http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
- 中文 을 입력하시려면 zhongwen을 입력하시고 space를누르시면됩니다.
- 北京 을 입력하시려면 beijing을 입력하시고 space를 누르시면 됩니다.
RISS 인기검색어
- 검색키워드
- 저자 : Hyeongwon Kim (검색결과 1 건)
-
Deep-Learning-Based Real-Time Monitoring of Laser Keyhole Welding Processes
Hyeongwon Kim Ulsan National Institute of Science and Technology 2021 국내박사
Recently, laser welding has been widely used in metal joining fields. Since the laser beam has a high energy density, the laser welding has high efficiency, small heat affected zone (HAZ), and good qualities of joining. In laser keyhole welding, a keyhole is formed by the rapid melting and evaporation of substrates. The interaction between the laser beam and keyhole exhibits inherently unstable behavior, and the variations in laser-beam absorptance are also highly unstable. For this reason, monitoring of the keyhole geometry and laser-beam absorptance inside the keyhole is significant in understanding the entire laser keyhole welding process. In this dissertation, a real-time monitoring system combined with the state-of-the-art (SOTA) deep learning models to accurately and promptly analyzed the entire lap joint laser welding process in iron (cold rolled steel and in this study) and nonferrous metal (aluminum alloy in this study) was proposed. In addition, high-speed object deep learning framework that detects defects on the weld bead which were frequently occurred during the laser keyhole welding of foil were developed. In chapter 1, the background of laser keyhole welding and the introduction of deep learning were presented. Analysis methods for laser keyhole welding introduced the prediction of laser-beam absorptance inside the keyhole and the observation of keyhole dynamics. The image classification and object-detection deep learning frameworks, which are the most commonly used methods in computer vision were described. In chapter 2, a data-based deep-learning model for predicting laser-beam absorptance in full-penetration laser keyhole welding was firstly proposed. The training dataset was prepared with laser-beam absorptances computed from various keyhole geometries by the ray-tracing model. A modified ResNet, which is based on a convolutional neural network (CNN) for image regression (predicting scalar values from input images), was employed to predict the laser-beam absorptance from the combined keyhole top and bottom aperture images. The trained model is applicable to real-time monitoring system since it can calculate the laser-beam absorptance inside the keyhole with extremely high accuracy and speed. In chapter 3, a real-time full-penetration laser keyhole welding monitoring system using a synchronized high-speed coaxial observation method and the state-of-the-art deep learning models was established. To observe the rapid variations in unstable keyhole dynamics, the high-speed observation method is essential, but analyzing large amounts of data obtained from the observations is still trouble in perspective of speed and accuracy. The proposed system consisted of a synchronized high-speed coaxial observation method for simultaneously pertaining the top and bottom surfaces of the welding process without any distortion and two deep learning models. In the deep learning models, an object detection deep learning model (YOLOv4) for automatically measuring the keyhole top and bottom apertures and a ResNet with image regression for predicting laser-beam absorptance inside the keyhole observed from the YOLOv4 model were employed. Through this system, it was possible to accurately figure out the tendency of variations in welding process under different experimental conditions (welding speed and laser power) and the weld defect generation process. In chapter 4, study of keyhole behavior and weld defect generation in full-penetration laser keyhole welding of aluminum (Al) alloy using the proposed deep learning frameworks was conducted. Al alloy laser welding has been a challenging task because Al alloy has a low absorptance to the laser beam and the alloying elements have a negative effect on the welding process. Using the developed real-time monitoring system presented in chapter 3, the Al alloy keyhole welding processes according to the experimental conditions and shielding gas (Ar and He) in terms of keyhole dynamics and energy absorptance inside the keyhole were analyzed. In addition, the weld defects (porosity and burn-through) generation processes and factors could be analyzed and comprehended. In chapter 5, the process quality of thin foil keyhole-mode welding was discussed and a deep learning model for detecting defects on the beads in laser keyhole welding of a 100 μm-thick stainless-steel foil onto a 0.5 mm-thick stainless-steel sheet was developed. Since thin metal foils are susceptible to deformation and defects, the joining of metal foils has always been a challenge task. Furthermore, if the penetration depth was deeper than the beam size, keyhole-mode welding was forced and then the process became more unstable. In this dissertation, keyhole-mode welding was found to be sensitive to the shield gas flow rate due to the tiny melt pool and keyhole, and a deep-learning-based model that detects defects on the beads generated during foil welding with high accuracy and speed was developed.
연관 검색어 추천
이 검색어로 많이 본 자료
활용도 높은 자료
