In-Situ monitoring of porosity based on static and dynamic molten pool features in laser powder bed fusion

Abstract

Porosity is one of the most serious defects for laser powder bed fusion (LPBF) which is a promising metal additive manufacturing (AM) technique. Reducing porosity is essential to improve the mechanical properties of parts in high-end applications. While existing research primarily focuses on static molten pool features, such as size and shape, these methods overlook critical local details, leading to imprecise porosity monitoring. To address this gap, this study proposes a novel attention-based method for in-situ monitoring of porosity (AMNet) from hybrid static and dynamic features (also referred to as motion features) extracted from the molten pool. Dynamic or motion features represent the molten pool by computing the distance from its edge to the centroid in the image. In the proposed AMNet, we adopt a self-attention mechanism to extract representative features from static and motion features separately. Furthermore, cross-attention is introduced to fuse two kinds of features in the latent space for porosity detection. To verify the effectiveness of AMNet, we collect a dataset consisting of molten pool and porosity data with three categories. The proposed AMNet achieves superior performance with an accuracy of 99.4 % on our dataset. Even 92 % accuracy is obtained with only 20 % of the training set by AMNet. The inference time of AMNet, averaging around 6.25 ms per sample has the potential to enhance closed-loop quality control systems by providing valuable real-time feedback. These results demonstrate the effectiveness and efficiency of the proposed method in porosity detection during LPBF.