Structural health monitoring for offshore wind turbine foundations through unsupervised and semi supervised machine learning methods

END OF DEGREE PROJECT Deg ee in Chemical Enginee ing STRUCTURAL HEALTH MONITORING FOR OFFSHORE WIND TURBINE FOUNDATIONS THROUGH UNSUPERVISED AND SEMI SUPERVISED MACHINE LEARNING METHODS Repo and Annexes Au ho : Cla a Rull Di ec o : Yolanda Vidal Call: May 2022 SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods i Abs ac The cu en clima e c isis equi es a shi owa ds enewable ene gies. Wind ene gy gene a ion will play a majo ole. O sho e wind ene gy can p o ide g ea e ou pu due o mo e p edic able wea he condi ions compa ed o onsho e wind ene gy and has one o he lowes li ecycle g eenhouse gas emissions o any sou ce o ene gy. Some o he di icul ies in hei ope a ion and main enance lie in he di icul y o accessing he si e. Al hough emo e moni o ing has become s anda d in he indus y, s uc u al heal h moni o ing and p edic i e main enance s ill p esen some challenges. No mally, mos o all he a ailable da a a e o egula ope a ion, hus me hods ha ocus on he da a leading o ailu es end up using only a small subse o he a ailable da a. Fu he mo e, when he e is no his o ical p eceden o a ype o damage, hose me hods canno be used. In addi ion, o sho e wind u bines wo k unde a wide a ie y o en i onmen al condi ions and egions o ope a ion in ol ing unknown inpu exci a ion gi en by he wind and wa es. Finally, supe ised app oaches ely on co ec ly labelling da a, which is no possible in p oduc ion condi ions. Conside ing he di icul ies, he s a ed s a egy in his wo k is based on unsupe ised and semi-supe ised app oaches and i wo ks unde di e en ope a ing and en i onmen al condi ions based only on he ou pu ib a ion da a ga he ed by accele ome e senso s. The p oposed s a egy has been es ed h ough expe imen al labo a o y es s on a down-scaled model. This p ojec applies spec al en opy, a non-s anda d pa ame e in ib a ion analysis, o he s udied models. O e all accu acies o 93,88% o Isola ion Fo es (a semi-supe ised me hod), and 88,67% o One Class Suppo Vec o Machine (a non-supe ised me hod) can be achie ed. The accu acies o bo h models inc ease o up o 100% when ained agains a la ge da ase o heal hy samples, howe e achie ing hese esul s equi es e uning o ea u es and hype pa ame e s. Fo all o his, he use o non-supe ised and semi-supe ised machine lea ning models is a ealis ic app oach o s uc u al heal h moni o ing o o sho e wind u bines and has ob ained p omising esul s when es ed agains an expe imen al da ase . Repo ii Resum La c isi climà ica ac ual eque eix un gi cap a les ene gies eno ables. La gene ació d'ene gia eòlica hi juga à un pape impo an . L'ene gia eòlica ma ina po p opo ciona una majo p oducció degu a condicions climà iques més p e isibles en compa ació amb l'ene gia eòlica e es e i é una de les emissions de gasos d'e ec e hi e nacle de cicle de ida més baixes en compa ació amb qualse ol on d'ene gia. Algunes de les di icul a s en el seu uncionamen i man enimen adiquen en la di icul a d'accés al lloc. Si bé la moni o i zació emo s'ha olgu es ànda d a la indús ia, la moni o i zació de la salu es uc u al i el man enimen p edic iu enca a p esen a algunes di icul a s. No malmen , la majo ia o o es les dades disponibles són de l’ope ació egula , pe an els mè odes en oca s en la u ili zació de les dades p eceden s a alles acaban u ili zan només un pe i subconjun de les dades disponibles. A més, quan no hi ha an eceden s his ò ics d'un ipus de dany, no es poden u ili za aques s mè odes. Enca a, les u bines eòliques ma ines uncionen en una amplia a ie a de condicions ambien als i egions d'ope ació que in oluc en una exci ació d'en ada desconeguda p opo cionada pel en i les onades. Finalmen , els en ocamen s supe isa s es basen en l'e ique a ge co ec e de les dades, que no és possible en condicions de p oducció. Tenin en comp e les di icul a s, l'es a ègia es able a en aques eball es basa en en ocamen s no supe isa s i semi-supe isa s i unciona so a di e en s condicions ambien als i ope a i es basan -se únicamen en les dades de ib ació de so ida ecopilades pels accele òme es. L’es a ègia ha es a p o ada a a és d’assajos expe imen als de labo a o i en un model a escala eduïda. Aques p ojec e aplica l'en opia espec al, un pa àme e no es ànda d en l'anàlisi de ib acions, als models es udia s. Es poden aconsegui p ecisions gene als del 93,88 % pe a ‘Isola ion Fo es ’ (un mè ode semi supe isa ) i del 88,67 % pe a ‘One Class Suppo Vec o Machine’ (un mè ode no supe isa ). Les p ecisions dels dos models augmen en ins al 100 % quan s'en enen amb un conjun de dades més g ans de mos es sanes; anma eix, pe aconsegui aques s esul a s és necessa i o na a ajus a les ‘ ea u es’ i els hipe pa àme es. Pe o això, l’ús de models no supe isa s i semi supe isa s és un en oc ealis a pe la moni o i zació es uc u al de les u bines de en ma ines ob enin esul a s p ome edo s quan s’ha p o a con a un conjun de dades expe imen al. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods iii Resumen La ac ual c isis climá ica equie e un gi o hacia las ene gías eno ables. La gene ación de ene gía eólica juga á un papel impo an e. La ene gía eólica ma ina puede p opo ciona una mayo p oducción debido a las condiciones climá icas más p edecibles en compa ación con la ene gía eólica e es e y iene una de las emisiones de gases de e ec o in e nade o de ciclo de ida más bajas en compa ación cualquie uen e de ene gía. Algunas de las di icul ades en su uncionamien o y man enimien o adican en la di icul ad de acceso al si io. Si bien el moni o eo emo o se ha uel o es ánda en la indus ia, el moni o eo de la salud es uc u al y el man enimien o p edic i o aún p esen a algunos desa íos. No malmen e, la mayo ía o odos los da os disponibles son de ope ación egula , po lo que los mé odos que se en ocan en los da os que conducen a allas e minan usando solo un pequeño subconjun o de los da os disponibles. Además, cuando no exis e un an eceden e his ó ico de un ipo de daño, no se pueden u iliza esos mé odos. Po añadido, las u binas eólicas ma inas uncionan en una amplia a iedad de condiciones ambien ales y egiones de ope ación que in oluc an una exci ación de en ada desconocida p opo cionada po el ien o y las olas. Finalmen e, los en oques supe isados se basan en el e ique ado co ec o de los da os, que no es posible en condiciones de p oducción. Teniendo en cuen a las di icul ades, la es a egia es ablecida en es e abajo se basa en en oques no supe isados y semi supe isados y unciona bajo di e en es condiciones ope a i as y ambien ales basadas solo en los da os de ib ación de salida ecopilados po los senso es del acele óme o. La es a egia p opues a ha sido p obada a a és de p uebas expe imen ales de labo a o io en un modelo a escala educida. Es e p oyec o aplica la en opía espec al, un pa áme o no es ánda en el análisis de ib aciones, a los modelos es udiados. Se pueden log a p ecisiones gene ales del 93,88 % pa a ‘Isola ion Fo es ’ (un mé odo semi supe isado) y del 88,67 % pa a ‘One Class Suppo Vec o Machine’ (un mé odo no supe isado). Las p ecisiones de ambos modelos aumen an has a un 100 % cuando se en enan con un conjun o de da os más g ande de mues as sanas; sin emba go, pa a log a es os esul ados es necesa io ol e a ajus a las ‘ ea u es’ y los hipe pa áme os. Po odo es o, el uso de modelos de ap endizaje au omá ico no supe isados y semi supe isados es un en oque ealis a pa a el moni o eo de la salud es uc u al de las u binas eólicas ma inas y ha ob enido esul ados p ome edo es cuando se p ueba con un conjun o de da os expe imen al. Repo i Glossa y AI: A i icial in elligence CF: C es ac o ML: Machine lea ning O&G: oil and gas O&M: ope a ion and main enance PP: Peak-peak RBF: Radial basis unc ion RBM: eliabili y-based main enance RMS: Roo mean squa ed SHM: S uc u al heal h moni o ing SVM: Suppo ec o machine TPM: o al p oduc i e main enance WT: Wind u bine ZP: Ze o-peak SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods Index ABSTRACT ___________________________________________________________ I GLOSSARY __________________________________________________________ IV INDEX OF FIGURES __________________________________________________ VII INDEX OF TABLES ___________________________________________________ VIII 1. INTRODUCTION _________________________________________________ 3 1.1. Goals o he p ojec .................................................................................................. 3 2. WIND ENERGY __________________________________________________ 5 2.1. O -Sho e wind u bines .......................................................................................... 6 2.2. Componen s o Wind Tu bine Ins alla ions............................................................. 8 2.2.1. D i e ain ................................................................................................................ 9 2.2.2. Founda ion ............................................................................................................. 9 2.2.3. S uc u al Suppo ................................................................................................ 10 2.2.4. Floa ing sys ems ................................................................................................... 11 3. MAINTENANCE THEORY _________________________________________ 12 3.1. Main enance S a egies ......................................................................................... 12 3.2. Key pe o mance indica o s in main enance ........................................................ 13 3.3. Main enance in o sho e wind u bines ................................................................ 14 4. VIBRATION ANALYSIS ___________________________________________ 15 4.1. Da a acquisi ion ..................................................................................................... 15 4.2. Vib a ion Signals ..................................................................................................... 15 4.2.1. Rele an ea u es o ib a ion signals .................................................................. 15 4.3. Complex me hods .................................................................................................. 16 5. APPLIED ARTIFICIAL INTELLIGENCE _________________________________ 17 5.1. Machine Lea ning................................................................................................... 17 5.2. Ca ying ou Machine Lea ning P ojec s ............................................................... 18 5.3. Models in Machine Lea ning ................................................................................. 20 5.3.1. One-Class SVM ...................................................................................................... 21 5.3.2. Isola ion o es ...................................................................................................... 23 5.4. Valida ion me ics in Machine Lea ning ................................................................ 26 5.5. Applica ions o Machine Lea ning in Enginee ing ................................................. 26 Repo i 5.5.1. Manu ac u ing Indus y ........................................................................................ 26 5.5.2. Ene gy indus y ..................................................................................................... 26 6. VIBRATION ANALYSIS AND ML MODEL APPLICATION TO EXPERIMENTAL DATA27 6.1. Da a collec ion ....................................................................................................... 27 6.2. Da a ans o ma ion .............................................................................................. 28 6.3. Fea u es ................................................................................................................. 29 6.4. Model aining ....................................................................................................... 29 6.5. Model alida ion .................................................................................................... 30 6.6. Model alida ion wi h eplica ba ......................................................................... 32 6.7. Re ained model alida ion wi h eplica ba ......................................................... 33 6.8. Re uning he model o eplica ba ....................................................................... 35 6.9. Conclusion .............................................................................................................. 37 7. ENVIRONMENTAL IMPACT ANALYSIS _______________________________ 38 CONCLUSIONS ______________________________________________________ 39 ECONOMIC ANALYSIS ________________________________________________ 41 BIBLIOGRAPHY _____________________________________________________ 43 SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods ii Index o igu es Figu e 1 Median emissions o selec ed elec ici y supply echnologies __________________________________ 5 Figu e 2 Wind elec ici y gene a ion, Wo ld (1990-2019) (IEA 2018) ___________________________________ 6 Figu e 3 ECMFW wind ield da a a e co ec ion o o og aphy and local oughness (Eu opean En i onmen Agency 2009) ______________________________________________________________________________ 7 Figu e 4 To al ins alled capaci y (IC) o o sho e wind ene gy by egion (Ba helmie and P yo 2021) _________ 8 Figu e 5 Componen s o an o sho e wind u bine (Li, e al. 2022) _____________________________________ 9 Figu e 6 Main ypes o o sho e wind u bine ounda ions (Xie and Lopez-Que ol 2021) ___________________ 10 Figu e 7 Jacke s uc u e. (A) Scheme (B) Jacke ounda ion anspo a ion (Alpha Ven us wind a m) (C) Jacke ounda ions ins alled (Alpha Ven us wind a m) (Manzano-Aguglia o, e al. 2020) _______________________ 11 Figu e 8 S eps o a Machine Lea ning p ojec ____________________________________________________ 19 Figu e 9 SVM decision bounda y and suppo ec o s (Wang, y o os 2019) ____________________________ 21 Figu e 10 O iginal and ke nelized ea u e space (Rizwan, e al. 2021) _________________________________ 22 Figu e 11 Illus a ion o non-linea ke nel ans o ma ions (Ez a Pila io, e al. 2020) _____________________ 23 Figu e 12 Example o a andom ee in an Isola ion Fo es Model ____________________________________ 24 Figu e 13 Sca e plo and decision bounda ies o a andom decision ee in an Isola ion T ees model _______ 25 Figu e 14 (a) The bench es de ailing he loca ion o he ba , and(b) Loca ion o he senso s (Hoxha, Vidal and Pozo 2020) _______________________________________________________________________________ 27 Figu e 15 Visualisa ion o Isola ion Model selec ed o u he analysis ________________________________ 31 Figu e 16 Visualisa ion o One Class SVM selec ed o u he analysis ________________________________ 32 Figu e 17 Selec ed Isola ion Fo es model alida ed agains eplica class ______________________________ 33 Figu e 18 Selec ed One Class SVM model alida ed agains eplica class _______________________________ 33 Figu e 19 Selec ed Isola ion Fo es model ained wi h eplica class __________________________________ 34 Figu e 20 Selec ed One Class SVM model ained wi h eplica class ___________________________________ 35 Figu e 21 Isola ion Fo es ained wi h eplica class, S anda d De ia ion, Ku osis and 50 es ima o s ________ 36 Figu e 22 Bes Pe o ming One Class SVM wi h eplica class ________________________________________ 36 Memo ia 6 Figu e 2 Wind elec ici y gene a ion, Wo ld (1990-2019) (IEA 2018) 2.1. O -Sho e wind u bines O sho e wind powe is a subse o wind powe , whe e wind u bines a e placed on bodies o wa e (usually seas o oceans, bu also in lakes). O sho e wind u bines bene i om highe and mo e p edic able wind speeds. Howe e , hey also p esen highe ope a ion and main enance (O&M) cos s compa ed o onsho e wind u bines. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 7 Figu e 3 ECMFW wind ield da a a e co ec ion o o og aphy and local oughness (Eu opean En i onmen Agency 2009) Globally, in 2020 o sho e wind capaci y passed 35GW and now ep esen s 4.8% o o al cumula i e wind capaci y. GWEC Ma ke In elligence expec s ha o e 469 GW o new onsho e and o sho e wind capaci y will be added in he nex i e yea s - ha is nea ly 94 GW o new ins alla ions annually un il 2025, based on p esen policies and pipelines. (Global Wind Ene gy Council 2021) Cu en ly Eu ope has o e 25GW o o sho e wind ene gy capaci y, wi h a o al o 5402 g id connec ed wind u bines deli e ing powe om 116 o sho e wind a ms in 12 Eu opean coun ies. Memo ia 8 Figu e 4 To al ins alled capaci y (IC) o o sho e wind ene gy by egion (Ba helmie and P yo 2021) 2.2. Componen s o Wind Tu bine Ins alla ions This p ojec ocuses on ho izon al axis upwind u bines ins alled o sho e. In ho izon al axis wind u bines, he axis ha is connec ed o he main bea ing o elec ici y gene a ion is pa allel o he g ound and he main o o is di ec ed owa ds he wind. The design o wind u bines in o sho e mus conside he ha she condi ions compa ed o onsho e wind u bines: • S ong cu en s and wa es. • Co osi e en i onmen s. • Ha sh clima ological condi ions, s onge s o ms, and winds. Typically, he u bine manu ac u e p o ides he o o-nacelle assembly and he owe . The suppo s uc u e and base a e chosen acco ding o he needs o he p ojec (Bha acha ya 2019). SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 9 Figu e 5 Componen s o an o sho e wind u bine (Li, e al. 2022) 2.2.1. D i e ain In wind u bines he powe is ansmi ed om he o o o he gene a o h ough he sys em composed o he main sha , ic ion connec ion, mul iplying gea box and a lexible coupling. This whole sys em is known as he d i e ain. (Michal, Gawa kiewicz and Wasilczuk 2015) The d i e ain may ha e a gea box be ween ha main o o and gene a o o inc ease he o a ional speed o he o o o gene a o speeds. This is he mos common design as i allows o use o s anda d componen s. Less equen ly, d i e ains may be gea less, equi ing a mul i-pole gene a o . (Ba szcz 2019) 2.2.2. Founda ion The ounda ions o wind u bines can be classi ied in wo main g oups: g ounded sys ems and loa ing sys ems. Founda ions can be classi ied as shallow base o deep base. Some examples a e he ollowing: • Monopile s uc u es a e deep base s uc u es, whe e a long s eel cylinde o 3 o 7m o diame e is placed up o 40m in o he ocean loo . These a e he mos common kind o ounda ion. • Shallow ounda ion s uc u es, designed o a oid ensions be ween he ounda ion s uc u e and he seabed, in o de o a oid o sion. Memo ia 10 • Suc ion based ounda ions a e mo e shallow han monopola s uc u es bu deepe han g a i y-based ones. They a e o med by a ubula s uc u e opped by a ci cula side ha ac s like a suc ion cup, a aching o he seabed. 2.2.3. S uc u al Suppo O sho e wind u bines equi e mo e obus suppo s uc u es han onsho e wind u bines, due o he ex eme condi ions a sea. Figu e 6 Main ypes o o sho e wind u bine ounda ions (Xie and Lopez-Que ol 2021) Suppo s uc u es may be: monopile (essen ially an ex ension o a pile ounda ion), ipile, ipod, g a i y based/shallow ounda ion o jacke ed/la iced. This p ojec ocuses on he SHM o jacke s uc u es, mo e in-dep h desc ip ion o hem can be ound in he nex sec ion. Jacke ed o la iced s uc u e The his o ical p eceden o jacke s uc u es in o sho e wind ounda ions a e gas and oil ex ac ion pla o ms. Howe e , hei use as a s uc u al suppo o wind u bines p esen s some speci ic pa icula challenges, he mos p ominen one being a signi ican con ibu ion o ib a ions due o he impac o wind, while in oil and gas (O&G) ex ac ion pla o ms wa es a e he mos signi ican ib a ion con ibu ion. These s uc u es ypically ha e 4 suppo s, which will ha e pile, g a i y bases o suc ion caissons. The e has been inc eased in e es in he use o 3-legged jacke s uc u es as hey p esen lowe cos s. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 11 As o he dimensions o he jacke s uc u e, mo e adi ional app oaches ely on in eg a ed ae oelas ic models wi h a simpli ied ep esen a ion o he ounda ion o calcula ions. (Agus yn, Nielsen and Pede sen 2017) Some models o a sys ema ic app oach o he p edesign phase ha e been de eloped, bu u he wo k om expe ienced p o essionals is s ill equi ed o a comple e design. (Hä ele, e al. 2018) Figu e 7 Jacke s uc u e. (A) Scheme (B) Jacke ounda ion anspo a ion (Alpha Ven us wind a m) (C) Jacke ounda ions ins alled (Alpha Ven us wind a m) (Manzano-Aguglia o, e al. 2020) 2.2.4. Floa ing sys ems The e has been inc easing in e es in loa ing sys ems o be used when he dep h exceeds a ound 60m. • Moo ing s abilised TLP ( ension leg pla o m) concep • Ballas s abilised Spa buoy • Buoyancy s abilised semi-subme sible is a combina ion o he p e ious wo app oaches. Al hough some o sho e wind p ojec s wi h loa ing sys ems ha e been deployed in Sco land (Sco land Hywind) and No way (Equino Tampen), hey a e s ill in he mino i y. Memo ia 12 3. Main enance heo y Main enance is a highly scoped subjec , ha includes bu is no limi ed o he main enance o buildings, he eme gency epai s o machines damaged du ing indus ial acciden s and he moni o isa ion o equipmen in any indus y. Many companies ha e s a ed o implemen me hodologies such as Six-Sigma, o Jus in Time in an e o o ul il he cus ome demands o high-quali y p oduc s in a imely manne . This has esul ed in a shi o hei manu ac u ing, o ganiza ional, and supply chain s a egies owa d agili y, quali y, au oma ion, and high pe o mance. This has esul ed in e y high in es men s in equipmen and people. To achie e he a ge ed a es o e u n-on-in es men equipmen mus be eliable and sa e o ope a e wi hou cos ly wo k s oppages and epai s. (Du uaa and Raou 2015) In he ene gy indus y, he g owing global ene gy demand, and he inabili y o s o e excess ene gy a a la ge scale ha e esul ed in he need o minimize down ime in ene gy p oduc ion sys ems, anging om nuclea eac o s o sola panels. These changes ha e shi ed he pe cep ion o main enance om a necessa y e il o a key ac i i y in manu ac u ing and ene gy p oduc ion. The equi emen s o agili y, quali y, au oma ion, and high pe o mance ha e led o he implemen a ion o main enance me hodologies like o al p oduc i e main enance (TPM), eliabili y cen ed main enance (RCM), o lean six sigma. 3.1. Main enance S a egies In his p ojec main enance will e e o conse a i e main enance, ha is, main enance ha is in ended o p ese e he unc ionali y o a sys em. Howe e , main enance may also include imp o emen main enance, o e haul main enance, eme gency main enance and o he s. Se e al di e en main enance s a egies ha e been de eloped since he indus ial e olu ion, wi h inc easing echnical complexi ies, le e aging he la es echnical de elopmen s in s a is ical analysis and moni o ing capabili ies. Mo e simple main enance s a egies mus no be dis ega ded as i is usual o se e al di e en s a egies o coexis in he main enance plan o any sys em. • Co ec i e main enance: Main enance ac ions a e ca ied ou a e a b eakdown. Up on cos s o his ype o main enance a e non-exis en , howe e , long e m and o expensi e pieces o equipmen , i may esul in e y high cos s. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 13 • P e en i e main enance: Main enance ac ions a e ca ied ou a p ede e mined in e als o ime o wea . This app oach leads o less equipmen down ime and longe asse li e, howe e i is also mo e labou -in ensi e and he e is po en ial o o e -main enance. • Condi ion-based main enance o p edic i e main enance: P e en i e main enance ha is ini ia ed because o knowledge o he condi ion equipmen h ough ou ine (discon inuous) o con inuous moni o ing. This app oach leads o a dec ease o main enance cos s o 30% on a e age (Schallehn, e al. 2018) and educes he equency o b eakdowns by abou 75% (PwC 2018). The complexi ies in he implemen a ion o p edic i e main enance sys ems in mos indus ies a ise om di icul ies in de eloping he models and implemen ing he in as uc u e equi ed o condi ion moni o ing acking. 3.2. Key pe o mance indica o s in main enance Di e en indus ies will ha e di e en de ini ions o success in main enance, a common way o de ine success in ela i ely s anda dised way a e Key Pe o mance Indica o s, o KPIs. Some common KPIs a e as ollows: • Mean ime Be ween Failu es (MTBF) is he a e age amoun o ime be ween b eakdowns. The de ini ion o a b eakdown can di e . In he case o O sho e WT his is a specially ele an me ic as se ice ips o he u bine a ms can be cos ly and ha e a high logis ical complexi y. 𝑀𝑇𝐵𝐹= 𝑇𝑜𝑡𝑎𝑙 𝑊𝑜𝑟𝑘𝑖𝑛𝑔 𝐻𝑜𝑢𝑟𝑠 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑎𝑖𝑙𝑢𝑟𝑒𝑠 • Mean ime o epai (MTTR) is he amoun o ime ha i akes, on a e age, o e u n a piece o equipmen o wo king condi ions a e a b eakdown. 𝑀𝑇𝑇𝑅= 𝑇𝑜𝑡𝑎𝑙 𝑟𝑒𝑝𝑎𝑖𝑟 𝑡𝑖𝑚𝑒 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑎𝑖𝑙𝑢𝑟𝑒𝑠 • A ailabili y measu es he pe cen age o ime ha equipmen is in wo king condi ions. I gi es an idea o he up ime o a piece o equipmen . I is especially ele an in enewable ene gy gene a ion (speci ically sola and wind) as a eadiness me ic o he use o a ou able wind condi ions, as he h oughpu elies on ex e nal a iable ac o s (e. g. me eo ology). 𝐴𝑣𝑎𝑖𝑙𝑖𝑏𝑖𝑙𝑖𝑡𝑦= 𝑀𝑇𝐵𝐹 𝑀𝑇𝐵𝐹+𝑀𝑇𝑇𝑅 Memo ia 14 • O e all equipmen e ec i eness (OEE): i is a measu e o quali y, pe o mance and a ailabili y equen ly used in manu ac u ing. The highes sco e (100%) is ob ained when equipmen is ope a ing a he highes pe o mance (numbe o pieces p oduces pe ime uni ), wi h no de ec i e pieces and no una ailabili y e en s. 𝑂𝐸𝐸=𝐴𝑣𝑎𝑖𝑙𝑖𝑏𝑖𝑙𝑖𝑡𝑦·𝑄𝑢𝑎𝑙𝑖𝑡𝑦·𝑃𝑒𝑟𝑓𝑜𝑟𝑚𝑎𝑛𝑐𝑒 These me ics o e a way o objec i ely compa e di e en main enance s a egies and he eliabili y o equipmen . 3.3. Main enance in o sho e wind u bines The inc ease in o sho e wind u bine ins alla ions has led o a enewed in e es o new and ad anced echniques o main enance o wind u bines. (Cos a, e al. 2021) Ope a ion and main enance cos s ep esen 25% o ene gy p oduc ion cos s o sho e wind u bine main enance. This is 15% mo e han O&M cos s o onsho e wind u bines. The eason behind his di e ence is he echnical and logis ical complexi y o main enance ope a ions o o sho e wind a ms, which his highe han o onsho e wind u bines. Se ice isi s o o sho e wind a ms occu app oxima ely once e e y 6 mon hs (Fauls ish, Hahn y Ta ne 2011) and up o 5 imes pe and equi e 40 o 80 o man-hou s o se ice. Two decades ago, he imp o emen s we e cen ed on imp o ing he main ainabili y o he u bines by acili a ing access h ough li ing imp o emen s and onsho e a ms and condi ion moni o ing happening discon inuously, wi h measu emen s aken du ing se ice isi s exclusi ely. ( an Bussel and Hende son 2001) Cu en ly, al hough he e olu ion o s a egies o main enances is ongoing, i is clea ly cen ed on emo e condi ion moni o ing o he u bines h ough he applica ion o ad anced models om ib a ion and acous ic signals. ( an Bussel and Hende son 2001) Speci ically ib a ion analysis ep esen s 58% o he ma ke sha e o condi ion moni o ing. (Ba szcz 2019) SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 15 4. Vib a ion analysis As s a ed in he p e ious sec ion, mos cu en condi ion moni o ing sys ems use ib a ion analysis, and he e is g ea in e es in i s applica ion o sma , emo e condi ion moni o ing. This sec ion will gi e an in oduc ion on how ib a ion signals a e acqui ed and p ocessed, and how condi ion moni o ing sys ems use ib a ion da a. 4.1. Da a acquisi ion The i s s ep o ib a ion analysis is he acquisi ion o he ib a ion da a. This is usually done by placing se e al accele ome e s h oughou he machine o a ea o be moni o ed. The exac placemen will depend on he machine, he componen s ha p esen mos wea , and a a ie y o o he ac o s. In wind u bines he senso s a e usually placed on he d i e ain, blades, and suppo s uc u e. 4.2. Vib a ion Signals Al hough he s udy o ib a ion signals may s a wi h simple, clean sine wa es, ib a ion signals eco ded in eal se ings a e o en much noisie , including se e al o e lapping signals o di e en ampli udes and phases. In any case, ib a ion signal analysis equen ly s a s by analysing he signal wa e o m i sel bu o he me hods such as equency analysis o en elope analysis may be used. Due o he scope o his p ojec , only ime domain ib a ion ea u es will be p esen ed. 4.2.1. Rele an ea u es o ib a ion signals The ea u es ha will be p esen ed in his sec ion a e “b oadband” ea u es because hey do no use any il e ing echniques. The e o e, he in o ma ion hey p o ide conside s all signal componen s om a la ge (o “b oad”) equency band and p o ide in o ma ion o he o e all sys em and no jus om he speci ic mechanical p oblem ha may be mal unc ioning. All hese ea u es can be easily calcula ed om he ib a ion signal. They a e: • S a is ical alues such as he mean, he s anda d de ia ion, and he ku osis o he signal. • Roo -mean-squa e (RMS) which desc ibes he a ea o he signal and he e o e i s ene gy 𝑅𝑀𝑆=√𝐸(𝑥2), whe e 𝐸 is he mean alue ope a o . • Peak alue, o peak-peak (PP) is a measu e o he dis ance o he maximum peaks o he signal 𝑃𝑃=𝑥𝑚𝑎𝑥−𝑥𝑚𝑖𝑛 Memo ia 22 Figu e 10 O iginal and ke nelized ea u e space (Rizwan, e al. 2021) One-Class SVM is ained wi h only one class o da a. In his case he samples a e p ojec ed in a highe dimensional space and he hype plane is se be ween he o igin and he samples, making he egion whe e he samples lie as small as possible. Poin s ha lie on he side o he o igin o he hype plane will be conside ed ou lie s. (Scholkop , e al. 1999) Ke nels This is he unc ion ha pe o ms he p ojec ion in o a highe dimensional space. Pa icula ly, his p ojec uses he polynomial, sigmoid and adial basis unc ion (RBF) ke nels. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 23 Figu e 11 Illus a ion o non-linea ke nel ans o ma ions (Ez a Pila io, e al. 2020) Nu Nu is he numbe o samples we allow ou side he decision bounda y du ing he ini ial aining. This hype pa ame e is use ul in he case o a noisy da ase , whe e al hough we expec mos samples o belong o he “no mal” class, we wan o allow some samples o lie ou side o he class. O he wise, he decision bounda y may include ou lie s. 5.3.2. Isola ion o es Isola ion o es is a semi-supe ised model o anomaly de ec ion based on he use o decision ees. Memo ia 24 The model will c ea e decision ees based on andom ea u es o he samples, se ing a andom h eshold o he sepa a ion c i e ia. As anomalies a e " ew and di e en " hey will be sepa a ed ea ly in he ee. Figu e 12 Example o a andom ee in an Isola ion Fo es Model SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 25 Figu e 13 Sca e plo and decision bounda ies o a andom decision ee in an Isola ion T ees model Ins ead o elying on one single decision ee, he model gene a es many isola ion ees, and he anomalies will be hose ha on a e age, o e all o he ees, ha e sho pa hs. The ensemble o hese isola ion ees is wha he name o he algo i hm e e s o. Con amina ion The pe cen age o samples ha a e expec ed o be anomalies. I is he one pa ame e ha makes his model in o a semi-supe ised model, as a leas an es ima ion o he p opo ion o he wo classes mus be known be o ehand. (Liu, Ting and Zhou 2008) Numbe o es ima o s The amoun o decision ees in he andom o es . The numbe o ees will impac he compu a ion ime signi ican ly. Howe e , a la ge numbe o ees will also p oduce mo e nuanced anomaly sco es when compa ed o a lowe numbe o ees. As his is no a de e minis ic model, wi h a lowe numbe o ees he esul s will also be less epea able. This can be sol ed by using a pseudo- andom gene a ion ha can be seeded wi h a epea able andomness s a e. Memo ia 26 5.4. Valida ion me ics in Machine Lea ning This sec ion will p esen some me ics ha can be used o he alida ion o ML models. In eg ession models some me ics like E o , Mean Squa e E o o Roo Mean Squa e E o may be used. As his p ojec ocuses on he sepa a ion o wo classes o da a, wo ele an me ics a e: Accu acy Accu acy is a ela i ely simple me ic o assess he pe o mance o a ML model. I is he pe cen age o co ec labels p edic ed o a da ase . Al hough i is a simple me ic i has some sho alls in he case o unbalanced da a. The model may be classi ying co ec ly only one la ge class and due o class imbalance, a high accu acy could s ill be ob ained. Con usion ma ix Con usion ma ixes a e a common way o ep esen he pe o mance o a ML model. Con usion ma ixes ha e ou boxes, and hey ep esen he p edic ed and ue label o a da ase . A good pe o ming model will pe ec ly map all he samples, so he p edic ed label ma ches he ue label. Con usion ma ixes also p o ide insigh in o alse posi i es and alse nega i es. 5.5. Applica ions o Machine Lea ning in Enginee ing Machine Lea ning has been applied o p oblems in he enginee ing domain o many decades now. Cu en ly, he ad ances in compu ing, senso echnology and new algo i hms a e acili a ing he implemen a ion o Machine Lea ning o new indus y p oblems. 5.5.1. Manu ac u ing Indus y Wi hin he manu ac u ing indus y, Machine Lea ning has been used o p ocess op imisa ion (Weiche , e al. 2019) and p edic i e main enance and compu e ision sys ems ha e been implemen ed o quali y con ol. (Wu and Sun 2013) 5.5.2. Ene gy indus y Wi hin he ene gy indus y, machine lea ning is cu en ly being applied o ene gy demand o ecas ing (Ahma and Chen 2018) and p edic i e main enance o bo h elec ical dis ibu ion (Ho man, e al. 2020) and ene gy p oduc ion asse s such as wind u bines. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 27 6. Vib a ion analysis and ML model applica ion o expe imen al da a In his sec ion an expe imen al da ase will be used o ain wo models (One Class SVM and Isola ion Fo es ) in o de o assess i i is easible o use non-supe ised and semi-supe ised ML models o SHM o o sho e wind u bines. 6.1. Da a collec ion The da ase used is he same as in Vidal e . al. In he a icle, eigh iaxial accele ome e s a e placed on a scaled down model o an o sho e wind u bine wi h a jacke s uc u e. The wind condi ions a e simula ed by a modal shake using se e al ampli udes (0.5, 1, 2 and 3A) o elec ical cu en as a p oxy o wind speeds. Fu he mo e, da a is eco ded in 4 scena ios: a heal hy ba , a ba wi h a loose bol , a ba wi h a c ack, and a eplica ba . (Vidal, Rubias and Pozo 2019) Figu e 14 (a) The bench es de ailing he loca ion o he ba , and(b) Loca ion o he senso s (Hoxha, Vidal and Pozo 2020) Memo ia 28 The da a consis s o 25 expe imen s o each ampli ude, amoun ing o a o al o 100 expe imen s: Ampli ude 0.5 A 1 A 2 A 3 A Heal hy ba 10 10 10 10 Replica ba 5 5 5 5 C acked ba 5 5 5 5 Loose bol in ba 5 5 5 5 Table 1 Numbe o expe imen s by s a e o ba and ampli ude In each expe imen , a ime window o 60 seconds is eco ded a a equency o 1651.6129 Hz. Thus, we ob ain 99097 da a measu emen s om each o he 24 senso s (8 accele ome e s wi h 3 axis each) o each expe imen . 6.2. Da a ans o ma ion As explained in he p e ious sec ion, o each o he 25 expe imen s pe o med we ob ain a ma ix o shape [999097x24]. Howe e , since he sampling equency is e y high compa ed o an indus y se ing, he da a is subsampled in a 1:6 a io. The e o e, ou new subsampled ma ixes a e o shape [166517x24] which is equi alen o a sampling equency o 256 Hz and a ime window o 60 seconds. [𝑥(1,1) ⋯ 𝑥(1,24) ⋮ ⋱ ⋮ 𝑥(999097,1) ⋯ 𝑥(999097,24)]𝑆𝑢𝑏𝑠𝑎𝑚𝑝𝑙𝑖𝑛𝑔 𝑡𝑜 𝑙𝑜𝑤𝑒𝑟 𝑓𝑟𝑒𝑞. → [𝑥(1,1) ⋯ 𝑥(1,24) ⋮ ⋱ ⋮ 𝑥(166517,1) ⋯ 𝑥(166517,24)] Howe e , we can expec o ob ain esul s wi h a sho e ime window, so he da a is eshaped in o de o ob ain 664 samples om each expe imen , which equa es using a ime window o 0.090361 seconds. The e o e, each ow (sample) will con ain 199 imes amps o 24 senso s, o a o al leng h o 4776 da apoin s). We can s ack he samples in a ma ix o shape [664x4776] o each expe imen , and u he mo e s acking samples o se e al expe imen s, al hough each sample will be p ocessed sepa a ely. [𝑥(1,1) ⋯ 𝑥(1,24) ⋮ ⋱ ⋮ 𝑥(166517,1) ⋯ 𝑥(166517,24)]𝑆𝑝𝑙𝑖𝑡𝑡𝑖𝑛𝑔 𝑡ℎ𝑒 𝑚𝑎𝑡𝑟𝑖𝑥 → [[𝑥(1,1) … 𝑥(199,24)] ⋮ [𝑥(166318,1) … 𝑥(166517,24)]] SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 29 The ma ixes a e hen scaled wi h a s anda d scale i ed column wise o he “heal hy” da ase . 6.3. Fea u es Fo each o he samples ob ained in he p e ious sec ion, we calcula e he a e age, s anda d de ia ion, ku osis, RMS, PP, ZP, CF (de ined in Sec ion 4.2.1) and spec al en opy. 6.4. Model aining The heal hy samples a e spli in o a aining se (80%) and a alida ion se (20%). In he case o One Class SVM all o he s a es a e used only as alida ion da a and no used o aining. In he case o Isola ion Fo es , a andom se o 8,73% he size o he heal hy sample aining se is d awn and included in he aining se , and he comple e se o o he s a es is used o aining. Then we ain he model wi h he aining se o a ange o alues on he hype pa ame e s o bo h models. Hype pa ame e Values Ke nel RBF, Polynomic, Sigmoid Nu 0.0001, 0.01, 0.1, 0.25 Tole ance 0.01, 0.001, 0.0001 Gamma Scale, Au oma ic Deg ee (only o polynomic ke nel) 2 Table 2 Hype pa ame e alues es ed o One Class SVM Hype pa ame e Values Numbe o es ima o s 5, 10, 50, 100 Con amina ion N andom ou lie s/N o al aining da a Random s a e 32 Table 3 Hype pa ame e alues es ed o Isola ion Fo es Memo ia 30 6.5. Model alida ion We pe o m alida ion agains he samples o classes 1, 3 and 4. Bo h models ob ain 100% accu acy in se e al cases. We enclose wo pa icula se s o hype pa ame e s and ea u es ha achie ed his accu acy. The ull esul s can be ound in he Gi Hub eposi o y in h ps://gi hub.com/cla a- 9/TFG_public. Pa ame e s Numbe o Es ima o s O e all accu acy S anda d De ia ion, Ku osis, Spec al En opy 50 100% RMS, Ze o Peak, Spec al En opy 100 100% S anda d De ia ion, Ze o Peak, Spec al En opy 100 100% S anda d De ia ion, Peak-Peak, Spec al En opy 100 100% Mean, RMS, Spec al En opy 50 100% Mean, Ku osis, Spec al En opy 100 100% Ku osis, Spec al En opy 50 100% Ku osis, Spec al En opy 100 100% S anda d De ia ion, Ku osis, Spec al En opy 100 100% Mean, Ku osis, Spec al En opy 50 100% Mean, RMS, Spec al En opy 100 100% Table 4 Selec ion o hype pa ame e s and ea u es o Isola ion Fo es models wi h 100% accu acy Pa ame e s Tole ance Nu Gamma O e all accu acy Ze o Peak, Spec al En opy 0.0010 0.0001 au o 100% Ze o Peak, Spec al En opy 0.0100 0.0001 au o 100% Ze o Peak, Spec al En opy 0.0001 0.0001 scale 100% Ze o Peak, Spec al En opy, C es Fac o 0.0010 0.0001 au o 100% Ze o Peak, Spec al En opy 0.0010 0.0001 scale 100% Ze o Peak, Spec al En opy, C es Fac o 0.0100 0.0001 au o 100% Ze o Peak, Spec al En opy 0.0001 0.0001 au o 100% Ze o Peak, Spec al En opy, C es Fac o 0.0001 0.0001 au o 100% Table 5 Selec ion o hype pa ame e s and ea u es o One Class SVM models ke nel RBF wi h 100% accu acy Pa ame e s Tole ance Nu Gamma O e all accu acy Mean, Spec al En opy 0.0001 0.0001 au o 100% RMS, Spec al En opy 0.0001 0.0001 au o 100% Ku osis, Spec al En opy, C es Fac o 0.0001 0.0001 au o 100% Ku osis, Spec al En opy, C es Fac o 0.0001 0.0001 scale 100% S anda d De ia ion, Spec al En opy 0.0001 0.0001 scale 100% Mean, Spec al En opy 0.0001 0.0001 scale 100% S anda d De ia ion, Spec al En opy, 0.0001 0.0001 au o 100% SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 31 C es Fac o RMS, Spec al En opy 0.0001 0.0001 scale 100% Table 6 Selec ion o hype pa ame e s and ea u es o One Class SVM models ke nel sigmoid wi h 100% accu acy Pa ame e s Tole ance Nu Gamma O e all accu acy Mean, Spec al En opy 0.0001 0.0001 scale 100% RMS, Spec al En opy 0.0001 0.0001 scale 100% RMS, Spec al En opy 0.0001 0.0001 au o 100% Mean, Spec al En opy 0.0001 0.0001 au o 100% RMS, Spec al En opy 0.0010 0.0001 scale 100% Mean, Spec al En opy 0.0010 0.0001 au o 100% Mean, Spec al En opy 0.0100 0.0001 au o 100% RMS, Spec al En opy 0.0100 0.0001 scale 100% Table 7 Selec ion o hype pa ame e s and ea u es o One Class SVM models ke nel polynomic wi h 100% accu acy We will u he analyse one o he se s o hype pa ame e s and ea u es ha achie ed a 100% accu acy o each model. Speci ically, o isola ion o es , we will analyse he pai o ea u es ku osis and spec al en opy wi h 100 es ima o s, and o one class SVM ke nel RBF, gamma “scale”, ole ance 0.0001, nu 0.0001 wi h ze o-peak and ku osis Figu e 15 Visualisa ion o Isola ion Model selec ed o u he analysis Memo ia 38 7. En i onmen al impac analysis Cu en ly o sho e wind ene gy gene a ion has a highe en i onmen al impac han onsho e wind ene gy gene a ion. Some es ima es indica e ha he emissions o g eenhouse gases amoun ed o less han 7 g CO2-eq/kWh o onsho e and 11 g CO2-eq/kWh o o sho e. (Bounou, Lau en and Olsen 2016) O sho e wind ene gy can also lead o ma ine habi a loss and ecosys em deg ada ion in a a ie y o ways. (He nandez, Shadman and Maali 2021) Howe e , app op ia e main enance can lead o a smalle en i onmen al oo p in o sys ems. In ac , badly main ained sys ems can lead o a highe ene gy consump ion. (Jasiulewicz-Kaczma ek and D ożyne 2013) In he case o o sho e wind u bines, he en i onmen al impac o an imp o ed main enance s a egy is wo old: - A highe a ailabili y o he wind u bines will lead o a highe gene a ion o enewable ene gy, enabling displacemen ossil uel-based ene gy gene a ion. (Snyde and Kaise 2009) - Imp o ed main enance leads o mo e eliable sys ems. A highe eliabili y would allow o a lowe equency o se icing, which is usually done by boa o some imes helicop e . The educ ion in hese se icing ips would educe ca bon emissions. Fo hese easons he applica ion o s uc u al heal h moni o ing o o sho e wind u bines could o e all lead o a dec ease in g eenhouse gas emissions and ha e a posi i e en i onmen al impac . SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 39 Conclusions The cu en clima e c isis equi es a shi owa ds enewable ene gies. Wind ene gy gene a ion Will play a majo ole. O sho e wind ene gy can p o ide g ea e ou pu due o mo e p edic able wea he condi ions compa ed o onsho e wind ene gy and has one o he lowes li ecycle g eenhouse gas emissions o any sou ce o ene gy. Fo hese easons, he e is inc eased in e es and in es men in o sho e wind u bines. Some o he di icul ies in hei ope a ion and main enance lie in he di icul y o accessing he si e. Al hough emo e moni o ing has become s anda d in he indus y, s uc u al heal h moni o ing and p edic i e main enance s ill p esen s some challenges. Mos p edic i e main enance s a egies in he indus y ely on ib a ion analysis, his wo k in oduces some o he mos s anda d, b oadband ea u es o s udy ib a ions in he indus y, bu i also in oduces some no el ea u es ha ha e shown p omising esul s in he ield o SHM o WT. Speci ically, his p ojec uses spec al en opy as an addi ional ea u e o he classical ea u es. Rega ding machine lea ning, his p ojec ea u es he use o Isola ion Fo es (a semi supe ised me hod) and One Class SVM (a non-supe ised me hod) as a mo e ealis ic app oach o SHM o WT compa ed o supe ised me hods due o he di icul y o labelling da a. The s a egy es ed in his p ojec (a ailable in h ps://gi hub.com/cla a-9/TFG_public) wo ks unde di e en ope a ing and en i onmen al condi ions and p o ides esul s based only on he ou pu ib a ion da a ga he ed by accele ome e senso s. When he s a egy is es ed agains an ini ial da ase o expe imen al da a accu acies o 100% a e achie ed wi h bo h models. Howe e , when he accu acies o a di e en , p e iously unseen heal hy da ase ob ains lowe accu acies and he ea u es and hype pa ame e s o he model mus be e uned. Excluding he ini ial heal hy s a e achie es o e all accu acies o 93,88% o Isola ion Fo es and 88,67% o One Class Suppo Vec o . Fo all o his, he use o non-supe ised and semi-supe ised machine lea ning models is ealis ic app oach o s uc u al heal h moni o ing o o sho e wind u bines and has ob ained good esul s when es ed agains an expe imen al da ase based on a scaled model. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 41 Economic analysis The esou ces equi ed o de elop his p ojec a e as ollows: Time Sala y Cos Resea che hou s 24 ECTS a 60h/ECTS 760€/80 h mon hly 12680€ Supe iso hou s 10% o esea che hou s 3040€/160 h mon hly 2736€ Compu e esou ces - - 700€ To al cos 16116€ Table 13 Economic analysis o he p ojec This calcula ion conside s he ne sala y o a20h/week esea che posi ion a UPC, mul iplied by 1.3 o ake in o accoun axes. In he case o he supe iso , he sala y has been calcula ed by doubling he hou ly a e. The expe imen al da a used o he de elopmen o he model was gene a ed in a p e ious s udy and will be eleased in an open-sou ce jou nal. All ci ed a icles we e accessed h ough open-sou ce jou nals o access was p o ided by he uni e si y. These cos s ha e no been conside ed. The economic impac o he esea ch goes beyond he di ec cos o he p ojec . Ope a ion and main enance cos s ep esen 25% o ene gy p oduc ion cos s o sho e wind u bine main enance. ( an Bussel and Hende son 2001) This is 15% han Ope a ion and Main enance cos s o onsho e wind u bines. Mo eo e , Tu nbull epo s ha up o 8% o hese cos s can be sa ed h ough ea ly main enance in e en ion. (Tu nbull and Ca ol 2021) SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 43 Bibliog aphy Agus yn, Dawid, Ma in B. Nielsen, and Ronnie R. Pede sen. “Desing o o sho e wind u bine jacke ounda ions.” Eu os eel, 2017. Ahma , Tan ee , and Huanxin Chen. “U ili y companies s a egy o sho - e m ene gy demand o ecas ing using machine lea ning based models.” Sus ainable Ci ies and Socie y, 2018: 401- 417. Ba szcz, Thomas. Vib a ion-Based Condi ion Moni o ing o Wind Tu bines. K akow: Sp inge , 2019. Ba helmie, Rebecca J., and Sa ah C. P yo . “Clima e Change Mi iga ion Po en ial o Wind Ene gy.” Clima e, 2021. Bha acha ya, Buhamoy. Design o Founda ions o O sho e Wind Tu bines. Wiley, 2019. Bounou, Alexand a, Alexis Lau en , and S ig I. Olsen. “Li e cycle assessmen o onsho e and o sho e wind ene gy- om heo y o applica ion.” Applied Ene gy, 2016. Cos a, Ángel M., José A. O osa, Diego Ve ga a, and Pablo Fe nández A ias. “New Tendencies in Wind Ene gy Ope a ion and Main enance.” Applied Sciences 11, no. 4 (2021). Du uaa, Salih O., and A. Raou . Planning and con ol o main enance sys ems. 2015. Eu opean En i onmen Agency. “Eu ope's onsho e and o sho e wind ene gy po en ial:An assessmen o en i onmen al and economic cons ain s.” 2009. Ez a Pila io, Ka l, Mahmood Sah iee, Yi Cao, Yao Liyun, and Yang Shuang-Hua. “A Re iew o Ke nel Me hods o Fea u e Ex ac ion in Nonlinea P ocess Moni o ing.” P ocesses, 2020. Fauls ish, S., Be hold Hahn, and P.J. Ta ne . “Wind Tu bine down ime and i s impo ance o o sho e deploymen .” Wind Ene gy, 2011. Global Wind Ene gy Council. “Global Wind Repo 2021.” 2021. G us, Joel. Da a Science om Sc a ch. O'Rielly, 2015. Haenlein, Michael, and And eas Kaplan. “A B ie His o y o A i icial In elligence: On he Pas , P esen , and Fu u e o A i icial In elligence.” Cali o nia Managemen Re iew, 2019. Memo ia 44 Hä ele, Jan, Rick R. Dammiani, Ryan N. King, C is ian G. Gebha d , and Raimund Rol es. “A sys ema ic app oach o o sho e wind u bine jacke p edesign and op imiza ion: geome y, cos , and su oga e s uc u al code check models.” Wind Ene gy Science 3 (2018): 553–572. He nandez, Mau icio, Milad Shadman, and Moj aba Maali. “En i onmen al impac s o o sho e wind ins alla ion, ope a ion and main enance, and decommissioning ac i i ies: A case s udy o B azil.” Renewable and Sus ainable Ene gy Re iews, 2021. Higgins, Lindley R., R. Kei h Mobley, and Smi hm Ricky. MAin enance Enginee ing Handbook. McG aw-Hill, 2002. Ho man, Ma in W., S ephan Wilde mu h, Ralph Gi zel, and Aydin Boyaci. “In eg a ion o No el Senso s and Machine Lea ning o P edic i e Main enance in Medium Vol age Swi chgea o Enable he Ene gy and Mobili y Re olu ions.” Senso s, 2020. Hoxha, E in, Yolanda Vidal, and F ancesc Pozo. “Damage Diagnosis o O sho e Wind Tu bine Founda ions Based on he F ac al Dimension.” Applied Sciences, 2020. IEA. Wind elec ici y gene a ion, Wo ld 1990-2019. 2019. h ps://www.iea.o g/da a-and- s a is ics/da a- b owse ?coun y=WORLD& uel=Renewables%20and%20was e&indica o =WindGen (accessed 10 17, 2021). —. “Wo ld Ene gy Ou look.” 2018. Jasiulewicz-Kaczma ek, Małgo za a, and P zemysław D ożyne . “The Role o Main enance in Reducing he Nega i e Impac o a Business on he En i onmen .” In Sus ainabili y App aisal: Quan i a i e Me hods and Ma hema ical Techniques o En i onmen al Pe o mance E alua ion, by Ma ina G E ech chouko a, Pe e A Khai e and Paulina Golinska, 141-166. Sp inge , 2013. Li, Jiawen, Zhenni Li, Yichen Jiang, and Tang Yougang. “Typhoon Resis ance Analysis o O sho e Wind Tu bines: A Re iew.” A hmosphe e, 2022. Liu, Fei Tony, Kai Ming Ting, and Zhi-Hua Zhou. “Isola ion Fo es .” 2008 Eigh h IEEE In e na ional Con e ence on Da a Mining. 2008. Manzano-Aguglia o, F ancisco, Miguel Sánchez-Cale o, Al edo Alcayde, Ca los San-An onio-Gómez, and Albe o Jesús, Same ón-Manzano, Es e Pe ea-Mo eno. “Wind Tu bines O sho e Founda ions and Connec ions o G id.” In en ions, 2020. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 45 Michal, Wasilczuk¸, Ra al Gawa kiewicz, and Filip Wasilczuk. “D i e ain o a Wind Tu bine.” Machine Dynamics Resea ch, 2015. No ig, Pe e , and S ua Russell. A i icial In elligence: A Mode n App oach. Pea son, 2021. Pu uncajas, B yan, Yolanda Vidal, and Ch i ian Tu i én. “Vib a ion-Response-Only S uc u al Heal h Moni o ing o O sho e Wind Tu bine Jacke Founda ions ia Con olu ional Neu al Ne wo ks.” Senso s, 2020. PwC. “P edic i e Main enance 4.0.” 2018. Rizwan, A i , Naeem Iqbal, Rashid Ahmad, and Kim Do-Hyeum. “WR-SVM Model Based on he Ma gin Radius App oach o Sol ing he Minimum Enclosing Ball P oblem in Suppo Vec o Machine Classi ica ion.” Applied Sciences, 2021. Sando al, Diego, U ko Le iu ondo, Pozo F nacesc, and Yolanda Vidal. “Low-Speed Bea ing Faul Diagnosis Based on Pe mu a ion and Spec al En opy Measu es.” Applied Sciences, 2020. Schallehn, Michael, Ch is ophe Scho ling, Pe e Bowen, and Oli e S aehle. Beyond P oo s o Concep : Scaling he Indus ial IoT. Bain & Company, 2018. Schlöme S., T. B uckne , L. Ful on, E. He wich, A. McKinnon, D. Pe czyk, J. Roy, R. Schae e , R. Sims, P. Smi h, and R. Wise. “Technology-speci ic cos and pe o mance pa ame e .” Chap. Annex III in Clima e Change, by T. B uckne , L. Ful on, E. He wich, A. McKinnon, D. Pe czyk, J. Roy, R. Schae e , R. Sims, P. Smi h, and R. Wise Schlöme S. Camb idge Uni e si y P ess, 2014. Scholkop , Be nha d, Robe Williamson, Alex Smola, Alex Shawe-Taylo , and John Pla . “Suppo Vec o Me hod o No el y De ec ion.” Neu ips, 1999: 582-588. Shoja, Sla ash. Guided wa e-based app oach o heal h moni o ing o composi e s uc u es, Applica ion o wind u bine blades. Depa men o Mechanics and Ma i ime Sciences, Chalme s Uni e si y o Technology, 2018. Snyde , B yan, and Ma k J. Kaise . “Ecological and economic cos -bene i analysis o o sho e wind ene gy.” Renewable Ene gy, 2009. S a co, Ad ian, Fa eme Dinmohammadi, Xingyu Zhao, and Valen in Robu. “Machine lea ning me hods o wind u bine condi ion moni o ing: A e iew.” Renewable Ene gy, 2019: 620-635. Taylo , Regan, Ch is ophe Beale, and Inalpola Mu a . “Wind Tu bine Blade Damage De ec ion Using Supe ised Machine Lea ning Algo i hms.” Jou nal o Vib a ion and Acous incs (ASME), 2017. Memo ia 46 Tu nbull, Alan, and James Ca ol. “Cos Bene i o Implemen ing Ad anced Moni o ing and P edic i e Main enance S a egies o O sho e Wind Fa ms.” Ene gies, 2021. an Bussel, G.J.W., and A. R. Hende son. “S a e o he a and Technology T ends o O sho e Wind Ene gy: Ope a ion and Main enance Issues.” O sho e Wind Ene gy, 2001. Vidal, Yolanda, José L. Rubias, and F ancesc Pozo. “Wind Tu bine Heal h Moni o ing Based on Accele ome e Da a.” IX ECCOMAS Thema ic Con e ence on Sma S uc u es and Ma e ials. 2019. Wang, Dahai, Jung Peng, Quian Yu, Yuanyuan Chen, and Hanghang Yu. “Suppo Vec o Machine Algo i hm o Au oma ically Iden i ying Deposi ional Mic o acies Using Well Logs.” Sus ainabili y, 2019. Weiche , Do ina, Pa ick Link, Anke S oll, S e an Rüping, S e en Ihlend eld , and S e an W obel. “A e iew o machine lea ning o he op imiza ion o p oduc ion p ocesses.” The In e na ional Jou nal o Ad anced Manu ac u ing Technology 104 (2019): 1889–1902. Wo ld Resou ces Ins i u e. “Clima e Wa ch His o ical GHG Emissions.” Washing o D.C., 2021. Wu, Di, and Da-Wen Sun. “Colou measu emen s by compu e ision o ood quali y con ol – A e iew.” T ends in Food Science & Technology 21 (2013): 5-20. Xie, Mian, and Susana Lopez-Que ol. “Nume ical Simula ions o he Mono onic and Cyclic Beha iou o O sho e Wind Tu bine Monopile Founda ions in Clayey Soils.” Jou nal o Ma ine Science and Enginee ing, 2021. SHM o O sho e Wind Tu bine Founda ions Th ough Unsupe ised and Semi Supe ised Machine Lea ning Me hods 47