Abstract Scope |
This work focuses on the analysis of laser powder bed fusion images and acoustic data as a means to monitor laser powder bed additive manufacturing processes for key outcomes. This would enable robust quality assurance, control and optimization of component properties, and improvement of process stability while reducing operator burden. Process mapping was employed in determining parameter sets that would reliably induce keyholing, lack-of-fusion, bead-up, laser spatter, as well as fully dense components. Using data collected during builds using these parameter sets, bag of words (BOG), support vector machines (SVM), convolutional neural networks, and long-short term memory (LSTM) networks were shown to be successful in classifying flaws. Future work will involve the use of more advanced deep learning architectures such as U-Net to localize the detection of laser spatter in fusion images as well as the application of the more advanced transformer architecture to the processing of acoustic signals. |