Analysis of unsupervised and semi-supervised machine learning techniques for print defect detection during laser powder bed fusion
J. Power, Denis P. Dowling, Shane Keaveney, Cathal Hoare
Abstract
Abstract This research examines the use of an unsupervised and semi-supervised machine learning approach to facilitate the extraction and analysis of process monitoring data during the printing of Ti-6Al-4 V alloy parts in real time. The detection of processing anomalies is achieved by analysing in situ photodiode sensor data gathered from a combination of laser power output data, along with photodiode data obtained from the laser-powder bed fusion (L-PBF) laser melt pool emissions. The machine learning techniques evaluated are the unsupervised Search and TRace AnomalY (STRAY) algorithm, along with the semi-supervised 1-dimensional autoencoder (1-DAE) technique. The latter approach is considered semi-supervised, as it requires an unlabelled training dataset for the model to learn inherent latent features of the data. To this end, the techniques were assessed on their ability to detect defects intentionally induced during the printing of lattice structures with a range of unit cell shapes and sizes. Based on the analysis of the print datasets, it was demonstrated that the 1-DAE achieved a maximum predictive accuracy (F1 score) of 0.94 and a minimum score of 0.80. Comparatively, the STRAY algorithm had a minimum F1 score of 0.09 and a maximum score of 0.92 when applied to the same datasets. While 1-DAE requires an initial training dataset before it can be applied, its implementation time post-training is only 1.92 ms. This speed along with its reliability makes it a potential candidate for the real-time detection of L-PBF processing anomalies.