Metal-based additive manufacturing processes are increasingly used in industry to produce complex-shaped components. In this regard, the laser-based Powder Bed Fusion process (PBF-LB/M) is one of the key technologies due to its capability to produce components in high spatial accuracy. The formation of porosity during manufacturing poses a serious risk to the safety of the printed parts. For quality assessment, in-situ monitoring technologies such as thermography can be used to capture the thermal history during production. It was shown that discontinuities within the thermal history can be correlated with the probability of porosity or defect formation. In this context, Machine Learning (ML) algorithms have achieved promising results for the task of porosity prediction based on thermographic in-situ monitoring data. One important technique is the use of thermogram features for porosity prediction that are extracted from the raw data (e.g., features related to the melt pool geometry or spatter generation). However, the reduction from large thermogram data to discrete features holds the risk of losing potentially important thermal information and, thereby, introducing bias in the model.

Therefore, we present a raw data-based deep learning approach that uses thermographic image sequences for the prediction of local porosity. The model takes advantage of the self-attention mechanism that considers not only the thermogram information but also its positional context within the sequence. The model is used to predict porosity in the form of a many-to-one regression. It is trained and tested on a dataset retrieved from the manufacturing of HAYNES282 cuboid specimens. The model results are compared against state-of-the-art thermogram feature-based ML models and artificial neural networks. The raw data model outperforms its feature-based counterparts in terms of prediction scores and, therefore, seems to make better use of the information available in the thermogram data.