B. Zhang, S. Liu, Y.C. Shin
Keywords: additive manufacturing, porosity, in-process monitoring
Summary:This work describes realtime in-process porosity monitoring for additive manufacturing processes based on deep learning and a real time weld pool monitoring system. A high-speed digital camera was mounted coaxially to the laser beam for in-process sensing of melt-pool data, and deep learning convolutional neural network models were designed to learn melt-pool features to predict the porosity attributes in built specimens during additive manufacturing. The convolutional neural network (CNN) models with a compact architecture, part of whose hyper parameters were selected through cross-validation analysis, achieved a classification accuracy of 91.2% for porosity occurrence detection in the direct laser deposition of sponge Titanium powders and presented predictive capacity for micro pores below 100 µm. For local volume porosity prediction, the model also achieved a root mean square error of 1.32% and exhibited high fidelity for both high porosity and low porosity specimens. The system is low cost and can be easily outfitted on many commercial additive manufacturing systems.