Predicting defects in SLM-produced parts based on melt pools clustering analysis

The paper proposes a machine learning (ML) approach for clustering analysis to classify images from video recordings of melt pools during selective laser melting (SLM) printing process. By employing this method, each moment in the video is categorized into specific groups based on thermal impact. This allows for an understanding of the heat released on the surface at any given moment during printing. The greater the thermal impact on the powder, the higher the likelihood of defect occurrence. t-SNE method was used for dimensionality reduction of the image vectors. Subsequently, k-means and DBSCAN algorithms were applied to cluster these reduced dimensions. The resulting clustered images were then analyzed by an expert to determine the type of class: overheating, underheating, or normal. By identifying the location on the part where the image was taken, it was possible to trace the origin of potential defects.