Π sikhulume ngokubikezela kochungechunge lwesikhathi. Ukuqhubeka okunengqondo kungaba i-athikili ekuboneni okudidayo.
Isicelo
Ukutholwa okudidayo kusetshenziswa ezindaweni ezifana nalezi:
1) Ukubikezelwa kokuwohloka kwemishini
Ngakho-ke, ngo-2010, ama-centrifuges ase-Iranian ahlaselwa yigciwane le-Stuxnet, elabeka imishini ekusebenzeni okungalungile futhi lakhubaza ezinye zemishini ngenxa yokuguga okusheshayo.
Ukube bekusetshenziswe ama-algorithms okuthola okudidayo emishinini, isimo sokwehluleka besingeke sigwenywe.
Ukusesha okungahambi kahle ekusebenzeni kwemishini akusetshenziswa embonini yenuzi kuphela, kodwa nasekusebenzeni kwensimbi nokusebenza kwezinjini zezindiza. Futhi kwezinye izindawo lapho ukusetshenziswa kokuxilongwa okubikezelwayo ishibhile kunokulahlekelwa okungenzeka ngenxa yokuwohloka okungalindelekile.
2) Ukubikezela ukukhwabanisa
Uma imali ikhishiwe ekhadini olisebenzisa e-Podolsk e-Albania, ukuthengiselana kungase kudingeke kuphinde kuhlolwe.
3) Ukukhonjwa kwamaphethini omthengi angajwayelekile
Uma amanye amakhasimende ebonisa ukuziphatha okungavamile, kungase kube nenkinga ongayazi.
4) Ukuhlonza isidingo esingavamile nomthwalo
Uma ukuthengisa esitolo se-FMCG kwehle ngaphansi kwesikhathi sokuzethemba sesibikezelo, kufanelekile ukuthola isizathu salokho okwenzekayo.
Izindlela zokukhomba okudidayo
1) Sekela umshini weVector oneClass One-Class SVM
Ifanele uma idatha ekusethi yokuqeqesha ilandela ukusatshalaliswa okuvamile, kodwa isethi yokuhlola iqukethe okudidayo.
Umshini we-vector wesigaba esisodwa wakha indawo engaqondile eduze komsuka. Kungenzeka ukusetha umkhawulo wokunqamuka lapho idatha ibhekwa njengeyindida.
Ngokusekelwe kokuhlangenwe nakho kwethimba lethu le-DATA4, i-One-Class SVM iyi-algorithm esetshenziswa kakhulu ukuxazulula inkinga yokuthola okudidayo.
2) Hlukanisa indlela yehlathi
Ngendlela "engahleliwe" yokwakha izihlahla, ukukhishwa kuzongena emaqabunga ngezigaba zakuqala (ekujuleni okungajulile kwesihlahla), i.e. ukukhishwa kwekhabhoni kulula "ukukuhlukanisa." Ukuhlukaniswa kwamanani amangalisayo kwenzeka ekuphindaphindweni kokuqala kwe-algorithm.
3) Imvilophu ye-elliptic nezindlela zezibalo
Isetshenziswa lapho idatha isatshalaliswa ngokuvamile. Lapho isilinganiso sisondela emsileni wengxube yokusabalalisa, inani elixakayo kakhulu.
Ezinye izindlela zezibalo nazo zingafakwa kuleli klasi.
Isithombe esivela ku-dyakonov.org
4) Izindlela zemethrikhi
Izindlela zifaka phakathi ama-algorithms afana nomakhelwane abaseduze no-k, umakhelwane oseduze no-k, i-ABOD (ukutholwa kokuphuma kwe-engeli) noma i-LOF (isici sasendaweni sangaphandle).
Kuyafaneleka uma ibanga eliphakathi kwamanani ezicini lilingana noma lijwayelekile (ukuze lingalinganisi i-boa constrictor kumapholi).
I-algorithm yomakhelwane abaseduze kakhulu ithatha ukuthi amanani ajwayelekile atholakala endaweni ethile yesikhala se-multidimensional, futhi ibanga lokuya kokudidayo lizoba likhulu kune-hyperplane ehlukanisayo.
5) Izindlela ze-Cluster
Ingqikithi yezindlela zeqoqo iwukuthi uma inani lingaphezu kwenani elithile kude nezikhungo zeqoqo, inani lingabhekwa njengeliyindida.
Into esemqoka ukusebenzisa i-algorithm ehlanganisa kahle idatha, okuncike emsebenzini othize.
6) Indlela yengxenye eyinhloko
Ifanele lapho izikhombisi-ndlela zoshintsho olukhulu kakhulu ekuhlakazweni zigqanyiswa khona.
7) Ama-algorithms asuselwa ekubikezelweni kochungechunge lwesikhathi
Umqondo uwukuthi uma inani liwela ngaphandle kwesikhawu sokuqiniseka sokubikezela, inani lithathwa njengeliyinqaba. Ukubikezela uchungechunge lwesikhathi, ama-algorithms afana nokushelela kathathu, i-S(ARIMA), i-boosting, njll. iyasetshenziswa.
Ama-algorithms okubikezela ochungechunge lwesikhathi kuxoxwe ngawo esihlokweni esandulele.
8) Ukufunda okugadiwe (ukuhlehla, ukuhlukaniswa)
Uma idatha ivumela, sisebenzisa ama-algorithms asuka ekuhlehleni komugqa kuye kumanethiwekhi avamile. Ake silinganise umehluko phakathi kokubikezela kanye nenani langempela, futhi sifinyelele isiphetho ngokuthi idatha ichezuka kangakanani kulokho okujwayelekile. Kubalulekile ukuthi i-algorithm ibe nekhono elanele lokuhlanganisa nokuthi isethi yokuqeqesha ayiqukethe amanani angavamile.
9) Izivivinyo zamamodeli
Ake sibhekane nenkinga yokusesha okudidayo njengenkinga yokufuna izincomo. Masibolise isici sethu se-matrix sisebenzisa i-SVD noma imishini ye-factorization, futhi sithathe amanani ku-matrix entsha ahluke kakhulu kunawasekuqaleni njengamangalisa.
Isithombe esivela ku-dyakonov.org
isiphetho
Kulesi sihloko, sibuyekeze izindlela eziyinhloko zokuthola i-anomaly.
Ukuthola okungaqondakali kungabizwa ngezindlela eziningi ngokuthi ubuciko. Ayikho i-algorithm ekahle noma indlela, ukusetshenziswa kwayo okuxazulula zonke izinkinga. Ngokuvamile isethi yezindlela isetshenziswa ukuxazulula icala elithile. Ukutholwa kwe-anomaly kwenziwa kusetshenziswa imishini ye-vector yokusekelwa yesigaba esisodwa, ukuhlukanisa amahlathi, izindlela zemethrikhi neqoqo, kanye nokusebenzisa izingxenye eziyinhloko nokubikezela kochungechunge lwesikhathi.
Uma wazi ezinye izindlela, bhala ngazo kumazwana esihlokweni.
Source: www.habr.com