Knowledge Constrained Deep Clustering for Melt Pool Anomaly Detection in Laser Powder Bed Fusion

Abstract

The rapid expansion of the manufacturing sector has brought laser-based metal additive manufacturing, like laser powder bed fusion, to the forefront of innovation. Yet, its widespread acceptance hinges on overcoming numerous obstacles, including uncertainties regarding part quality when employing standardized materials in additive manufacturing procedures. Clustering techniques are essential in uncovering patterns within data sets, particularly in the field of additive manufacturing, where understanding the behavior of meltpool images is crucial for process optimization. Traditional hierarchical clustering methods often lack the ability to incorporate domain-specific knowledge, limiting their effectiveness in this field. In this study, we propose a novel approach that integrates knowledge-constrained hierarchical clustering with encoded meltpool image sequences. By incorporating domain-specific constraints, our approach aims to enhance clustering accuracy and provide more interpretable cluster assignments. Specifically, our approach demonstrates improvements in clustering performance, as measured by the Calinski-Harabasz index. The base model achieved an index of 2.07, while the Constrained model attained 5.47, indicating a substantial improvement in clustering structure. This enhanced performance enables a more accurate evaluation of printed parts’ quality based on melt pool image sequences.