Directed Graphical Model for Real-Time Process Monitoring in Additive Manufacturing

One of the grand challenges for additive manufacturing and 3D printing process technologies is accurate and repeatable fabrication quality. Largely, approaching this challenge is focused on the deposition process. Current approaches tightly control the process parameters and input material quality. When material properties may be accurately controlled in additive manufacturing the full potential of this process will be unleashed. As of yet, this is still an open question. This paper proposes the use of statistical learning techniques in conjunction with iterative material study to identify and compute the sources of defects and local material properties in additively manufacturing goods. The model makes use of the element-by-element fabrication of additive manufacturing and the time-series material changes. The deposition plan for a part is segmented into volume elements, called voxels. Each voxel of deposited material is treated as an independent sample of the process parameter effects. The time series of deposition is treated as a Hidden Markov Chain where the control parameters and measurable emissions as known quantities. The state of the material is a hidden variable with occasional observation. The hidden variable is approximated using material models and updated with available post-fabrication testing results to train the distribution embedded in the Hidden Markov Chain. The results indicated that a physics-based material state transition matrix can be trained to give insight into the real-time quality of the deposited material. As noted broadly, thermal emissions are useful for understanding the mass and heat transfer phenomena in these processes. In addition, acoustic emissions give insight for welding and are added here. The thermal and acoustic emissions of the material transition are used as observable data to update the prediction as the process proceeds. When the state transition matrix is used in conjunction with final material properties process variability and control errors may be better identified and strategies employed to correct them. These results have wide ranging implications as a computationally effective means of iterative process improvement for additive manufacturing, designing new control strategies, and revealing the real-time state of volume elements as they are deposited. The method is limited only to processes that can be approximated by voxel deposition and the size of those voxels. The scalable formulation is limited by real-time data collection capacity. Data collection capacity is inversely proportional to the voxel size. The state estimation calculation also depends on the efficiency of Baum-Welch technique. This approach moves closer to a predictive model which includes current information on the state of the process to update the prediction.