Monitoring of Laser Welding via Coaxial Imaging and Acoustic Emission

Abstract

Laser Welding (LW) offers multiple advantages compared to previous welding technologies and is a key technology in battery manufacturing and aerospace industry. However, process discontinuities hinder broader market adoption for use cases such as welding thick plates for structural steel. Quality inspection systems and automated process control are a hot topic that promise to increase the automation level in LW and deliver higher quality manufactured components. The aforementioned automated systems require computer software that measures process parameters from sensors, and in this context previous studies demonstrated the potential of coaxial imaging and Acoustic Emission (AE) for LW monitoring, of which the main advantages are the online process monitoring capability and ease of adaptation into existing LW systems. Specifically, laser-illuminated cameras offer great advantages for monitoring since they use band pass filters, removing process noise and delivering high quality images; In parallel, the development of optical microphones opened vast possibilities for acoustic monitoring of signals in the order of MHz. This thesis studies two different Machine Learning (ML) methods for predicting process parameters in LW. In the first study, coaxial images were used to predict the penetration depth using supervised learning and Optical Coherence Tomography as ground truth data. In the second study, AE was used for laser power prediction, where two alternatives were compared: a Deep Learning model and a statistical feature extraction library combined with out-of-the-shelf ML regressor. For evaluating both alternatives, train-test validation was used, as well as regression scoring functions. The Coefficient of Determination (R-squared) was applied for both methodologies. For image-based prediction of the keyhole depth, the R-squared was of 0.59 for the best strategy, and for the AE-based laser power prediction, an R-squared of 0.93 was obtained in the best strategy. Both results are successful indicators of using coaxial imaging and optically sourced AE with ML to predict LW parameters. In the case of imaging, improvements are suggested to obtain better results, while in the case of AE, the results strongly indicate the success in capturing the data variance. In a wider perspective, this work presents advancements in process monitoring for LW that could be applied in quality control, self-correcting systems, and parameter optimization.