Spatiotemporal analysis of powder bed fusion melt pool monitoring videos using deep learning

For several years now, in-situ process monitoring has been proposed as the enabler of fast and flexible part qualification in powder bed fusion manufacturing. However, the predictive tools built using in-situ sensor data have so far lacked the accuracy and precision needed for this purpose. In this paper, we introduce a convolutional recurrent neural network to extract spatiotemporal features from melt pool monitoring videos and accurately detect adverse surface quality in powder bed fusion specimens. Experimental data gathered at the National Institute of Standards and Testing (NIST) was used in the study, consisting of two separate builds of a component containing overhanging topographical features. Model training and validation were performed using data from separate builds. The model was found to accurately detect scan positions corresponding with poor surface roughness, achieving an overall accuracy of over 99% on the full test set and a low false positive rate, below 0.01, on selected build layers. Further, the model was found to be computationally efficient, with the optimal model predicting on a full layer in 107 s. The results demonstrate the potential of extracting both spatial and temporal features from monitoring video data, with the temporal dimension largely neglected by the field to date. The presented model enables real-time layerwise detection of unacceptable surface finish in powder bed fusion components.