Targeted disentangling of melt pool features for layer-wise printing quality assessment in L-PBF

Ensuring the stability and consistency of laser-powder bed fusion (L-PBF) additive manufacturing remains a persistent challenge in the industry, which has led to growing research interest in process monitoring in recent years. Among various monitoring techniques, coaxial melt pool imaging stands out as one of the most promising approaches. However, achieving stable and scalable print quality assessment based on coaxial melt pool images remains challenging. Deep learning methods often suffer from limited interpretability and weak generalization, while traditional image processing approaches tend to lack flexibility and exhibit low discriminative accuracy. To address these issues, this paper proposes an interpretable directional feature disentanglement framework designed to enable the extraction of strongly-correlated physical features with structures from melt pool images for printing quality assessment. Specifically, a feature anchoring module is incorporated into a variational autoencoder (VAE) generation framework to stabilize the position of the disentangled target features in the latent space. A multi-stage, multi-task training strategy is then introduced to progressively accomplish melt pool image reconstruction, feature anchoring, and feature disentanglement. Finally, the effectiveness of the proposed framework is verified by cross-device and cross-material experiments involving unsupported overhang structures, which proves that it is a technology worthy of engineering promotion.