Defect Prediction in Directed Energy Deposition using an Ensemble of Clustering Models

Directed energy deposition (DED) stands as a pivotal additive manufacturing technique, revolutionizing the landscape of modern manufacturing. However, process-related defects hin-der its broad application across different sectors. In this paper, we propose a novel methodology for the in-situ prediction of defects in DED processes based on the thermal images of the melt pools. Initially, multiple features, summarizing the thermal and geometric characteristics of the melt pool, were extracted. Based on these features, an ensemble of unsupervised clustering models was constructed to distinguish anomalies - images with defects - from the defect-free images. Roughly 3% of the acquired images were predicted to include defects. Upon visual inspection, these images exhibited distinctive thermal distributions and geometric configurations compared to the remaining dataset. Furthermore, a 2D approximate visualization of the feature space revealed the clustering structure of the thermal images in their feature space. This visualization showed that the anomalies could be distinguished from the majority of normal images, which further validates the prediction model's effectiveness.