AM Analytics: Automated in-situ defect detection in powder bed-based additive manufacturing
The highly complex Laser Powder Bed Fusion (L-PBF) process, which only runs stably within narrow process windows, currently exhibits instabilities that are insufficiently detected by sensors. These instabilities are the cause of the occurrence of component defects and significantly determine the resulting component quality.
Up to now, the component quality can only be realised downstream of the L-PBF process destructively on accompanying samples to determine the mechanical properties or non-destructively via μCT scans as a high-resolution option for three-dimensional integrity testing of the components. Both methods involve an additional step in the process chain, are limited in their informative value (e.g. μCT scans are not applicable for high-strength alloys) and, due to the process, cause considerable costs while at the same time requiring a lot of time. An early in-situ defect detection procedure can make downstream quality assurance more efficient and thus reduce costs. Such a reliable in-situ process monitoring and quality assurance does not exist today and is therefore to be developed in this project. The following scientific problems still need to be addressed for such a solution:
Challenge I: In-situ sensor technology
Currently, there is no suitable in-situ sensor concept that can detect all typical component defects. Although there are many different in-situ measurement methods, such as melt pool monitoring (MPM), optical tomography (OT) and optical powder bed monitoring (PBM) from the L-PBF plant manufacturers, there is a lack of reliable measurement data and analysis methods. This means that it is not possible to achieve a well-founded statement about the process and component quality. Among other things, the sensor technology currently available is not suitable for recording the height (topography) of powder layers. Especially in the layer structure a local or global height deviation is critical to quality and would therefore have to be recorded with suitable sensor technology.
Challenge II: Data analysis
Despite many developments and innovations in the field of machine learning, no theory exists today that allows a correct choice and configuration of a learning procedure for the given problem. As easy as it often is to achieve first quick results in data analysis, it is difficult to generate learned models in the quality necessary for engineering applications. Existing data analysis methods do not yet allow classification of sensor data with regard to product quality in the AM sector. In addition to the lack of sensor technology mentioned above, there are challenges in the area of machine learning methods:
1) Integration of domain-specific a-priori knowledge: In order to secure the learning results and to reduce the necessary amount of data, a-priori knowledge such as material science and physics knowledge as well as process knowledge must be integrated into the learning methods. Although there is preliminary work in this area, there are no approaches for AM technology so far.
2) Explainability of learning outcomes: Learning outcomes, such as neural networks, are difficult to comprehend for humans. There is a need for an explanatory component that makes the models interpretable for humans and thus achieves user acceptance. Approaches to solving this problem also exist and need to be further developed for this domain.
3) Network architecture: Neural networks should be used for data analysis. For this purpose, an architecture suitable for the sensor technology and the task must be developed based on CNNs or gated neural networks, such as GRU, and based on generative networks, such as VAE.
Project duration: 10.2021 – 08.2023
Funding programme: IGF
Letzte Änderung: 2. August 2022