Drift Detection in Selective Laser Melting (SLM) Using a Machine Learning Approach

Abstract

Selective laser melting has seen a growing demand in niche industry applications for its complexity-free manufacturing capabilities. Besides, all the advantages over traditional manufacturing techniques, reliability, and repeata-bility is still a challenge. In-situ monitoring systems comprising high precision sensors such as photodiodes, High-speed IR cameras both in co-axial and off-axis positions have been installed in commercially available machines to improve the overall performance of the process. However, understanding the correlation between the acquired data from sensors and build quality is a challenge and time-consuming. Thus, the sensitivity analysis of sensors installed on the machine to various defects and easy detectability of the drift during the process is the need of the time. In this work, a sensitivity analysis of the melt pool monitoring system installed in the commercial SLM 280 HL machine is presented. Also, a supervised machine learning approach for in-line detection of the drift in the final part is used. A balanced labeled dataset for training the support vector machine algorithm by inducing drifts in the parts artificially (overheating and lack of fusion) was prepared. Then the trained classifier successfully classifies the layers of the unlabeled case study samples into "drift" and "no-drift" labels. Here, drift classification is based on overheating and lack of fusion defects. Thus, the machine learning approach is robust and cost-effective for techniques like selective laser melting, where to obtain labeled data is time-consuming and expensive.