In this study, we developed a process that allows for evaluating artificial intelligence-aided detection and self-healing of defects in the proximity of standard process conditions. The build parameters were manipulated to seed defects in a tailor-made notch area of the test specimen. A near-infrared Scientific Complementary Metal-Oxide-Semiconductor (sCMOS)-based optical tomography (OT) system and a dual-photodiode-based meltpool monitoring (MPM) system were used to capture layer-wise thermal signatures of these process variations. Based on these data, a lightweight late-fusion ensemble deep learning pipeline was developed to evaluate defect detection using modality-specific predictions, and the results were verified by optical metallography. This work examines whether two widely available, imaging in-situ sensing modalities can provide distinct and complementary information for defect detection. The results show that OT and MPM capture different aspects of the process, and that their score-level fusion improves prediction confidence relative to single-modality models in the present controlled experimental setting. Separately, defects in the notch area comprising up to 7 defective layers were healed when the region was subsequently exposed to under near-standard process parameters, confirming the self-healing phenomenon. These self-healing observations motivate the need for reliable in-situ defect identification as a prerequisite for future detect-and-heal process control. Overall, the study highlights a practical route toward accessible, cost-effective imaging quality monitoring in resource-constrained additive manufacturing environments, hence lowering the barrier to entry in this domain. • Self-induced porosity in laser-based powder bed fusion, by varying volumetric energy density close to natural process. • Capture of distinct process cues via optical tomography (OT) and meltpool monitoring (MPM) imaging sensors. • In-situ defect detection via OT and MPM modalities as cost-effective in-situ monitoring for resource-constrained environments. • High-confidence complementary global (OT) and local (MPM) defect signature quantification using multimodal ensemble learning. • Investigation of self-healing defective regions up to 7 layers, when exposed by the standard VED, motivates ”detect-and-heal” closed-loop control.
Doosti et al. (Sun,) studied this question.