h1

%0 Thesis
%A Knaak, Christian
%T Echtzeitüberwachung und -optimierung der Nahtqualität beim Laserstrahlschweißen mittels bildgebender Sensorik und künstlicher Intelligenz; 1. Auflage
%I RWTH Aachen University
%V Dissertation
%C Aachen
%M RWTH-2025-04385
%@ 978-3-98555-277-1
%B Ergebnisse aus der Lasertechnik
%P 1 Online-Ressource : Illustrationen
%D 2025
%Z Druckausgabe: 2025. - Auch veröffentlicht auf dem Publikationsserver der RWTH Aachen University
%Z Dissertation, RWTH Aachen University, 2025
%X Laser-based manufacturing technology is an indispensable tool for the cost-efficient production of products, such as battery electric vehicles, which are of utmost importance for meeting current societal challenges. However, decreasing cycle times, growing demands on product quality, and increasing flexibility requirements as well as desired increases in cost efficiency represent growing challenges from the perspective of manufacturing technology. In addition to process-specific possibilities for improvement within the framework of the respective applications, the digitalization and automation of manufacturing processes in particular offer the potential to further facilitate market access with regard to resource-saving end products. In order to tap this potential, it is necessary to collect and intelligently process extensive machine, process and production data so that data-supported recommendations for action can be derived. In the context of this work, the holistic evaluation and optimization of product quality during production is in the foreground. In this context, an AI-based process monitoring system is developed and evaluated using the example of laser-welded seams on galvanized car body components, which is able to distinguish between different seam irregularities and process deviations during the process. In addition, the neural network-based AI system will be extended to extract characteristic process features from the image data, which will provide the informational foundation for downstream process control. Application-specific optimizations of the neural network architecture are also part of the investigations. The metrological basis for the quality assurance system is provided by image-based sensors in different observation configurations, which also allow a comparison of individual image features with regard to their detection performance. An additional evaluation of the seam quality takes place in the form of a hybrid modelling of the weld penetration depth with subsequent calibration on the basis of specific image features. The hybrid model allows the weld penetration depth to be calculated based on current image data and the process parameters used during the process. Since an evaluation of the uncertainty of the used AI system is crucial in the context of this application, an approach is presented that allows the estimation of the epistemic uncertainty of the neural network based on outlier detection. Ultimately, a process control system will be implemented and tested using algorithms from the field of reinforcement learning, which promise a high degree of adaptability to new process conditions. The overall system is examined with respect to its real-time capability and finally evaluated on the basis of experimental investigations to determine the achievable defect detection and mitigation performance.
%F PUB:(DE-HGF)11 ; PUB:(DE-HGF)3
%9 Dissertation / PhD ThesisBook
%R 10.18154/RWTH-2025-04385
%U https://publications.rwth-aachen.de/record/1010863