ ๋ถ„ํฌ๋˜์–ด ์žˆ๋Š” ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ณ  ๊ตฐ์ง‘(Clustering)๋œ ๋ถ„๋ฅ˜๋ฅผ ์ฐพ์•„๋‚ด๊ธฐ ์œ„ํ•ด์„œ๋Š” ์•„๋ž˜ Fig. 6๊ณผ ๊ฐ™์ด ์ฒจ๋„(Kurtosis)๊ฐ€ ์ ํ•ฉํ•œ Features๋ผ๊ณ  ํŒ๋‹จํ•˜์˜€๋‹ค. ์ฒจ๋„(Kurtosis)๋Š” ํ™•๋ฅ  ๋ถ„ํฌ์˜ ๋พฐ์กฑํ•œ ์ •๋„๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ฒ™๋„๋กœ์„œ ์ž๋ฃŒ์˜ ๋ถ„ํฌ๊ฐ€ ์ค‘์‹ฌ ๊ฒฝํ–ฅ๊ฐ’์„ ์ค‘์‹ฌ์œผ๋กœ ์ง‘์ค‘์ ์œผ๋กœ ๋ถ„ํฌ๋˜์–ด ์žˆ๋Š” ์ •๋„ ์ฆ‰, ๊ด€์ธก์น˜๋“ค์ด ์–ด๋Š ์ •๋„ ์ง‘์ค‘์ ์œผ๋กœ ์ค‘์‹ฌ์— ๋ชฐ๋ ค์žˆ๋Š”๊ฐ€๋ฅผ ์ธก์ •ํ•  ๋•Œ ์‚ฌ์šฉ๋˜๋Š” ํ†ต๊ณ„์ธ์ž๋กœ ์ถœ์ž…๋ฌธ ๋ชจํ„ฐ ์ „๋ฅ˜์˜ ์ตœ๋Œ€ Torque์น˜๋ฅผ ํŒŒ์•…ํ•˜์—ฌ ๊ณผ๋ถ€ํ•˜(Overload) ์—ฌ๋ถ€๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ๋Š” ํŠน์ง•(Features)์ธ์ž์ด๋‹ค. ์ฐธ๊ณ ๋กœ ์ฒจ๋„๊ฐ€ 0์ด๋ฉด ์ •๊ทœ ๋ถ„ํฌ์ด๊ณ  ์ฒจ๋„๊ฐ€ 0๋ณด๋‹ค ํฌ๋ฉด ์ •๊ทœ๋ถ„ํฌ๋ณด๋‹ค ๊ธด ๊ผฌ๋ฆฌ๋ฅผ ๊ฐ€์ง€๋ฉฐ ๋ถ„ํฌ๊ฐ€ ์ค‘์•™๋ถ€๋ถ„์— ๋œ ์ง‘์ค‘๋˜์–ด ๋พฐ์กฑํ•œ ๋ชจ์–‘์„ ๊ฐ€์ง€๊ฒŒ ๋œ๋‹ค.

๊ฒ€์ฆ์€ ์‹ค์ œ ์ธก์ •ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ MATHLABMathWorks ์‚ฌ์—์„œ ๊ฐœ๋ฐœํ•œ ์ˆ˜์น˜ ํ•ด์„ ๋ฐ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ํ™˜๊ฒฝ์„ ์ œ๊ณตํ•˜๋Š” ๊ณตํ•™์šฉ ์†Œํ”„ํŠธ์›จ์–ด ์˜ Diagnostic Features Designer ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ์‹œ๊ฐ„ ์˜์—ญ ๋ฐ ์ŠคํŽ™ํŠธ๋Ÿผ ๊ธฐ๋Šฅ์„ ์ถ”์ถœํ•˜๊ณ  One way ANOVA ๊ธฐ๋ฒ•์œผ๋กœ ๋ถ„์„ํ•ด ์ค‘์š”๋„๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์ •๋ ฌํ•ด๋ณธ ๊ฒฐ๊ณผ ์•„๋ž˜ Fig. 7๊ณผ ๊ฐ™์ด ์ฒจ๋„(Kurtosis)๊ฐ€ ์ ํ•ฉํ•œ ํŠน์ง•(Features)์œผ๋กœ ํ™•์ธ๋˜์—ˆ๋‹ค.

์ฐธ๊ณ ๋กœ One way ANOVA๊ธฐ๋ฒ•์€ ์ธก์ •๊ฐ’์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋Š” ์š”์ธ์ด 1๊ฐœ(์˜ˆ:๋ชจํ„ฐ์ „๋ฅ˜์น˜)์ธ ๊ฒฝ์šฐ์— ์ ํ•ฉํ•œ ๊ธฐ๋ฒ•์ด๊ณ  ๋˜ํ•œ ์ „๋ฅ˜์น˜ ๋ถ„๋ฅ˜๊ฐ€ 3๊ฐœ ์ด์ƒ(์—ด๋ฆผ ์ •์ƒ, ์—ด๋ฆผ ๋น„์ •์ƒ, ๋‹ซํž˜ ์ •์ƒ, ๋‹ซํž˜ ๋น„์ •์ƒ)์ธ ๊ฒฝ์šฐ์—๋Š” ๋ถ„์‚ฐ๋ถ„์„(ANOVA)์„ ์‚ฌ์šฉํ•˜๋ฉด ํšจ์œจ์ ์ด๋‹ค.

๋ชจ๋“  ์žฅ์น˜๋“ค์€ ์žฅ์น˜๋งˆ๋‹ค ํŠน์ง•์ด ์žˆ๊ณ  ํŠน์ง•์„ ์ถ”์ถœํ•˜๊ธฐ ์œ„ํ•œ ํŠน์„ฑ ํŒŒ์•…๊ณผ ํŠน์ง•์ธ์ž์˜ ๋ฐ์ดํ„ฐ ์ทจ๋“ ๋ฐ ์ „์ฒ˜๋ฆฌ(Preprocessing) ํ›„ Machine Learning ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šตํ•˜๊ณ  ๋ถ„์„ํ•˜๋Š” ๊ฒƒ์ด ์ƒํƒœ๊ธฐ๋ฐ˜ ์œ ์ง€๋ณด์ˆ˜ (Condition Based Maintenance) ํ™œ๋™์˜ ์ฃผ๋œ ๋‚ด์šฉ์ด๋‹ค.

Fig. 6. Time Domain Analysis when door mortor open or close

../../Resources/kiee/KIEE.2021.70.1.096/fig6.png

Fig. 7. Selection of suitable statistical factors using One way ANOVA technique

../../Resources/kiee/KIEE.2021.70.1.096/fig7.png

2.4 ์ถœ์ž…๋ฌธ ๋ชจํ„ฐ์˜ ํŠน์ง• ์„ ํƒ

์ถœ์ž…๋ฌธ ๋ชจํ„ฐ์˜ ์ „๋ฅ˜์น˜๋ฅผ ์ธก์ •ํ•œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ฐ€์ง€๊ณ  ์ง€๋„ํ•™์Šต ๋ฐฉ๋ฒ•์„ ํ™œ์šฉํ•˜์—ฌ Machine Learning ๊ธฐ๋ฒ•์œผ๋กœ ๋ถ„์„ํ•ด๋ณธ ๊ฒฐ๊ณผ ์กฐ๋ฐ€ KNN(Dense K-Nearest Neighbor) ์•Œ๊ณ ๋ฆฌ์ฆ˜์ด ์ •ํ™•๋„ 99.7%๋กœ ์ ํ•ฉํ•œ ๊ฒƒ์œผ๋กœ ๋ถ„์„๋˜์—ˆ๋‹ค.

์กฐ๋ฐ€ KNN์€ ์ธก์ •๋ฐ์ดํ„ฐ๊ฐ€ ๋งŽ๊ณ  ๋ฐ์ดํ„ฐ ์ธก์ •์ฃผ๊ธฐ๊ฐ€ ์กฐ๋ฐ€ํ•œ ๊ฒฝ์šฐ์— ์‚ฌ์šฉํ•˜๋Š” ๋Œ€ํ‘œ์ ์ธ Machine Learning ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ค‘ ํ•˜๋‚˜๋กœ Training Data(ํ•™์Šต๋ฐ์ดํ„ฐ)์™€ Test Data(์ธก์ •๋ฐ์ดํ„ฐ)์‚ฌ์˜์˜ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•œ ํ›„ Test Data์—์„œ ๊ฐ€์žฅ ์ˆ˜๊ฐ€ ๋งŽ์€ ํด๋ž˜์Šค๋กœ ์ถ”์ •ํ•˜๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ๊ทผ์ ‘ํ•œ K๊ฐ’๊ณผ ์ด์›ƒ์˜ ํด๋ž˜์Šค๋ฅผ Test Data์˜ ํด๋ž˜์Šค๋กœ ์ถ”์ •ํ•˜๋Š” ๋ฐฉ์‹์œผ๋กœ 5๊ฒน ๊ต์ฐจ ๊ฒ€์ฆ์„ ํ†ตํ•ด ์ด์›ƒK์™€ ๊ฑฐ๋ฆฌ์ธก์ •๋ฐฉ์‹์œผ๋กœ ๊ฒฐ์ •ํ•œ๋‹ค.

์กฐ๋ฐ€KNN ๋ถ„๋ฅ˜ ๊ฒฐ๊ณผ๋Š” ๊ทผ์ ‘ ๊ฐœ์ˆ˜๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” K๊ฐ’๊ณผ ๊ฑฐ๋ฆฌ๋ฅผ ์ธก์ •ํ•˜๋Š” ๋ฐฉ์‹์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๊ฒŒ ๋˜๋ฉฐ ์ด๋“ค์„ Hyper-parameter ๋ผ ํ•œ๋‹ค. ๊ฑฐ๋ฆฌ ์ธก์ • ๋ฐฉ์‹์€ ๋Œ€ํ‘œ์ ์œผ๋กœ Euclidean๊ณผ Mahalonbis ๋ฐฉ์‹์ด๋ฉฐ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ์ˆ˜์‹์œผ๋กœ ๊ตฌ์„ฑ๋œ๋‹ค.

$d_{Euclidean(X,\: Y)}=\sqrt{(X-Y)^{T}(X-Y)} $

$d_{Mahalanobis(X,\: Y)}=\sqrt{(X-Y)^{T}\sum^{-1}(X-Y)} $

์—ฌ๊ธฐ์„œ X์™€ Y๋Š” ๊ฐ๊ฐ T์ธก์ •๋ฐ์ดํ„ฐ์™€ ํ•™์Šต๋ฐ์ดํ„ฐ์ด๊ณ  ฮฃ๋Š” ํ•™์Šต๋ฐ์ดํ„ฐ ์ „์ฒด์— ๋Œ€ํ•ด ๊ตฌํ•ด์ง„ ๊ฐ’์˜ ๋ฒกํ„ฐ๊ฐ„์˜ ๊ณต๋ถ„์‚ฐ ํ–‰๋ ฌ(Covariance Matrix)์„ ์˜๋ฏธํ•œ๋‹ค.

๋ณธ ์—ฐ๊ตฌ์—์„œ๋Š” MATHLAB์˜ Classification Learner ๊ธฐ๋Šฅ์„ ํ™œ์šฉํ•˜์—ฌ ๊ฐ€๋กœ์ถ•(x)์€ ์ฒจ๋„(Kurtosis), ์„ธ๋กœ์ถ•(y)์€ ์ „๋ฅ˜ ํ‰๊ท ์น˜๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ์‚ฐ์ ๋„(Scatter Plot)๋ฅผ ๊ทธ๋ ค ๋ณด๋ฉด ์•„๋ž˜ Fig. 8๊ณผ ๊ฐ™๋‹ค.

Fig. 8. Classification of kurtosis and current average values using dense knn method

../../Resources/kiee/KIEE.2021.70.1.096/fig8.png

Fig. 8์—์„œ ์ •์ƒ๋ชจํ„ฐ์— ๋น„ํ•˜์—ฌ ๋น„์ •์ƒ ๋ชจํ„ฐ์˜ ๊ฒฝ์šฐ๋Š” ํ‰๊ท ์ „๋ฅ˜์น˜(Average Current)์™€ ์ฒจ๋„(Kurtosis)๊ฐ€ ๋†’๊ฒŒ ๋‚˜ํƒ€๋‚จ์œผ๋กœ์จ ๋ชจํ„ฐ ๊ถŒ์„ ์˜ ๋‹จ๋ฝ(Short Circuit) ๋“ฑ์œผ๋กœ ๊ณผ๋ถ€ํ•˜๊ฐ€ ๋‚˜ํƒ€๋‚˜ ๊ณ ์žฅ์œผ๋กœ ์ง„ํ–‰๋  ํ™•๋ฅ ์ด ๋†’๋‹ค๊ณ  ํ•  ์ˆ˜ ์žˆ๋‹ค.

2.5 ์˜ˆ์ง€์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ํ‰๊ฐ€

์˜ˆ์ง€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ •ํ™•๋„ ํ‰๊ฐ€ ์ฒ™๋„๋Š” ํ˜ผํ•ฉ๋งคํŠธ๋ฆญ์Šค(Confusion matrix)๋ฅผ ์‚ฌ์šฉํ•˜์˜€์œผ๋ฉฐ ์„ธ๋กœ์ถ•์€ ํ•ด๋‹น๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ํด๋ž˜์Šค์ด๋ฉฐ ๊ฐ€๋กœ์ถ•์€ ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ์˜ˆ์ธก๋œ ํด๋ž˜์Šค์ด๋‹ค.

๋Œ€๊ฐ ์„ฑ๋ถ„ (TN ๊ณผ TP)์€ ๋ฐ์ดํ„ฐ์˜ ์‹ค์ œ ํด๋ž˜์Šค์™€ ์˜ˆ์ธกํด๋ž˜์Šค๊ฐ€ ๋™์ผํ•œ ๊ฒฝ์šฐ์˜ ์ˆ˜์ด๊ณ  ๋‚˜๋จธ์ง€ ์„ฑ๋ถ„(FP ์™€ FN)์€ ์‹ค์ œํด๋ž˜์Šค์™€ ์˜ˆ์ธก๋œ ํด๋ž˜์Šค๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ ๊ฒฝ์šฐ์˜ ์ˆ˜์ด๋‹ค

์ •ํ™•๋„๋Š” ์ „์ฒด ๊ฒฝ์šฐ์˜ ์ˆ˜์— ๋Œ€ํ•œ ๋Œ€๊ฐ ์„ฑ๋ถ„์˜ ์ˆ˜์ด๋‹ค.

Accuracy(์ •ํ™•๋„)=(์ฐธ์˜ˆ์ธก์œจTP+์ฐธ๋ฐœ๊ฒฌ์œจTN)/(์ฐธ์˜ˆ์ธก์œจTP+๊ฑฐ์ง“์˜ˆ์ธก์œจFP+๊ฑฐ์ง“๋ฐœ๊ฒฌ์œจFN+์ฐธ๋ฐœ๊ฒฌ์œจTN)๋ฅผ ๋‚˜ํƒ€๋‚ธ๋‹ค.

์ž์„ธํ•œ ๋‚ด์šฉ์€ ์•„๋ž˜ Table 2์™€ ๊ฐ™๋‹ค.

Table 2. Confusion matrix for Classification accuracy measurement

Model

Training class

Remark

-

+

True(False) Negative

Test Class

-

TN

FN

+

FP

TP

Precision

Remark

True(false)Positive

Recall

Accuracy

โ€ปIndex: TN: True Negative, TP : True Positive, FN: False Negative, FP: False Positive, Precision=TP/(TP+TN);The percentage of the results predicted as actual positives that are truly positive, Recall=TP/(TP+FN);Percentage of correctly predicting a positive target as a positive

ํ•™์Šต๊ฒฐ๊ณผ๋Š” ์•„๋ž˜ Table 3๊ณผ ๊ฐ™์ด ์กฐ๋ฐ€ knn ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ •ํ™•๋„๋Š” 99.7% ๋กœ ๋‚˜ํƒ€๋‚ฌ๋‹ค.

Table 3. Accuracy Measurement result of dense KNN algorithm

Training

Test

True Positive Rate(TPR)*, (False Negative Rate(FNR)**

1

2

3

4

1

99.9%*

0.0%**

0.1%**

0.0%**

2

0.0%**

99.6%*

0.1%**

0.4%**

3

0.3%**

0.0%**

99.6%*

0.0%**

4

0.0%**

0.5%**

0.0%**

99.5%*

์ƒ๊ธฐ Table 3์—์„œ ๊ฐ€๋กœ์ถ•์€ ์˜ˆ์ธกํด๋ž˜์Šค๋กœ์„œ ๋ฐ์ดํ„ฐ๋ฅผ ํ•™์Šตํ•˜์—ฌ 4๊ฐœ ํด๋ž˜์Šค(1์ •์ƒ์—ด๋ฆผ, 2์—ด๋ฆผ๋น„์ •์ƒ, 3์ •์ƒ๋‹ซํž˜, 4๋‹ซํž˜๋น„์ •์ƒ)๋กœ ๊ตฌ๋ถ„ํ•˜์˜€์œผ๋ฉฐ ์„ธ๋กœ์ถ•์€ ์‹ค์ œ๋ฐ์ดํ„ฐ๋ฅผ 4๊ฐœํด๋ž˜์Šค๋กœ ๊ตฌ๋ถ„ํ•˜์˜€๋‹ค.

3. ๊ฒฐ ๋ก 

์ถœ์ž…๋ฌธ ๋ชจํ„ฐ์˜ ์ „๋ฅ˜๊ฐ’์„ ์‹œ๊ฐ„์˜์—ญ(Time Domain)์˜ ํ†ต๊ณ„์  ์œ ์˜์„ฑ์„ ๋ถ„์„ํ•œ ๊ฒฐ๊ณผ ์ฒจ๋„(Kutorsis) ์ธ์ž๊ฐ€ ์œ ์˜ํ•œ ๊ฒƒ์œผ๋กœ ํ™•์ธ๋˜์—ˆ์œผ๋ฉฐ ์ง€๋„ํ•™์Šต์„ ํ™œ์šฉํ•˜์—ฌ ๋ถ„๋ฅ˜(Classification) ๊ธฐ๋ฒ•์ธ Machine Learing์ค‘์— ์กฐ๋ฐ€ knn ์•Œ๊ณ ๋ฆฌ์ฆ˜์œผ๋กœ ์ถœ์ž…๋ฌธ ๋ชจํ„ฐ ์ „๋ฅ˜์น˜๋ฅผ 4๊ฐœ ํด๋ž˜์Šค๋กœ ๋ถ„๋ฅ˜ํ•˜์—ฌ(์ •์ƒ์—ด๋ฆผ, ๋น„์ •์ƒ ์—ด๋ฆผ, ์ •์ƒ ๋‹ซํž˜, ๋น„์ •์ƒ ๋‹ซํž˜) ๋ถ„์„ํ•ด ๋ณธ ๊ฒฐ๊ณผ ์ •์ƒํด๋ž˜์Šค์— ๋น„ํ•˜์—ฌ ๋น„์ •์ƒํด๋ž˜์Šค์— ์žˆ๋Š” ๋ฐ์ดํ„ฐ๊ฐ€ ํ‰๊ท  ์ „๋ฅ˜์น˜๊ฐ€ ๋†’์€ ๊ฒƒ์œผ๋กœ ํ™•์ธ๋˜์—ˆ๋‹ค.

ํ‰๊ท ์ „๋ฅ˜์น˜๊ฐ€ ๋†’์€ ๊ฒฝ์šฐ๋Š” ํ–ฅํ›„ ๋ชจํ„ฐ์˜ ๋‹จ๋ฝ(Short Circuit)๋“ฑ์œผ๋กœ ์ธํ•œ ๊ณผ๋ถ€ํ•˜๊ฐ€ ๋ฐœ์ƒํ•˜์—ฌ ๋ชจํ„ฐ์˜ ๊ณ ์žฅ์„ ์œ ๋ฐœํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋‹ค.

ํ–ฅํ›„์—๋Š” ๋ถˆ๋Ÿ‰ ๋ชจํ„ฐ์˜ ์ „๋ฅ˜ ํŒŒํ˜•์„ ๋ถ„์„ํ•˜์—ฌ ๋ชจํ„ฐ์˜ ๊ณ ์žฅMode๋ณ„ ๊ณ ์žฅ ํ˜„์ƒ๊ณผ Mappingํ•˜์—ฌ ๊ณ ์žฅ์ด ๊ฐ์ง€๋˜์ง€ ์•Š์•˜๋Š”๋ฐ ๊ณ ์ „๋ฅ˜ ์ €์ „๋ฅ˜ ๋“ฑ์˜ ์ด์ƒ ํ˜„์ƒ์ด ์ง€์†๋˜๋Š” ๊ฒฝ์šฐ ์ด์ƒ ์‹œ์ ์˜ ์ฃผํŒŒ์ˆ˜๋ฅผ ๋ถ„์„ํ•˜์—ฌ ๋ชจํ„ฐ์˜ ๋‹จ๋ฝ(Short Circuit) ๋“ฑ ์ด์ƒ ์ƒํƒœ๋ฅผ ํ™•์ธํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋‹ค.

์ด๋ ‡๊ฒŒ ์˜ˆ์ธกํ•œ ๋ฐ์ดํ„ฐ์™€ ์‹ค์ œ ๊ฒฝํ—˜ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ถ„์„ํ•œ ๋ชจํ„ฐ์˜ ๋ถ€ํ’ˆ, ์ฆ‰ ์ •๋ฅ˜์ž, ๋ธŒ๋Ÿฌ์‹œ, ์ „๊ธฐ์ž ๋“ฑ์˜ ์–ด๋–ค ๋ถ€ํ’ˆ์ด ๊ณ ์žฅ์ธ์ง€ ๋Œ€์šฉ๋Ÿ‰์˜ ๊ณ ์žฅ ํ˜„์ƒ๋ฐ์ดํ„ฐ์™€ ์ƒํ˜ธ ๊ด€๋ จ์„ฑ์„ ๊ฒ€์ฆํ•˜์—ฌ ์ƒํƒœ๊ธฐ๋ฐ˜ ์˜ˆ์ง€ ์ •๋น„๋ฅผ ํ•˜๊ณ  ์šด์˜ ๋ฐ์ดํ„ฐ ๊ธฐ๋ฐ˜์˜ ๋Œ€์šฉ๋Ÿ‰ ๋ฐ์ดํ„ฐ๋ฅผ ์ถ•์ ํ•˜๋ฉด ์ž”์กด ์ˆ˜๋ช…์„ ์˜ˆ์ธก ํ•  ์ˆ˜ ์žˆ์„ ๊ฒƒ์ด๋‹ค.

References

1 
Kwangsung Choi, 2018, Maintenance application study through analysis of electric RAM system on Metro Railway Line No.2, Seoul National University of Science and Techno- logy Masterโ€™s ThesisGoogle Search
2 
Sangjin Lee, 2018, Research on KTX high-speed Rolling Stock traction Inverter failure analysis and defect detection method, Seoul National University of Science and Tech- nology Master's ThesisGoogle Search
3 
Hyunkyu Lee, 2016, Design of maintenance support system based on Rolling Stock major parts failure forecast and soundness management technology, Electronics and Tele- communications Research Institute of KoreaGoogle Search
4 
Kyusung Jung, Jooho Choi, 2017, A Study on the model-based failure of electric models using current signal analysis, Collection of the Spring Conference in the Reliability department of the Korea Society of MachineryGoogle Search
5 
Jongman Kim, 2014, Development of FMMEA procedures for assessing the soundness of fault deposits in power modules, Myongji University Master's ThesisGoogle Search
6 
SeokJoo Ham, others 5 people, 2019, Development of soundness diagnosis algorithm for electric Rolling Stock door using motor current signal, PHM Conference ThesisGoogle Search
7 
Euyseok Hong, Unsupervised Learning Model for Fault Prediction Using Representative Clustering Algorithms, KIPS Tr. Software and Data Eng, Vol. 3, No. 2, pp. 57-64DOI
8 
Adebena Oluwasegun, Jae-Cheon Jung, The application of machine learning for the prognostics and health manage- ment of control element drive system, Nuclear Engineering and Technology, Vol. 52, pp. 2267-2273DOI
9 
Jungyeon Sim, 2016, PHM concepts and practices, Journal of the Korean Society of ReliabilityGoogle Search
10 
Mangi Lee, 2020, Door failure investigation by door type, Seoul National University of Science and Technology Masterโ€™s ThesisGoogle Search
11 
Sanghuun Kim, 2020, A styudy on the failure of plug-in type electric Rolling Stock door using moving average method, Doctoral Thesis, Graduate School of Transportation, Korea University of TransportationGoogle Search
12 
Byungchul Jeon, 2014, Statistical Approach to Diagnostic Rules for Various Malfunctions of Journal Bearing System Using Fisher Discriminant Analysis, European Conference of the Prognostics and Health Management SocietyGoogle Search

์ €์ž์†Œ๊ฐœ

Jong-Kook Lim
../../Resources/kiee/KIEE.2021.70.1.096/au1.png

PhD completion at Seoul National University of Science and Technology.

Consulting Div.

Executive Director, Bizpeer Co., LTD.Professional Engineer Railroad Rolling Stock.

E-mail : jklim@seoultech.ac.kr

Hee-Jung Yoon
../../Resources/kiee/KIEE.2021.70.1.096/au2.png

Director, Daeyun Engineering Co., LTD Rolling Stock Door Control Unit Responsible Researcher DCU Data Acquisition Responsibilities

E-mail : yhj@dataok.co.kr