3788 KB225 KB0 KB2021Brili, NikaFicko, MirkoKlančnik, SimonXII, 114 str., 30 cmCOBISSID:86333443URN:URN:NBN:SI:doc-YRY5TZCOslN. Brilivisokošolska delaartificial intelligencecuttingcutting tooldeep learningDisertacijedoktorske disertacijegloboko učenjeIndustrija 4.0industry 4.0Obrabaobraba orodjaodrezavanjeOrodjarezalna orodjastruženjetermografijathermographytool wearturningumetna inteligencaModel inteligentnega nadzora obrabe in poškodb rezalnih orodij z uporabo termografije| doktorska disertacija|In turning, wear control of a cutting tool serves to increase product quality, optimise tool-related costs, and avoid undesirable events. In small series and single item production, the machine operator decides when a cutting tool needs to be changed based on his experience. Incorrect decisions often lead to higher costs, production downtime and scrap. In this dissertation, a control system for a tool condition monitoring system that automatically detects tool wear during turning is presented. For process control an infrared camera was used, which - in contrast to conventional cameras - detects not only the visual but also the thermographic condition. Despite difficult environmental conditions (e.g. hot chips), we protected the camera and placed it right up to the cutting knife, so that the machining could be observed closely. To create a dataset of 18,486 images, we machined on a lathe with tool inserts of different wear levels. Using a Deep Learning and a Convolutional Neural Network (CNN), we developed a model for tool wear and tool damage prediction. The image database was divided into two groups: images created during the turning process and images created after the turning process (thermographic images of a cutting tool). The model is successful on both image databases. Based on the thermographic images, the learned model automatically determines the condition of the cutting tool in terms of its suitability for further use in machining (no wear, low wear, high wear). The accuracy of classification for the combined set of all images is 99.92%, which confirms the suitability of the proposed method. The model was tested on unknown images (changed machining conditions), confirming the robustness of the intelligent system for the classification of unknown materials (more than 98% accuracy of tool wear classification for unknown material). Such a system allows immediate action in case of cutting tool wear or breakage, regardless of the operator's knowledge and competence.%Nadzor obrabe rezalnega orodja pri struženju prispeva k izboljšanju kakovosti izdelkov, optimizaciji stroškov orodja in zmanjšanju števila neželenih dogodkov. Pri maloserijski in posamični proizvodnji se operater stroja na podlagi izkušenj odloča, kdaj zamenjati rezalno orodje. Slabe odločitve lahko vodijo do povišanja stroškov, zastojev proizvodnje in izmeta. V disertaciji smo predstavili sistem nadzora stanja rezalnega orodja, ki med in po struženju samodejno prepozna obrabo rezalnega orodja. Za nadzor procesa smo uporabili infrardečo (IR) kamero, ki za razliko od uporabe navadnih industrijskih kamer ne spremlja zgolj vizualnega stanja procesa, ampak zajema še termografsko stanje. Kljub zahtevnemu okolju (vroči ostružki) smo kamero ustrezno zaščitili in namestili tik ob rezalno ploščico, kar omogoča spremljanje obdelave iz neposredne bližine. Material smo obdelovali z različno obrabljenimi rezalnimi ploščicami in ustvarili bazo 18.486 slik, ki so bile namenjene učenju in testiranju modela. Z uporabo globokega učenja in konvolucijske nevronske mreže (CNN) smo razvili napovedni model obrabe in poškodb rezalnega orodja. Pripravljeno bazo slik smo razdelili na slike, ki so nastale med procesom struženja (slike procesa), in na slike, ki so nastale po struženju (termografske slike rezalnega orodja). Ugotovili smo, da je model uspešen na obeh bazah slik. Naučen model na podlagi termografske slike procesa samodejno razvrsti stanje rezalnega orodja glede na primernost za nadaljnjo uporabo pri struženju (brez obrabe, majhna obraba, velika obraba). Točnost klasifikacije za združeno množico vseh slik je 99,92 % in potrjuje ustreznost predlagane metode. Model smo testirali tudi na nepoznanih slikah (spremenjeni obdelovalni pogoji), s čimer smo z več kot 98 % točnostjo klasifikacije potrdili robustnost naučenega sistema pri uporabi za nepoznani material obdelovanca. Takšen sistem omogoča takojšnje ukrepanje v primeru obrabe ali zloma rezalnega orodja, ne glede na znanje in usposobljenost operaterjaTEXTvisokošolska delatheses and dissertationsSlovenian National E-content AggregatorNational and University Library of SloveniaUniverza v Mariboru, Fakulteta za strojništvo