20.500.12556/DKUM-67936 Napovedovanje odpovedi izdelkov z metodami globokega učenja Predicting the failures of products using deep learning methods Diplomsko delo obravnava razvoj in optimizacijo modelov za analizo garancijskih podatkov in napovedovanje odpovedi z metodami globokega učenja. Globoko učenje je redko uporabljeno v tovrstne namene, zato so raziskave na tem področju pomembne, a obenem težavne, saj obstaja manj predhodnih virov, s katerimi si lahko pomagamo. Na drugi strani pa se tehnologija v zadnjih letih razvija izjemno hitro, tako da lahko modele globokega učenja implementiramo tudi brez detajlnega poznavanja vseh elementov globokega učenja, kar je omogočilo razcvet uporabe in aplikacijo globokega učenja na široko paleto problemov. V nalogi smo preizkusili več različnih modelov, od prilagojenega enoslojnega perceptrona do konvolucijske nevronske mreže, in večje število optimizacijskih metod. Z uporabljenimi metodami smo dosegli 30–40-% stopnjo natančnosti, kar odstopa od želene 10-% stopnje napake. Pri tem moramo upoštevati majhen nabor vhodnih podatkov. Metode globokega učenja se ob zastavljenem zahtevnem pogoju niso izkazale kot primerne za uporabo, iz pridobljenih informacij pa zaključujemo, da bodo metode najverjetneje uporabne v prihodnje, ko bo na voljo več podatkov, ki bodo tudi bolj kvalitetni. The thesis deals with the development and optimization of models for analyzing warranty data and predicting failure with deep learning methods. Deep learning is rarely used for such purposes, which is why research in this field is important but difficult, given that there are fewer sources we can rely on for previous experience. On the other hand, technology has been developing rapidly in recent years, so we are able to implement deep learning even without detailed knowledge of all the elements of the technology. This enabled the increase in use and application of deep learning to a wide range of problems. In the thesis, we tested several different models from the adapted single-layer perceptron to the convolutional neural network and a number of optimization methods. With the methods used, we achieved a 30-40% accuracy rate, which deviates from the desired error rate of 10%. However, we need to take into consideration a small set of input data we had available. The methods of deep learning did not prove to be suitable for use in this specific problem set, but from the acquired information we can conclude that the methods will most likely be useful in the future, when more data of higher quality is available. Garancijski podatki Strojno učenje Nevronske mreže Globoko učenje Python Tensorflow Warranty data Machine learning Neural networks Deep learning Python Tensorflow true true false Slovenski jezik Angleški jezik Diplomsko delo/naloga 2017-09-04 11:48:12 2017-09-14 12:52:26 2022-08-01 14:38:28 2019-07-11 16:36:07 2017 0 Kranj 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 7962387 URN:SI:UM:DK:2RKNENYB UN_Sasek_Blaz_2017.pdf UN_Sasek_Blaz_2017.pdf 1 7089AB4CEB90D87B0CA1784AB0459EA8 d005cf79d685a2e126b98d6c1aac9838250a09ba3126b4c12664436f2dc07fc9 a0e79268-7c0e-11eb-bb7a-00155d0001ca 20.500.12556/dkum/dc51ca82-4c01-49ee-b78d-bc21922a1a59 https://dk.um.si/Dokument.php?lang=slv&id=116958 Fakulteta za organizacijske vede 0 0 0