Search

A procedure to label Laser Powder Bed Fusion video data and leverage it to train a convolutional neural network is demonstrated in article number 1800136, by Bodi Yuan, Brian Giera, and co‐workers. Testing the neural network reveals it can predict track continuity and the average and standard deviations of track width from high speed video alone, reducing the need for post‐build quality assessment. [ABSTRACT FROM AUTHOR]

A two‐step machine learning approach to monitoring laser powder bed fusion (LPBF) additive manufacturing is demonstrated that enables on‐the‐fly assessments of laser track welds. First, in situ video melt pool data acquired during LPBF is labeled according to the (1) average and (2) standard deviation of individual track width and also (3) whether or not the track is continuous, measured postbuild through an ex situ height map analysis algorithm. This procedure generates three ground truth labeled datasets for supervised machine learning. Using a portion of the labeled 10 ms video clips, a single convolutional neural network architecture is trained to generate three distinct networks. With the remaining in situ LPBF data, the trained neural networks are tested and evaluated and found to predict track width, standard deviation, and continuity without the need for ex situ measurements. This two‐step approach should benefit any LPBF system – or any additive manufacturing technology – where height‐map‐derived properties can serve as useful labels for in situ sensor data. A procedure to label laser powder bed fusion video data and leverage it to train a convolutional neural network is demonstrated. Testing the neural network reveals that it can predict track continuity and the average and standard deviations of track width from high‐speed video alone, reducing the need for postbuild quality assessment. [ABSTRACT FROM AUTHOR]