作者
Berkay Bostan,Shawn Hinnebusch,David H. Anderson,Albert C. To
摘要
Porosity critically impacts the reliability and performance of metal laser powder bed fusion (LPBF) parts, affecting properties like fracture toughness and fatigue life. This work proposes a deep learning (DL) framework to predict local porosity in LPBF Inconel 718 parts using in-situ infrared (IR) camera imaging where parts are produced under standard conditions, resulting in 0.03 % overall porosity. The framework achieves over 90 % balanced accuracy for detecting pores above 34 μm at a 360 μm sensor resolution. First, input features include six physics-based IR signatures (cooling rate, heat intensity, interpass temperature, relative melt pool area, spatter generation, and maximum predeposition temperature) and local scan vector length, all linked to porosity generation mechanisms. Second, the framework considers feature interactions across the current pixel and its 26 nearest neighbors. Third, special convolutional filters are developed to filter heat intensity and cooling rate features at edges and stripe boundaries, compensating for limited camera resolution in those regions. Ground truth data on pore size and locations are gathered through serial sectioning and optical microscopy. In unseen parts with varying geometrical features, the framework achieves a true positive rate above 88 % and a false negative rate below 4 % for pores over 34 μm. The proposed DL framework is rigorously compared to traditional machine learning models, demonstrating its superiority in terms of faster training, higher prediction speed, smaller size, and robust performance on unseen test blocks. Additionally, Shapley Additive Explanations analysis elucidates pore formation mechanisms, revealing complex feature interactions across different regimes. Results align well with known pore formation mechanisms, indicating that the developed algorithm interprets complex relationships between features and porosity. This work enhances in-situ porosity detection in LPBF and advances the understanding of pore formation mechanisms. • The proposed deep learning framework predicts local porosity in LPBF Inconel 718 parts using physics-based features from in-situ IR imaging. • Tested on unseen blocks, the model achieves over 90 % balanced accuracy in detecting pores larger than 34 µm, despite 0.03 % overall porosity. • SHAP analysis shows the model interprets complex pore formation mechanisms, linking physics-based IR features to porosity formation.