Abstract
EDF's industrial objective is to better control welding operations to improve the quality of the assembly of large components. Currently, all these parts are inspected after manufacture and any identified defect must be repaired. For many years, the in-situ control of welds during the manufacturing process has been a crucial point for industries that have high quality requirements for their welded parts. These constraints impose the development of a control chain to ensure weld quality during the manufacturing process. This control is mainly based on monitoring the process parameters and the weld pool dynamics. In this thesis, several measurement methods are developed according to a well-defined specification to answer the industrial challenges.A contactless sensor system allowing to show the process and data processing from these sensors is developed. This instrumentation contains several sensors for monitoring and analyzing the influence of welding parameters (voltage, current, ...). In parallel, cameras are set up giving access to weld pool geometric data. This allows us to analyze the influence of the weld pool dynamics on the final shape of the bead with respect to variations of the heat input or the welding position. A database is created on the one hand for the storage and labeling of the data and on the other hand to feed machine learning algorithms in its training, validation, and prediction steps. Two different approaches are explored using machine learning. A first approach is developed to guarantee the conformity of the welding procedure description. A second one is used to anticipate the appearance of defects and the detection of deviation during the manufacturing process. Finally, all the developed instrumentation and methods will be tested on an industrial application to showcase online control on a real configuration.