Estimating Pore Location of PBF-LB/M Processes with Segmentation Models

TL;DR

This work tackles the challenge of precise, within-layer pore localization in PBF-LB/M by reframing localization as estimating a pore probability field. It introduces Gaussian kernel density estimation (KDE) to convert discrete CT-derived pore positions into a continuous label $x_{PP}$, enabling segmentation models to learn mappings from in-situ monitoring data (HR and OT) to pore probability: $\mathrm{KDE}(\mathbf{v}) = \frac{1}{\beta Q} \sum_{i=1}^{Q} N\left(\frac{\mathbf{v}-\mathbf{v_i}}{\beta}\right)$ with $\beta=20$. The authors train and compare five segmentation architectures (Unet, FPN, LinkNet, DeepLabV3, MA-Net) on data from two geometries across varied process parameters, using MAE as the loss. Results show comparable performance across models, with parameter settings generally having a larger impact than geometry, and demonstrate the feasibility of accurate pore probability estimation from in-situ data, paving the way for real-time defect control and more precise pore detection systems. Limitations include labeling accuracy and potential misalignment between CT and in-situ data, suggesting future work on improved labeling and sequential layer predictions.

Abstract

Reliably manufacturing defect free products is still an open challenge for Laser Powder Bed Fusion processes. Particularly, pores that occur frequently have a negative impact on mechanical properties like fatigue performance. Therefore, an accurate localisation of pores is mandatory for quality assurance, but requires time-consuming post-processing steps like computer tomography scans. Although existing solutions using in-situ monitoring data can detect pore occurrence within a layer, they are limited in their localisation precision. Therefore, we propose a pore localisation approach that estimates their position within a single layer using a Gaussian kernel density estimation. This allows segmentation models to learn the correlation between in-situ monitoring data and the derived probability distribution of pore occurrence. Within our experiments, we compare the prediction performance of different segmentation models depending on machine parameter configuration and geometry features. From our results, we conclude that our approach allows a precise localisation of pores that requires minimal data preprocessing. Our research extends the literature by providing a foundation for more precise pore detection systems.

Figures (7)

• Figure 1: Geometries selected for our experiments are (a) Complex Geometry and (b) XYZ Cube. (c) shows the relative arrangement of part positions $p = 1, \dots , \Pi$ within our building platform.
• Figure 2: PBF-LB/M experiment and monitoring setup, displayed (a) schematically with its corresponding (b) real-world implementation at the Digital Additive Production chair (DAP).
• Figure 3: Samples of the training dataset, showing images of two parts $p=2$ and $p=6$. Five exemplary layers are shown for each part, with their corresponding data modality HR, OT, CT, and PP arranged from left to right respectively. For the pore probability distribution, blue indicates low and red high probability values.
• Figure 4: Box plots of MAE for all experiments. Model architectures are color-coded. The circles above the whiskers can be considered outliers. The smaller the MAE the better the model performance.
• Figure 5: Box plots of MAE values for each model over Geometry Features derived from the complex geometry. Geometry features are derived from layers with different geometry features.