Optical tomography and machine learning for in-situ defects detection in laser powder bed fusion: A self-organizing map and U-Net based approach

Abstract

The inherently stochastic nature of the laser powder bed fusion (LPBF) process presents a significant challenge in developing dependable and efficient defect detection algorithms that can flexibly adjust to a variety of machine configurations and process parameters. The present study proposes a machine learning-based approach that employs the optical tomography (OT) data acquired during LPBF for identifying defects, specifically lack of fusion and keyhole porosity. This novel approach includes a machine learning framework consisting of a self-organizing map (SOM) and a custom U-Net model, to predict porosity effectively and automatically across different process parameter choices, making it robust and computationally efficient. The proposed approach was validated with various sets of experiments, including analyzing the influence of process parameter changes, as well as intentionally and randomly mimicking these defects. These defects predicted by the developed model are validated using computed tomography (CT)-scanning to evaluate the algorithm’s performance. The proposed approach effectively predicted porosity caused by lack of fusion or keyhole for different process parameter choices by comparing the similarity based on Euclidean distance between the normalized porosity curves for CT scan and the developed model using Dynamic Time Wrapping (DTW) technique, leading to an average distance score of 0.243 for randomized lack of fusion pores and an average score of 0.6 for randomized keyhole pores. These results demonstrate the effectiveness of the proposed approach in predicting defects during the LPBF process, and its potential to improve in-situ monitoring and quality assurance.