Logistic regression classification to predict regional anomalies in nominally printed volume of separate test pieces

Supervised machine learning techniques have struggled to accurately predict voxel-wise occurrence of anomalies in metal powder bed parts printed with optimal processing parameters. This work discusses a method to visualize machine learning model predictions in 3D to interrogate patterns in the predictions. A simple logistic regression classifier, with cross validation and an optimized classification threshold, is trained using synthetic in situ features, a machine parameter, and post-process output labels. The developed classifier is shown to outperform deep learning and boosted classifiers on the datasets used. Voxel-wise prediction performance is very low, but 3D representation of model predictions shows the developed model can predict anomalies in the correct region of the printed part. The practical use of the developed method is demonstrated by predicting the occurrence of anomalies in nominally printed volume using a model that had been trained on a dataset printed with induced defects.