Machine learning approaches for process monitoring in powder bed fusion of polymers

Abstract

In recent years, laser-based powder bed fusion technology has become an integral part of the manufacturing industry, enabling the production of highly complex component structures with strength comparable to that of conventionally manufactured components. In particular, the reduction in component volume makes a significant contribution to resource-efficient production and use. This opens up new design possibilities that are simply not possible with conventional manufacturing processes. For example, functions can be integrated into components. For industrial applications, highly customized products can be produced that are only used once because it is not worthwhile to use conventional manufacturing processes for them. However, as advanced as this technology is, the nature of the process itself means that uncontrollable phenomena can occur during the process which can cause defects in the powder bed and consequently affect the quality of the component. This thesis deals with approaches to monitor the process of laser-based powder bed fusion of polymers (PBF-LB/P) using appropriate sensor technology and machine learning. The experimental investigations were carried out on an EOS powder bed fusion machine (FORMIGA P 110), which sinters PA12 polyamide powder layer by layer into component structures using a 30 W CO2 laser with a wavelength of 10.6 μm. As defects can occur unpredictably and at irregular intervals during this process, this work has developed a method to artificially induce different types of defects at different stages of the process. Curling, part shifting and particle drag could be simulated in a sustainable way during the running process without interrupting it. The next step was to evaluate different sensor technologies and integrate them into the machine to measure and record these defects. In this case, infrared thermal imaging, laser triangulation and an RGB camera, mounted both outside and even inside the machine, proved to be suitable methods. This has shown that it is possible to detect the defects at least once using these sensor concepts, which leads to the final step of this thesis, automated process monitoring using machine learning. The extensive datasets were then used to train convolutional neural networks based on deep learning, demonstrating that automated and intelligent defect detection is possible in the PBF-LB/P process. Different networks were applied, compared and tuned by adding features and tuning hyperparameters to make a statement about the performance of the networks in combination with the available data. Simulated defects in the process were predicted with over 99% accuracy using different CNNs and different data formats. In summary, this work combines comprehensive investigations into the potential of integrating different sensor technologies for targeted defect detection in the powder bed with extensive experiments on the sustained simulation of different defects and their prediction using deep learning-based computer vision. In the future, the empirically collected data could be used to determine what type of defect it is, which parameters in the software need to be adjusted during the process when predicting a defect, in order to take countermeasures in real time.