| Abstract [eng] |
The aim of this work is to design and investigate a 3D printing defect image recognition system that would be able to adjust the 3D printer settings depending on the defect found. To achieve the goal it is necessary to: analyse 3D printing and defect detection methods, define the main types of 3D printing defects and video camera calibration methods, analyse the methods of artificial neural network training and its application in 3D printing, develop a control algorithm, select hardware and software for 3D printing defective image recognition system. The first chapter of this project describes 3D printing methods: fused deposition modeling, stereolithography and laser powder bed fusion, and explains the principle of operation of each method. The second chapter analyzes the most common 3D printing defects, from which 5 main ones were selected: stringing, object deformation, over and under extrusion, layer separation (cracking). The third section examines defect detection methods using multi-camera image recognition. In the fourth chapter, the neural network training methods are analyzed, from which supervised learning is selected for the application used. The fifth chapter describes the use of a neural network in 3D printing to monitor procedures, design, and apply correlations between process parameters and the final properties of the resulting component. In the sixth chapter, the methods of camera calibration are discussed: automatic calibration, non-planar calibration and using a checkerboard. The seventh chapter describes and selects the necessary equipment to create a 3D printing defect image recognition system. The following defect monitoring equipment components have been selected: “Raspberry Pi” 4b microcomputer, which stores photos and supports communication with a 3D printer, “Raspberry Noir” V2 8MP camera suitable for use with a microcomputer. Selected software: “Raspbian Debian” 10, 32-bit operating system, “Tensorflow” 2.0.0. In the eighth chapter, tests with 3D printed objects of various dimensions were performed, the decisions made were analyzed, and the parameter control algorithm for making the right decisions was adjusted. The system is trained to detect 5 different types of defects. During the tests, parameter sets were created that replicate the defect of over or under extrusion. Algorithm optimization was performed to identify each defect separately. An algorithm for photo collection and defect inspection has been developed, which is used to check communication with the 3D printer, process photos and make decisions to correct defects. Photos are processed and tagged at a rate of 2 frames per second at a photo resolution of 1280 × 720 pixels. When using colorful filament, the subject properties and defects are found to be most visible when using green lighting. |
