Krmiljenje moči laserskega varjenja na osnovi konvolucijske nevronske mreže

magistrsko delo magistrskega študijskega programa II. stopnje Strojništvo

Alex Božič (Author), Matija Jezeršek (Mentor), Aleksander Sadikov (Co-mentor)

Abstract

V magistrskem delu razvijemo krmilnik moči varilnega laserja z maksimalno močjo 400 W na osnovi slikovnih zaznaval, kjer proces varjenja sprotno ocenjujemo s pomočjo konvolucijske nevronske mreže. Na podlagi naučenih modelov s skupno točnostjo 94% na testnih podatkih smo sposobni oceniti vnos energije v pločevino iz nerjavečega jekla AISI 304 s skupno debelino prekrivnih pločevin 1,5 mm. Krmilnik je s pomočjo CNN-modela in PID-krmilnika zmožen krmiliti moč laserja, pri čemer je vselej poudarek na odzivnosti in stabilnosti sistema. Krmilnik se pri določenih parametrih stabilizira že v 0,46 s. S pomočjo dodatnih testov nakažemo smer nadaljnjega dela za izboljšavo in pohitritev krmilnika.

Keywords

magistrske naloge;krmilniki;nevronske mreže;laserska izhodna moč;kamere;lasersko varjenje;

Data

Language:	Slovenian
Year of publishing:	2019
Typology:	2.09 - Master's Thesis
Organization:	UL FS - Faculty of Mechanical Engineering
Publisher:	[A. Božič]
UDC:	621.791.725:004.8:681.5(043.2)
COBISS:	16956443
Views:	909
Downloads:	192
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Laser welding power control based on convolutional neural networks
Secondary abstract:	In the master thesis we develop and implement power controller of welding laser with maximum power of 400 W based on image sensor, where the process of welding is being continuously estimated with convolutional neural network. Based on learnt models with total accuracy of 94 % on test dataset we are able to estimate heat input in overlaying AISI 304 metal sheet with cumulative thickness of 1,5 mm. Controller developed from PID controller and convolutional neural network is responsively and stably controlling laser power output. Controller in certain cases stabilizes within 0,46 s. Based on additional tests we propose additional possible improvements to increase controller's performance.
Secondary keywords:	controllers;neural networks;laser output power;cameras;laser welding;
Type (COBISS):	Master's thesis/paper
Study programme:	0
Embargo end date (OpenAIRE):	1970-01-01
Thesis comment:	Univ. Ljubljana, Fak. za strojništvo
Pages:	XXIII, 72 str.
ID:	11270680